Daily Papers

The project automatically fetches the latest papers from arXiv based on keywords.

The subheadings in the README file represent the search keywords.

Only the most recent articles for each keyword are retained, up to a maximum of 20 papers.

You can click the 'Watch' button to receive daily email notifications.

Last update: 2026-04-03

LLM Reasoning

Title	Date	Abstract	Comment
Online Reasoning Calibration: Test-Time Training Enables Generalizable Conformal LLM Reasoning	2026-04-01	Show While test-time scaling has enabled large language models to solve highly difficult tasks, state-of-the-art results come at exorbitant compute costs. These inefficiencies can be attributed to the miscalibration of post-trained language models, and the lack of calibration in popular sampling techniques. Here, we present Online Reasoning Calibration (ORCA), a framework for calibrating the sampling process that draws upon conformal prediction and test-time training. Specifically, we introduce a meta-learning procedure that updates the calibration module for each input. This allows us to provide valid confidence estimates under distributional shift, e.g. in thought patterns that occur across different stages of reasoning, or in prompt distributions between model development and deployment. ORCA not only provides theoretical guarantees on conformal risks, but also empirically shows higher efficiency and generalization across different reasoning tasks. At risk level $δ=0.1$, ORCA improves Qwen2.5-32B efficiency on in-distribution tasks with savings up to 47.5% with supervised labels and 40.7% with self-consistency labels. Under zero-shot out-of-domain settings, it improves MATH-500 savings from 24.8% of the static calibration baseline to 67.0% while maintaining a low empirical error rate, and the same trend holds across model families and downstream benchmarks. Our code is publicly available at https://github.com/wzekai99/ORCA.	20 pages
Reasoning Shift: How Context Silently Shortens LLM Reasoning	2026-04-01	Show Large language models (LLMs) exhibiting test-time scaling behavior, such as extended reasoning traces and self-verification, have demonstrated remarkable performance on complex, long-term reasoning tasks. However, the robustness of these reasoning behaviors remains underexplored. To investigate this, we conduct a systematic evaluation of multiple reasoning models across three scenarios: (1) problems augmented with lengthy, irrelevant context; (2) multi-turn conversational settings with independent tasks; and (3) problems presented as a subtask within a complex task. We observe an interesting phenomenon: reasoning models tend to produce much shorter reasoning traces (up to 50%) for the same problem under different context conditions compared to the traces produced when the problem is presented in isolation. A finer-grained analysis reveals that this compression is associated with a decrease in self-verification and uncertainty management behaviors, such as double-checking. While this behavioral shift does not compromise performance on straightforward problems, it might affect performance on more challenging tasks. We hope our findings draw additional attention to both the robustness of reasoning models and the problem of context management for LLMs and LLM-based agents.	Prepr... Preprint, work in progress
PluriHopRAG: Exhaustive, Recall-Sensitive QA Through Corpus-Specific Document Structure Learning	2026-04-01	Show Retrieval-Augmented Generation (RAG) has been used in question answering (QA) systems to improve performance when relevant information is in one (single-hop) or multiple (multi-hop) passages. However, many real life scenarios (e.g. dealing with financial, legal, medical reports) require checking all documents for relevant information without a clear stopping condition. We term these pluri-hop questions, and formalize them by 3 conditions - recall sensitivity, exhaustiveness, and exactness. To study this setting, we introduce PluriHopWIND, a multilingual diagnostic benchmark of 48 pluri-hop questions over 191 real wind-industry reports, with high repetitiveness to reflect the challenge of distractors in real-world datasets. Naive, graph-based, and multimodal RAG methods only reach up to 40% statement-wise F1 on PluriHopWIND. Motivated by this, we propose PluriHopRAG, which learns from synthetic examples to decompose queries according to corpus-specific document structure, and employs a cross-encoder filter at the document level to minimize costly LLM reasoning. We test PluriHopRAG on PluriHopWIND and the Loong benchmark built on financial, legal and scientific reports. On PluriHopWIND, our method shows 18-52% F1 score improvement across base LLMs, while on Loong, we show 33% improvement over long-context reasoning and 52% improvement over naive RAG.
Demystifying Chains, Trees, and Graphs of Thoughts	2026-04-01	Show The field of natural language processing (NLP) has witnessed significant progress in recent years, with a notable focus on improving large language models' (LLM) performance through innovative prompting techniques. Among these, prompt engineering coupled with structures has emerged as a promising paradigm, with designs such as Chain-of-Thought, Tree of Thoughts, or Graph of Thoughts, in which the overall LLM reasoning is guided by a structure such as a graph. As illustrated with numerous examples, this paradigm significantly enhances the LLM's capability to solve numerous tasks, ranging from logical or mathematical reasoning to planning or creative writing. To facilitate the understanding of this growing field and pave the way for future developments, we devise a general blueprint for effective and efficient LLM reasoning schemes. For this, we conduct an in-depth analysis of the prompt execution pipeline, clarifying and clearly defining different concepts. We then build the first taxonomy of structure-enhanced LLM reasoning schemes. We focus on identifying fundamental classes of harnessed structures, and we analyze the representations of these structures, algorithms executed with these structures, and many others. We refer to these structures as reasoning topologies, because their representation becomes to a degree spatial, as they are contained within the LLM context. Our study compares existing prompting schemes using the proposed taxonomy, discussing how certain design choices lead to different patterns in performance and cost. We also outline theoretical underpinnings, relationships between prompting and other parts of the LLM ecosystem such as knowledge bases, and the associated research challenges. Our work will help to advance future prompt engineering techniques.
RefineRL: Advancing Competitive Programming with Self-Refinement Reinforcement Learning	2026-04-01	Show While large language models (LLMs) have demonstrated strong performance on complex reasoning tasks such as competitive programming (CP), existing methods predominantly focus on single-attempt settings, overlooking their capacity for iterative refinement. In this paper, we present RefineRL, a novel approach designed to unleash the self-refinement capabilities of LLMs for CP problem solving. RefineRL introduces two key innovations: (1) Skeptical-Agent, an iterative self-refinement agent equipped with local execution tools to validate generated solutions against public test cases of CP problems. This agent always maintains a skeptical attitude towards its own outputs and thereby enforces rigorous self-refinement even when validation suggests correctness. (2) A reinforcement learning (RL) solution to incentivize LLMs to self-refine with only standard RLVR data (i.e., problems paired with their verifiable answers). Extensive experiments on Qwen3-4B and Qwen3-4B-2507 demonstrate that our method yields substantial gains: after our RL training, these compact 4B models integrated with the Skeptical-Agent not only outperform much larger 32B models but also approach the single-attempt performance of 235B models. These findings suggest that self-refinement holds considerable promise for scaling LLM reasoning, with significant potential for further advancement.
KnowDiffuser: A Knowledge-Guided Diffusion Planner with LLM Reasoning	2026-04-01	Show Recent advancements in Language Models (LMs) have demonstrated strong semantic reasoning capabilities, enabling their application in high-level decision-making for autonomous driving (AD). However, LMs operate over discrete token spaces and lack the ability to generate continuous, physically feasible trajectories required for motion planning. Meanwhile, diffusion models have proven effective at generating reliable and dynamically consistent trajectories, but often lack semantic interpretability and alignment with scene-level understanding. To address these limitations, we propose \textbf{KnowDiffuser}, a knowledge-guided motion planning framework that tightly integrates the semantic understanding of language models with the generative power of diffusion models. The framework employs a language model to infer context-aware meta-actions from structured scene representations, which are then mapped to prior trajectories that anchor the subsequent denoising process. A two-stage truncated denoising mechanism refines these trajectories efficiently, preserving both semantic alignment and physical feasibility. Experiments on the nuPlan benchmark demonstrate that KnowDiffuser significantly outperforms existing planners in both open-loop and closed-loop evaluations, establishing a robust and interpretable framework that effectively bridges the semantic-to-physical gap in AD systems.	10pages, 1 figure
Learning to Reason in Structured In-context Environments with Reinforcement Learning	2026-04-01	Show Large language models (LLMs) have achieved significant advancements in reasoning capabilities through reinforcement learning (RL) via environmental exploration. As the intrinsic properties of the environment determine the abilities that LLMs can learn, the environment plays a important role in the RL finetuning process. An ideal LLM reasoning environment should possess three core characteristics: scalability, generalizable reasoning, and verifiability. However, existing mathematical and coding environments are difficult to scale due to heavy reliance on expert annotation, while the skills learned in game-based environments are too specialized to generalize. To bridge this gap, we introduce the \textbf{S}tructured \textbf{I}n-context \textbf{E}nvironment (SIE) framework. SIE achieves scalability by automatically constructing reasoning environments from large-scale structured data, where the rich compositional patterns naturally support generalizable reasoning. Moreover, the explicit schemas and reasoning chains in structured data provide a foundation for rule-based verifiability. Experimental results show that SIE framework not only achieves substantial improvements in in-domain structured reasoning, but also enables the learned compositional reasoning skills to generalize effectively to out-of-domain mathematical and logical reasoning tasks. We further explored learning in information-limited partial SIEs and found that LLMs can infer the missing information through exploring the environment, leading to robust reasoning improvements and generalization performance.
LocationReasoner: Evaluating LLMs on Real-World Site Selection Reasoning	2026-03-31	Show Recent advances in large language models (LLMs), particularly those enhanced through reinforced post-training, have demonstrated impressive reasoning capabilities, as exemplified by models such as OpenAI o1 and DeepSeek-R1. However, these capabilities are predominantly benchmarked on domains like mathematical problem solving and code generation, leaving open the question of whether such reasoning skills generalize to complex real-world scenarios. In this paper, we introduce LocationReasoner, a benchmark designed to evaluate LLMs' reasoning abilities in the context of real-world site selection, where models must identify feasible locations by reasoning over diverse and complicated spatial, environmental, and logistic constraints. The benchmark covers carefully crafted queries of varying difficulty levels and is supported by a sandbox environment with in-house tools for constraint-based location search. Automated verification further guarantees the scalability of the benchmark, enabling the addition of arbitrary number of queries. Extensive evaluations on real-world site selection data from Boston, New York, and Tampa reveal that state-of-the-art reasoning models offer limited improvement over their non-reasoning predecessors in real-world contexts, with even the latest OpenAI o4 model failing on 30% of site selection tasks. Moreover, agentic strategies such as ReAct and Reflexion often suffer from over-reasoning, leading to worse outcomes than direct prompting. With key limitations of LLMs in holistic and non-linear reasoning highlighted, we release LocationReasoner to foster the development of LLMs and agents capable of robust, grounded reasoning in real-world decision-making tasks. Codes and data for our benchmark are available at https://github.com/miho-koda/LocationReasoner.	ICLR ... ICLR 2026 Workshop on Efficient Spatial Reasoning
Hierarchical Chain-of-Thought Prompting: Enhancing LLM Reasoning Performance and Efficiency	2026-03-31	Show Chain-of-Thought (CoT) prompting has significantly improved the reasoning capabilities of large language models (LLMs). However, conventional CoT often relies on unstructured, flat reasoning chains that suffer from redundancy and suboptimal performance. In this work, we introduce Hierarchical Chain-of-Thought (Hi-CoT) prompting, a structured reasoning paradigm specifically designed to address the challenges of complex, multi-step reasoning. Hi-CoT decomposes the reasoning process into hierarchical substeps by alternating between instructional planning and step-by-step execution. This decomposition enables LLMs to better manage long reasoning horizons and maintain logical coherence. Extensive evaluations across diverse LLMs and mathematical reasoning benchmarks show that Hi-CoT consistently improves average accuracy by 6.2% (up to 61.4% on certain models and tasks) while reducing reasoning trace length by 13.9% compared to CoT prompting. We further show that accuracy and efficiency are maximized when models strictly adhere to the hierarchical structure. Our code is available at https://github.com/XingshuaiHuang/Hi-CoT.
Can Commercial LLMs Be Parliamentary Political Companions? Comparing LLM Reasoning Against Romanian Legislative Expuneri de Motive	2026-03-31	Show This paper evaluates whether commercial large language models (LLMs) can function as reliable political advisory tools by comparing their outputs against official legislative reasoning. Using a dataset of 15 Romanian Senate law proposals paired with their official explanatory memoranda (expuneri de motive), we test six LLMs spanning three provider families and multiple capability tiers: GPT-5-mini, GPT-5-chat (OpenAI), Claude Haiku 4.5 (Anthropic), and Llama 4 Maverick, Llama 3.3 70B, and Llama 3.1 8B (Meta). Each model generates predicted rationales evaluated through a dual framework combining LLM-as-Judge semantic scoring and programmatic text similarity metrics. We frame the LLM-politician relationship through principal-agent theory and bounded rationality, conceptualizing the legislator as a principal delegating advisory tasks to a boundedly rational agent under structural information asymmetry. Results reveal a sharp two-tier structure: frontier models (Claude Haiku 4.5, GPT-5-chat, GPT-5-mini) achieve statistically indistinguishable semantic closeness scores above 4.6 out of 5.0, while open-weight models cluster a full tier below (Cohen's d larger than 1.4). However, all models exhibit task-dependent confabulation, performing well on standardized legislative templates (e.g., EU directive transpositions) but generating plausible yet unfounded reasoning for politically idiosyncratic proposals. We introduce the concept of cascading bounded rationality to describe how failures compound across bounded principals, agents, and evaluators, and argue that the operative risk for legislators is not stable ideological bias but contextual ignorance shaped by training data coverage.	12 Figures
Expressive Power of Implicit Models: Rich Equilibria and Test-Time Scaling	2026-03-31	Show Implicit models, an emerging model class, compute outputs by iterating a single parameter block to a fixed point. This architecture realizes an infinite-depth, weight-tied network that trains with constant memory, significantly reducing memory needs for the same level of performance compared to explicit models. While it is empirically known that these compact models can often match or even exceed the accuracy of larger explicit networks by allocating more test-time compute, the underlying mechanism remains poorly understood. We study this gap through a nonparametric analysis of expressive power. We provide a strict mathematical characterization, showing that a simple and regular implicit operator can, through iteration, progressively express more complex mappings. We prove that for a broad class of implicit models, this process lets the model's expressive power scale with test-time compute, ultimately matching a much richer function class. The theory is validated across four domains: image reconstruction, scientific computing, operations research, and LLM reasoning, demonstrating that as test-time iterations increase, the complexity of the learned mapping rises, while the solution quality simultaneously improves and stabilizes.
SemLoc: Structured Grounding of Free-Form LLM Reasoning for Fault Localization	2026-03-31	Show Fault localization identifies program locations responsible for observed failures. Existing techniques rank suspicious code using syntactic spectra--signals derived from execution structure such as statement coverage, control-flow divergence, or dependency reachability. These signals collapse for semantic bugs, where failing and passing executions follow identical code paths and differ only in whether semantic intent is satisfied. Recent LLM-based approaches introduce semantic reasoning but produce stochastic, unverifiable outputs that cannot be systematically cross-referenced across tests or distinguish root causes from cascading effects. We present SemLoc, a fault localization framework based on structured semantic grounding. SemLoc converts free-form LLM reasoning into a closed intermediate representation that binds each inferred property to a typed program anchor, enabling runtime checking and attribution to program structure. It executes instrumented programs to construct a semantic violation spectrum--a constraint-by-test matrix--from which suspiciousness scores are derived analogously to coverage-based methods. A counterfactual verification step further prunes over-approximate constraints and isolates primary causal violations. We evaluate SemLoc on SemFault-250, a corpus of 250 Python programs with single semantic faults. SemLoc outperforms five coverage-, reduction-, and LLM-based baselines, achieving Top-1 accuracy of 42.8% and Top-3 of 68%, while reducing inspection to 7.6% of executable lines. Counterfactual verification provides an additional 12% accuracy gain and identifies primary causal semantic constraints.
The Model Says Walk: How Surface Heuristics Override Implicit Constraints in LLM Reasoning	2026-03-30	Show Large language models systematically fail when a salient surface cue conflicts with an unstated feasibility constraint. We study this through a diagnose-measure-bridge-treat framework. Causal-behavioral analysis of the ``car wash problem'' across six models reveals approximately context-independent sigmoid heuristics: the distance cue exerts 8.7 to 38 times more influence than the goal, and token-level attribution shows patterns more consistent with keyword associations than compositional inference. The Heuristic Override Benchmark (HOB) -- 500 instances spanning 4 heuristic by 5 constraint families with minimal pairs and explicitness gradients -- demonstrates generality across 14 models: under strict evaluation (10/10 correct), no model exceeds 75%, and presence constraints are hardest (44%). A minimal hint (e.g., emphasizing the key object) recovers +15 pp on average, suggesting the failure lies in constraint inference rather than missing knowledge; 12/14 models perform worse when the constraint is removed (up to -39 pp), revealing conservative bias. Parametric probes confirm that the sigmoid pattern generalizes to cost, efficiency, and semantic-similarity heuristics; goal-decomposition prompting recovers +6 to 9 pp by forcing models to enumerate preconditions before answering. Together, these results characterize heuristic override as a systematic reasoning vulnerability and provide a benchmark for measuring progress toward resolving it.
Masked IRL: LLM-Guided Reward Disambiguation from Demonstrations and Language	2026-03-30	Show Robots can adapt to user preferences by learning reward functions from demonstrations, but with limited data, reward models often overfit to spurious correlations and fail to generalize. This happens because demonstrations show robots how to do a task but not what matters for that task, causing the model to focus on irrelevant state details. Natural language can more directly specify what the robot should focus on, and, in principle, disambiguate between many reward functions consistent with the demonstrations. However, existing language-conditioned reward learning methods typically treat instructions as simple conditioning signals, without fully exploiting their potential to resolve ambiguity. Moreover, real instructions are often ambiguous themselves, so naive conditioning is unreliable. Our key insight is that these two input types carry complementary information: demonstrations show how to act, while language specifies what is important. We propose Masked Inverse Reinforcement Learning (Masked IRL), a framework that uses large language models (LLMs) to combine the strengths of both input types. Masked IRL infers state-relevance masks from language instructions and enforces invariance to irrelevant state components. When instructions are ambiguous, it uses LLM reasoning to clarify them in the context of the demonstrations. In simulation and on a real robot, Masked IRL outperforms prior language-conditioned IRL methods by up to 15% while using up to 4.7 times less data, demonstrating improved sample-efficiency, generalization, and robustness to ambiguous language. Project page: https://MIT-CLEAR-Lab.github.io/Masked-IRL and Code: https://github.com/MIT-CLEAR-Lab/Masked-IRL	Accep... Accepted to ICRA 2026
Multi-Agent LLMs for Adaptive Acquisition in Bayesian Optimization	2026-03-30	Show The exploration-exploitation trade-off is central to sequential decision-making and black-box optimization, yet how Large Language Models (LLMs) reason about and manage this trade-off remains poorly understood. Unlike Bayesian Optimization, where exploration and exploitation are explicitly encoded through acquisition functions, LLM-based optimization relies on implicit, prompt-based reasoning over historical evaluations, making search behavior difficult to analyze or control. In this work, we present a metric-level study of LLM-mediated search policy learning, studying how LLMs construct and adapt exploration-exploitation strategies under multiple operational definitions of exploration, including informativeness, diversity, and representativeness. We show that single-agent LLM approaches, which jointly perform strategy selection and candidate generation within a single prompt, suffer from cognitive overload, leading to unstable search dynamics and premature convergence. To address this limitation, we propose a multi-agent framework that decomposes exploration-exploitation control into strategic policy mediation and tactical candidate generation. A strategy agent assigns interpretable weights to multiple search criteria, while a generation agent produces candidates conditioned on the resulting search policy defined as weights. This decomposition renders exploration-exploitation decisions explicit, observable, and adjustable. Empirical results across various continuous optimization benchmarks indicate that separating strategic control from candidate generation substantially improves the effectiveness of LLM-mediated search.	Proce... Proceedings of the IISE Annual Conference & Expo 2026
Distilling LLM Reasoning into Graph of Concept Predictors	2026-03-30	Show Deploying Large Language Models (LLMs) for discriminative workloads is often limited by inference latency, compute, and API costs at scale. Active distillation reduces these costs by querying an LLM oracle to train compact discriminative students, but most pipelines distill only final labels, discarding intermediate reasoning signals and offering limited diagnostics of what reasoning is missing and where errors arise. We propose Graph of Concept Predictors (GCP), a reasoning-aware active distillation framework that externalizes the teacher's decision process as a directed acyclic graph and mirrors it with modular concept predictors in the student. GCP enhances sample efficiency through a graph-aware acquisition strategy that targets uncertainty and disagreement at critical reasoning nodes. Additionally, it improves training stability and efficiency by performing targeted sub-module retraining, which attributes downstream loss to specific concept predictors and updates only the most influential modules. Experiments on eight NLP classification benchmarks demonstrate that GCP enhances performance under limited annotation budgets while yielding more interpretable and controllable training dynamics. Code is available at: https://github.com/Ziyang-Yu/GCP.
Multi-Agent Home Energy Management Assistant	2026-03-30	Show Existing home energy management systems conceptualize occupants as passive recipients of energy information and control, which limits their ability to effectively support informed decision-making and sustained engagement. This paper presents Home Energy Management Assistant (HEMA), the first open-source, multi-agent system enabling sustained human-AI collaboration - multi-turn conversational interactions with preserved context - across diverse home energy management (HEM) tasks - from energy analysis and educational support to smart device control. HEMA combines large language model (LLM) reasoning capabilities with 36 purpose-built domain-specific tools through a three-layer architecture: a web-based conversational interface, a backend API server, and a multi-agent system. The system features three specialized agents - Analysis (energy consumption patterns and cost optimization), Knowledge (educational queries and rebate information), and Control (smart device management and scheduling) - coordinated through a self-consistency classifier that routes user queries using chain-of-thought reasoning. This architecture enables various energy analyses, adaptive explanations, and streamlined device control. HEMA also includes a comprehensive evaluation framework using an LLM-as-simulated-user methodology with 23 objective metrics across task performance, factual accuracy, interaction quality, and system efficiency, allowing systematic testing across diverse scenarios and user personas without requiring extensive human subject testing. Through demonstrations using real-world household energy consumption data, we show how HEMA supports informed decision-making and active engagement in HEM, highlighting its potential as a user-friendly, adaptable tool for residential deployment and as a research platform for HEM innovation.	32 pages, 9 figures
FGTR: Fine-Grained Multi-Table Retrieval via Hierarchical LLM Reasoning	2026-03-30	Show With the rapid advancement of large language models (LLMs), growing efforts have been made on LLM-based table retrieval. However, existing studies typically focus on single-table query, and implement it by similarity matching after encoding the entire table. These methods usually result in low accuracy due to their coarse-grained encoding which incorporates much query-irrelated data, and are also inefficient when dealing with large tables, failing to fully utilize the reasoning capabilities of LLM. Further, multi-table query is under-explored in retrieval tasks. To this end, we propose a hierarchical multi-table query method based on LLM: Fine-Grained Multi-Table Retrieval FGTR, a new retrieval paradigm that employs a human-like reasoning strategy. Through hierarchical reasoning, FGTR first identifies relevant schema elements and then retrieves the corresponding cell contents, ultimately constructing a concise and accurate sub-table that aligns with the given query. To comprehensively evaluate the performance of FGTR, we construct two new benchmark datasets based on Spider and BIRD . Experimental results show that FGTR outperforms previous state-of-the-art methods, improving the F_2 metric by 18% on Spider and 21% on BIRD, demonstrating its effectiveness in enhancing fine-grained retrieval and its potential to improve end-to-end performance on table-based downstream tasks.	work ... work in process;10pages, 5 figures, 4 tables
LLM as an Algorithmist: Enhancing Anomaly Detectors via Programmatic Synthesis	2026-03-30	Show Existing anomaly detection (AD) methods for tabular data usually rely on some assumptions about anomaly patterns, leading to inconsistent performance in real-world scenarios. While Large Language Models (LLMs) show remarkable reasoning capabilities, their direct application to tabular AD is impeded by fundamental challenges, including difficulties in processing heterogeneous data and significant privacy risks. To address these limitations, we propose LLM-DAS, a novel framework that repositions the LLM from a data processor'' to an algorithmist''. Instead of being exposed to raw data, our framework leverages the LLM's ability to reason about algorithms. It analyzes a high-level description of a given detector to understand its intrinsic weaknesses and then generates detector-specific, data-agnostic Python code to synthesize ``hard-to-detect'' anomalies that exploit these vulnerabilities. This generated synthesis program, which is reusable across diverse datasets, is then instantiated to augment training data, systematically enhancing the detector's robustness by transforming the problem into a more discriminative two-class classification task. Extensive experiments on 36 TAD benchmarks show that LLM-DAS consistently boosts the performance of mainstream detectors. By bridging LLM reasoning with classic AD algorithms via programmatic synthesis, LLM-DAS offers a scalable, effective, and privacy-preserving approach to patching the logical blind spots of existing detectors.	Accep... Accepted by the Fourteenth International Conference on Learning Representations (ICLR 2026)
Unveiling Implicit Advantage Symmetry: Why GRPO Struggles with Exploration and Difficulty Adaptation	2026-03-30	Show Reinforcement Learning with Verifiable Rewards (RLVR), particularly GRPO, has become the standard for eliciting LLM reasoning. However, its efficiency in exploration and difficulty adaptation remains an open challenge. In this work, we argue that these bottlenecks stem from an implicit advantage symmetry inherent in Group Relative Advantage Estimation (GRAE). This symmetry induces two critical limitations: (i) at the group level, strict symmetry in weights between correct and incorrect trajectories leaves unsampled action logits unchanged, thereby hindering exploration of novel correct solution. (ii) at the sample level, the algorithm implicitly prioritizes medium-difficulty samples, remaining agnostic to the non-stationary demands of difficulty focus. Through controlled experiments, we reveal that this symmetric property is sub-optimal, yielding two pivotal insights: (i) asymmetrically suppressing the advantages of correct trajectories encourages essential exploration. (ii) learning efficiency is maximized by a curriculum-like transition-prioritizing simpler samples initially before gradually shifting to complex ones. Motivated by these findings, we propose Asymmetric GRAE (A-GRAE), which dynamically modulates exploration incentives and sample-difficulty focus. Experiments across seven benchmarks demonstrate that A-GRAE consistently improves GRPO and its variants across both LLMs and MLLMs.

Chain of Thought

Title	Date	Abstract	Comment
Therefore I am. I Think	2026-04-01	Show We consider the question: when a large language reasoning model makes a choice, did it think first and then decide to, or decide first and then think? In this paper, we present evidence that detectable, early-encoded decisions shape chain-of-thought in reasoning models. Specifically, we show that a simple linear probe successfully decodes tool-calling decisions from pre-generation activations with very high confidence, and in some cases, even before a single reasoning token is produced. Activation steering supports this causally: perturbing the decision direction leads to inflated deliberation, and flips behavior in many examples (between 7 - 79% depending on model and benchmark). We also show through behavioral analysis that, when steering changes the decision, the chain-of-thought process often rationalizes the flip rather than resisting it. Together, these results suggest that reasoning models can encode action choices before they begin to deliberate in text.
Brainstacks: Cross-Domain Cognitive Capabilities via Frozen MoE-LoRA Stacks for Continual LLM Learning	2026-04-01	Show We present Brainstacks, a modular architecture for continual multi-domain fine-tuning of large language models that packages domain expertise as frozen adapter stacks composing additively on a shared frozen base at inference. Five interlocking components: (1) MoE-LoRA with Shazeer-style noisy top-2 routing across all seven transformer projections under QLoRA 4-bit quantization with rsLoRA scaling; (2) an inner loop performing residual boosting by freezing trained stacks and adding new ones; (3) an outer loop training sequential domain-specific stacks with curriculum-ordered dependencies; (4) null-space projection via randomized SVD constraining new stacks to subspaces orthogonal to prior directions, achieving zero forgetting in isolation; (5) an outcome-based sigmoid meta-router trained on empirically discovered domain-combination targets that selectively weights stacks, enabling cross-domain composition. Two boundary experiments: (6) PSN pretraining on a randomly initialized model; (7) per-domain RL (DPO/GRPO) validating compatibility with post-SFT alignment. Validated on TinyLlama-1.1B (4 domains, 9 stacks) and Gemma 3 12B IT (5 domains, 10 stacks), MoE-LoRA achieves 2.5x faster convergence than parameter-matched single LoRA, residual boosting breaks through the single-stack ceiling, and the routed system recovers generation quality destroyed by ungated stack accumulation. The central finding: the outcome-based router discovers that domain stacks encode transferable cognitive primitives (instruction-following clarity, numerical reasoning, procedural logic, chain-of-thought structure) rather than domain-specific knowledge, with medical prompts routing to chat+math stacks in 97% of cases despite zero medical data in those stacks.	26 pa... 26 pages, 13 figures, 4 tables
Automated Framework to Evaluate and Harden LLM System Instructions against Encoding Attacks	2026-04-01	Show System Instructions in Large Language Models (LLMs) are commonly used to enforce safety policies, define agent behavior, and protect sensitive operational context in agentic AI applications. These instructions may contain sensitive information such as API credentials, internal policies, and privileged workflow definitions, making system instruction leakage a critical security risk highlighted in the OWASP Top 10 for LLM Applications. Without incurring the overhead costs of reasoning models, many LLM applications rely on refusal-based instructions that block direct requests for system instructions, implicitly assuming that prohibited information can only be extracted through explicit queries. We introduce an automated evaluation framework that tests whether system instructions remain confidential when extraction requests are re-framed as encoding or structured output tasks. Across four common models and 46 verified system instructions, we observe high attack success rates (> 0.7) for structured serialization where models refuse direct extraction requests but disclose protected content in the requested serialization formats. We further demonstrate a mitigation strategy based on one-shot instruction reshaping using a Chain-of-Thought reasoning model, indicating that even subtle changes in wording and structure of system instructions can significantly reduce attack success rate without requiring model retraining.
View-oriented Conversation Compiler for Agent Trace Analysis	2026-04-01	Show Agent traces carry increasing analytical value in agentic systems and context engineering, yet most prior work treats conversation format as a trivial implementation detail. Modern agent conversations, however, contain deeply structured content, including nested tool calls and results, chain-of-thought reasoning blocks, sub-agent invocations, context-window compaction boundaries, and harness-injected system directives, whose complexity far exceeds that of simple user-assistant exchanges. Feeding such traces to a reflector or other analytical mechanism in plain text, JSON, YAML, or via grep can materially degrade analysis quality. This paper presents VCC (View-oriented Conversation Compiler), a compiler (lex, parse, IR, lower, emit) that transforms raw agent JSONL logs into a family of structured views: a full view (lossless transcript serving as the canonical line-number coordinate system), a user-interface (UI) view (reconstructing the interaction as the user actually perceived it), and an adaptive view (a structure-preserving projection governed by a relevance predicate). In a context-engineering experiment on AppWorld, replacing only the reflector's input format, from raw JSONL to VCC-compiled views, leads to higher pass rates across all three model configurations tested, while cutting reflector token consumption by half to two-thirds and producing more concise learned memory. These results suggest that message format functions as infrastructure for context engineering, not as an incidental implementation choice.	Code:... Code: https://github.com/lllyasviel/VCC
Beyond Symbolic Solving: Multi Chain-of-Thought Voting for Geometric Reasoning in Large Language Models	2026-04-01	Show Geometric Problem Solving (GPS) remains at the heart of enhancing mathematical reasoning in large language models because it requires the combination of diagrammatic understanding, symbolic manipulation and logical inference. In existing literature, researchers have chiefly focused on synchronising the diagram descriptions with text literals and solving the problem. In this vein, they have either taken a neural, symbolic or neuro-symbolic approach. But this solves only the first two of the requirements, namely diagrammatic understanding and symbolic manipulation, while leaving logical inference underdeveloped. The logical inference is often limited to one chain-of-thought (CoT). To address this weakness in hitherto existing models, this paper proposes MARS-GPS, that generates multiple parallel reasoning rollouts augmented with Python code execution for numerical verification, ranks them using token-level entropy as a confidence signal, and aggregates answers through a multi-stage voting and self-verification pipeline. Empirical results show that MARS-GPS with 8 parallel rollouts achieves 88.8% on Geometry3K, a nearly +11% improvement over the prior state-of-the-art, with accuracy scaling consistently as the number of rollouts increases from 1 to 16 (+6.0% on ablation subset). We provide our code and data in an anonymous repository: https://anonymous.4open.science/r/MARS-GPS-DE55.	Under... Under review, 4 figures, 7 tables
Demystifying Chains, Trees, and Graphs of Thoughts	2026-04-01	Show The field of natural language processing (NLP) has witnessed significant progress in recent years, with a notable focus on improving large language models' (LLM) performance through innovative prompting techniques. Among these, prompt engineering coupled with structures has emerged as a promising paradigm, with designs such as Chain-of-Thought, Tree of Thoughts, or Graph of Thoughts, in which the overall LLM reasoning is guided by a structure such as a graph. As illustrated with numerous examples, this paradigm significantly enhances the LLM's capability to solve numerous tasks, ranging from logical or mathematical reasoning to planning or creative writing. To facilitate the understanding of this growing field and pave the way for future developments, we devise a general blueprint for effective and efficient LLM reasoning schemes. For this, we conduct an in-depth analysis of the prompt execution pipeline, clarifying and clearly defining different concepts. We then build the first taxonomy of structure-enhanced LLM reasoning schemes. We focus on identifying fundamental classes of harnessed structures, and we analyze the representations of these structures, algorithms executed with these structures, and many others. We refer to these structures as reasoning topologies, because their representation becomes to a degree spatial, as they are contained within the LLM context. Our study compares existing prompting schemes using the proposed taxonomy, discussing how certain design choices lead to different patterns in performance and cost. We also outline theoretical underpinnings, relationships between prompting and other parts of the LLM ecosystem such as knowledge bases, and the associated research challenges. Our work will help to advance future prompt engineering techniques.
Let the Model Distribute Its Doubt: Confidence Estimation through Verbalized Probability Distribution	2026-04-01	Show Knowing the reliability of a model's response is essential in practical applications. Given the strong generation capabilities of large language models (LLMs), research has focused on generating verbalized confidence. This approach is further enhanced by integrating chain-of-thought reasoning, which provides logical and transparent estimates. However, how reasoning strategies affect the estimated confidence remains under-explored. In this work, we demonstrate that predicting a verbalized probability distribution effectively promotes reasoning for confidence estimation. It requires an LLM to consider all possible answers rather than relying on a single guess, and the requirement of producing a distribution elicits more careful confidence assignment. We conduct systematic experiments comparing different verbalization-based methods across multiple LLMs and tasks. Our method consistently shows advantages, whether in the simple prompting setup or after optimization via reinforcement learning (RL). Notably, it achieves higher reasoning efficacy during inference-time scaling, saving nearly 6$\times$ the computation to reach the best Brier score of the strongest baseline on MMLU-Pro. Additionally, we reveal its limitations on specific tasks and discuss possible solutions for broader applicability.
From Hindsight to Foresight: Self-Encouraged Hindsight Distillation for Knowledge-based Visual Question Answering	2026-04-01	Show Knowledge-based Visual Question Answering (KBVQA) necessitates external knowledge incorporation beyond cross-modal understanding. Existing KBVQA methods either utilize implicit knowledge in multimodal large language models (MLLMs) via in-context learning or explicit knowledge via retrieval augmented generation. However, their reasoning processes remain implicit, without explicit multi-step trajectories from MLLMs. To address this gap, we provide a Hindsight Distilled Reasoning (HinD) framework with Knowledge Encouragement Preference Optimization, aiming at self-encouraging the knowledge reasoning ability inside the MLLM. First, we construct the Hindsight Teacher by prompting the MLLM to complete the reasoning process with knowing the right answer, obtaining Hindsight-Zero training data. Then, the Foresight Student, without knowing the answer, learns the golden trajectories from Hindsight: (1) Hindsight Distillation Fine-Tuning (HDFT) to self-distill the Hindsight-Zero into a modularized Chain-of-Thought (CoT) Generator and a Knowledge Generator for sequential steps and discrete facts generation, respectively; (2) Knowledge Encouragement Preference Optimization (KEPO) to encourage the under-confident but relevant knowledge inside the MLLM and suppress the over-confident but irrelevant one. Experiments on OK-VQA and A-OKVQA validate the effectiveness of HinD, showing that HinD with 7-8B MLLM achieves superior performance without commercial model APIs or retrieved knowledge.
SCPatcher: Automated Smart Contract Code Repair via Retrieval-Augmented Generation and Knowledge Graph	2026-04-01	Show Smart contract vulnerabilities can cause substantial financial losses due to the immutability of code after deployment. While existing tools detect vulnerabilities, they cannot effectively repair them. In this paper, we propose SCPatcher, a framework that combines retrieval-augmented generation with a knowledge graph for automated smart contract repair. We construct a knowledge graph from 5,000 verified Ethereum contracts, extracting function-level relationships to build a semantic network. This graph serves as an external knowledge base that enhances Large Language Model reasoning and enables precise vulnerability patching. We introduce a two-stage repair strategy, initial knowledge-guided repair followed by Chain-of-Thought reasoning for complex vulnerabilities. Evaluated on a diverse set of vulnerable contracts, SCPatcher achieves 81.5% overall repair rate and 91.0% compilation pass rate, substantially outperforming existing methods.	6 pages, 3 figures
Klear-Reasoner: Advancing Reasoning Capability via Gradient-Preserving Clipping Policy Optimization	2026-04-01	Show We present Klear-Reasoner, a model with long reasoning capabilities that demonstrates careful deliberation during problem solving, achieving outstanding performance across multiple benchmarks. Although there are already many excellent works related to inference models in the current community, there are still many problems with reproducing high-performance inference models due to incomplete disclosure of training details. This report provides an in-depth analysis of the reasoning model, covering the entire post-training workflow from data preparation and long Chain-of-Thought supervised fine-tuning (long CoT SFT) to reinforcement learning (RL), along with detailed ablation studies for each experimental component. For SFT data, our experiments show that a small number of high-quality data sources are more effective than a large number of diverse data sources, and that difficult samples can achieve better results without accuracy filtering. In addition, we investigate two key issues with current clipping mechanisms in RL: Clipping suppresses critical exploration signals and ignores suboptimal trajectories. To address these challenges, we propose Gradient-Preserving clipping Policy Optimization (GPPO) that gently backpropagates gradients from clipped tokens. GPPO not only enhances the model's exploration capacity but also improves its efficiency in learning from negative samples. Klear-Reasoner exhibits exceptional reasoning abilities in mathematics and programming, scoring 90.5% on AIME 2024, 83.2% on AIME 2025, 66.0% on LiveCodeBench V5 and 58.1% on LiveCodeBench V6.
Speech LLMs are Contextual Reasoning Transcribers	2026-04-01	Show Despite extensions to speech inputs, effectively leveraging the rich knowledge and contextual understanding of large language models (LLMs) in automatic speech recognition (ASR) remains non-trivial, as the task primarily involves direct speech-to-text mapping. To address this, this paper proposes chain-of-thought ASR (CoT-ASR), which constructs a reasoning chain that enables LLMs to first analyze the input speech and generate contextual analysis, thereby fully exploiting their generative capabilities. With this contextual reasoning, CoT-ASR then performs more informed speech recognition and completes both reasoning and transcription in a single pass. Moreover, CoT-ASR naturally supports user-guided transcription: while designed to self-generate reasoning, it can also seamlessly incorporate user-provided context to guide transcription, further extending ASR functionality. To reduce the modality gap, this paper introduces a CTC-guided Modality Adapter, which uses CTC non-blank token probabilities to weight LLM embeddings, efficiently aligning speech encoder outputs with the LLM's textual latent space. Experiments show that, compared to standard LLM-based ASR, CoT-ASR achieves a relative reduction of 8.7% in word error rate (WER) and 16.9% in entity error rate (EER).
Structured Prompts Improve Evaluation of Language Models	2026-04-01	Show As language models (LMs) are increasingly adopted across domains, high-quality benchmarking frameworks are essential for guiding deployment decisions. In practice, however, frameworks such as Holistic Evaluation of Language Models (HELM) typically evaluate models under a single static prompt configuration, even though model behavior depends strongly on prompt choice. As a result, reported scores can reflect prompt choice as much as model capability. Declarative prompting frameworks such as DSPy offer a scalable way to evaluate models under a set of structured prompting strategies rather than a static prompt configuration. We present a reproducible DSPy+HELM framework for studying how prompt choice impacts reported benchmark outcomes. Using five prompting methods, we evaluate four frontier and two open-source LMs across seven benchmarks against existing HELM baseline scores. By evaluating LMs across a family of prompt configurations, we find that prompt choice can materially impact leaderboard outcomes. In particular, structured prompting improves performance (by 6% on average), alters comparisons (leaderboard rankings shift on 5/7 benchmarks), with most gains coming from introducing chain-of-thought, and little additional benefit from more advanced optimizers. To our knowledge, this is the first study to systematically integrate structured prompting into an established evaluation framework and quantify how prompt choice alone can impact benchmark conclusions. We open-source (i) DSPy+HELM Evaluation (stanford-crfm/helm#3893) and (ii) Prompt Optimization Pipeline (https://github.com/StanfordMIMI/dspy-helm).
Competition and Cooperation of LLM Agents in Games	2026-04-01	Show Large language model (LLM) agents are increasingly deployed in competitive multi-agent settings, raising fundamental questions about whether they converge to equilibria and how their strategic behavior can be characterized. In this paper, we study LLM agent interactions in two standard games: a network resource allocation game and a Cournot competition game. Rather than converging to Nash equilibria, we find that LLM agents tend to cooperate when given multi-round prompts and non-zero-sum context. Chain-of-thought analysis reveals that fairness reasoning is central to this behavior. We propose an analytical framework that captures the dynamics of LLM agent reasoning across rounds and explains these experimental findings.
MemeMind: A Large-Scale Multimodal Dataset with Chain-of-Thought Reasoning for Harmful Meme Detection	2026-04-01	Show As a multimodal medium combining images and text, memes frequently convey implicit harmful content through metaphors and humor, rendering the detection of harmful memes a complex and challenging task. Although recent studies have made progress in detection accuracy and interpretability, large-scale, high-quality datasets for harmful memes remain scarce, and current methods still struggle to capture implicit risks and nuanced semantics. Thus, we construct MemeMind, a large-scale harmful meme dataset. Aligned with the international standards and the context of internet, MemeMind provides detailed Chain-of-Thought (CoT) reasoning annotations to support fine-grained analysis of implicit intentions in memes. Based on this dataset, we further propose MemeGuard, a reasoning-oriented multimodal detection framework that significantly improves both the accuracy of harmful meme detection and the interpretability of model decisions. Extensive experimental results demonstrate that MemeGuard outperforms existing state-of-the-art methods on the MemeMind dataset, establishing a solid foundation for future research in harmful meme detection. The complete dataset and code will be released upon acceptance.
Omni-MMSI: Toward Identity-attributed Social Interaction Understanding	2026-03-31	Show We introduce Omni-MMSI, a new task that requires comprehensive social interaction understanding from raw audio, vision, and speech input. The task involves perceiving identity-attributed social cues (e.g., who is speaking what) and reasoning about the social interaction (e.g., whom the speaker refers to). This task is essential for developing AI assistants that can perceive and respond to human interactions. Unlike prior studies that operate on oracle-preprocessed social cues, Omni-MMSI reflects realistic scenarios where AI assistants must perceive and reason from raw data. However, existing pipelines and multi-modal LLMs perform poorly on Omni-MMSI because they lack reliable identity attribution capabilities, which leads to inaccurate social interaction understanding. To address this challenge, we propose Omni-MMSI-R, a reference-guided pipeline that produces identity-attributed social cues with tools and conducts chain-of-thought social reasoning. To facilitate this pipeline, we construct participant-level reference pairs and curate reasoning annotations on top of the existing datasets. Experiments demonstrate that Omni-MMSI-R outperforms advanced LLMs and counterparts on Omni-MMSI. Project page: https://sampson-lee.github.io/omni-mmsi-project-page.	Accep... Accepted to CVPR 2026. Project page: https://sampson-lee.github.io/omni-mmsi-project-page
Can Large Language Models Self-Correct in Medical Question Answering? An Exploratory Study	2026-03-31	Show Large language models (LLMs) have achieved strong performance on medical question answering (medical QA), and chain-of-thought (CoT) prompting has further improved results by eliciting explicit intermediate reasoning; meanwhile, self-reflective (self-corrective) prompting has been widely claimed to enhance model reliability by prompting LLMs to critique and revise their own reasoning, yet its effectiveness in safety-critical medical settings remains unclear. In this work, we conduct an exploratory analysis of self-reflective reasoning for medical multiple-choice question answering: using GPT-4o and GPT-4o-mini, we compare standard CoT prompting with an iterative self-reflection loop and track how predictions evolve across reflection steps on three widely used medical QA benchmarks (MedQA, HeadQA, and PubMedQA). We analyze whether self-reflection leads to error correction, error persistence, or the introduction of new errors. Our results show that self-reflective prompting does not consistently improve accuracy and its impact is highly dataset- and model-dependent: it yields modest gains on MedQA but provides limited or negative benefits on HeadQA and PubMedQA, and increasing the number of reflection steps does not guarantee better performance. These findings highlight a gap between reasoning transparency and reasoning correctness, suggesting that self-reflective reasoning is better viewed as an analytical tool for understanding model behavior rather than a standalone solution for improving medical QA reliability.
Teaching AI to Handle Exceptions: Supervised Fine-Tuning with Human-Aligned Judgment	2026-03-31	Show Large language models (LLMs), initially developed for generative AI, are now evolving into agentic AI systems, which make decisions in complex, real-world contexts. Unfortunately, while their generative capabilities are well-documented, their decision-making processes remain poorly understood. This is particularly evident when testing targeted decision-making: for instance, how models handle exceptions, a critical and challenging aspect of decision-making made relevant by the inherent incompleteness of contracts. Here we demonstrate that LLMs, even ones that excel at reasoning, deviate significantly from human judgments because they adhere strictly to policies, even when such adherence is impractical, suboptimal, or even counterproductive. We then evaluate three approaches to tuning AI agents to handle exceptions: ethical framework prompting, chain-of-thought reasoning, and supervised fine-tuning. We find that while ethical framework prompting fails and chain-of-thought prompting provides only slight improvements, supervised fine-tuning - specifically with human explanations - yields markedly better results. Surprisingly, in our experiments, supervised fine-tuning even enabled models to generalize human-like decision-making to novel scenarios, demonstrating transfer learning of human-aligned decision-making across contexts. Furthermore, fine-tuning with explanations, not just labels, was critical for alignment, suggesting that aligning LLMs with human judgment requires explicit training on how decisions are made, not just which decisions are made. These findings highlight the need to address LLMs' shortcomings in handling exceptions in order to guide the development of agentic AI toward models that can effectively align with human judgment and simultaneously adapt to novel contexts.
Q-Mask: Query-driven Causal Masks for Text Anchoring in OCR-Oriented Vision-Language Models	2026-03-31	Show Optical Character Recognition (OCR) is increasingly regarded as a foundational capability for modern vision-language models (VLMs), enabling them not only to read text in images but also to support downstream reasoning in real-world visual question answering (VQA). However, practical applications further require reliable text anchors, i.e., accurately grounding queried text to its corresponding spatial region. To systematically evaluate this capability, we introduce TextAnchor-Bench (TABench), a benchmark for fine-grained text-region grounding, which reveals that both general-purpose and OCR-specific VLMs still struggle to establish accurate and stable text anchors. To address this limitation, we propose Q-Mask, a precise OCR framework built upon a causal query-driven mask decoder (CQMD). Inspired by chain-of-thought reasoning, Q-Mask performs causal visual decoding that sequentially generates query-conditioned visual masks before producing the final OCR output. This visual CoT paradigm disentangles where the text is from what the text is, enforcing grounded evidence acquisition prior to recognition and enabling explicit text anchor construction during inference. To train CQMD, we construct TextAnchor-26M, a large-scale dataset of image-text pairs annotated with fine-grained masks corresponding to specific textual elements, encouraging stable text-region correspondences and injecting strong spatial priors into VLM training. Extensive experiments demonstrate that Q-Mask substantially improves text anchoring and understanding across diverse visual scenes.
Hierarchical Chain-of-Thought Prompting: Enhancing LLM Reasoning Performance and Efficiency	2026-03-31	Show Chain-of-Thought (CoT) prompting has significantly improved the reasoning capabilities of large language models (LLMs). However, conventional CoT often relies on unstructured, flat reasoning chains that suffer from redundancy and suboptimal performance. In this work, we introduce Hierarchical Chain-of-Thought (Hi-CoT) prompting, a structured reasoning paradigm specifically designed to address the challenges of complex, multi-step reasoning. Hi-CoT decomposes the reasoning process into hierarchical substeps by alternating between instructional planning and step-by-step execution. This decomposition enables LLMs to better manage long reasoning horizons and maintain logical coherence. Extensive evaluations across diverse LLMs and mathematical reasoning benchmarks show that Hi-CoT consistently improves average accuracy by 6.2% (up to 61.4% on certain models and tasks) while reducing reasoning trace length by 13.9% compared to CoT prompting. We further show that accuracy and efficiency are maximized when models strictly adhere to the hierarchical structure. Our code is available at https://github.com/XingshuaiHuang/Hi-CoT.
Aligned, Orthogonal or In-conflict: When can we safely optimize Chain-of-Thought?	2026-03-31	Show Chain-of-Thought (CoT) monitoring, in which automated systems monitor the CoT of an LLM, is a promising approach for effectively overseeing AI systems. However, the extent to which a model's CoT helps us oversee the model - the monitorability of the CoT - can be affected by training, for instance by the model learning to hide important features of its reasoning. We propose and empirically validate a conceptual framework for predicting when and why this occurs. We model LLM post-training as an RL environment where the reward decomposes into two terms: one term depending on final outputs and another term depending on the CoT. Our framework allows us to classify these two terms as "aligned", "orthogonal", or "in-conflict" before training. We predict that training with in-conflict terms will reduce monitorability, orthogonal terms will not affect it, and aligned terms will improve it. To validate our framework, we use it to classify a set of RL environments, train LLMs within those environments, and evaluate how training affects CoT monitorability. We find that (1) training with "in-conflict" reward terms reduces CoT monitorability and (2) optimizing in-conflict reward terms is difficult.

LLM Interpretability

Title	Date	Abstract	Comment
Discovering Decoupled Functional Modules in Large Language Models	2026-03-18	Show Understanding the internal functional organization of Large Language Models (LLMs) is crucial for improving their trustworthiness and performance. However, how LLMs organize different functions into modules remains highly unexplored. To bridge this gap, we formulate a functional module discovery problem and propose an Unsupervised LLM Cross-layer MOdule Discovery (ULCMOD) framework that simultaneously disentangles the large set of neurons in the entire LLM into modules while discovering the topics of input samples related to these modules. Our framework introduces a novel objective function and an efficient Iterative Decoupling (IterD) algorithm. Extensive experiments show that our method discovers high-quality, disentangled modules that capture more meaningful semantic information and achieve superior performance in various downstream tasks. Moreover, our qualitative analysis reveals that the discovered modules show semantic coherence, correspond to interpretable specializations, and a clear spatial and hierarchical organization within the LLM. Our work provides a novel tool for interpreting the functional modules of LLMs, filling a critical blank in LLM's interpretability research.	AAAI-26 Oral
LaMoGen: Language to Motion Generation Through LLM-Guided Symbolic Inference	2026-03-12	Show Human motion is highly expressive and naturally aligned with language, yet prevailing methods relying heavily on joint text-motion embeddings struggle to synthesize temporally accurate, detailed motions and often lack explainability. To address these limitations, we introduce LabanLite, a motion representation developed by adapting and extending the Labanotation system. Unlike black-box text-motion embeddings, LabanLite encodes each atomic body-part action (e.g., a single left-foot step) as a discrete Laban symbol paired with a textual template. This abstraction decomposes complex motions into interpretable symbol sequences and body-part instructions, establishing a symbolic link between high-level language and low-level motion trajectories. Building on LabanLite, we present LaMoGen, a Text-to-LabanLite-to-Motion Generation framework that enables large language models (LLMs) to compose motion sequences through symbolic reasoning. The LLM interprets motion patterns, relates them to textual descriptions, and recombines symbols into executable plans, producing motions that are both interpretable and linguistically grounded. To support rigorous evaluation, we introduce a Labanotation-based benchmark with structured description-motion pairs and three metrics that jointly measure text-motion alignment across symbolic, temporal, and harmony dimensions. Experiments demonstrate that LaMoGen establishes a new baseline for both interpretability and controllability, outperforming prior methods on our benchmark and two public datasets. These results highlight the advantages of symbolic reasoning and agent-based design for language-driven motion synthesis.	Accep... Accepted by CVPR 2026. Supplementary material included. Project page: https://jjkislele.github.io/LaMoGen/
Pneuma-Seeker: A Relational Reification Mechanism to Align AI Agents with Human Work over Relational Data	2026-03-11	Show When faced with data problems, many data workers cannot articulate their information need precisely enough for software to help. Although LLMs interpret natural-language requests, they behave brittly when intent is under-specified, e.g., hallucinating fields, assuming join paths, or producing ungrounded answers. We present Pneuma-Seeker, a system built around a central idea: relational reification. Pneuma-Seeker represents a user's evolving information need as a relational schema: a concrete, analysis-ready data model shared between user and system. Rather than answering prompts directly, Pneuma-Seeker iteratively refines this schema, then discovers and prepares relevant sources to construct a relation and executable program that compute the answer. Pneuma-Seeker employs an LLM-powered agentic architecture with conductor-style planning and macro- and micro-level context management to operate effectively over heterogeneous relational corpora. We evaluate Pneuma-Seeker across multiple domains against state-of-the-art academic and industrial baselines, demonstrating higher answer accuracy. Deployment in a real organization highlights trust and inspectability as essential requirements for LLM-mediated data systems.
PhysLLM: Harnessing Large Language Models for Cross-Modal Remote Physiological Sensing	2026-03-05	Show Remote photoplethysmography (rPPG) enables non-contact physiological measurement but remains highly susceptible to illumination changes, motion artifacts, and limited temporal modeling. Large Language Models (LLMs) excel at capturing long-range dependencies, offering a potential solution but struggle with the continuous, noise-sensitive nature of rPPG signals due to their text-centric design. To bridge this gap, we introduce the PhysLLM, a collaborative optimization framework that synergizes LLMs with domain-specific rPPG components. Specifically, the Text Prototype Guidance (TPG) strategy is proposed to establish cross-modal alignment by projecting hemodynamic features into LLM-interpretable semantic space, effectively bridging the representational gap between physiological signals and linguistic tokens. Besides, a novel Dual-Domain Stationary (DDS) Algorithm is proposed for resolving signal instability through adaptive time-frequency domain feature re-weighting. Finally, rPPG task-specific cues systematically inject physiological priors through physiological statistics, environmental contextual answering, and task description, leveraging cross-modal learning to integrate both visual and textual information, enabling dynamic adaptation to challenging scenarios like variable illumination and subject movements. Evaluation on four benchmark datasets, PhysLLM achieves state-of-the-art accuracy and robustness, demonstrating superior generalization across lighting variations and motion scenarios. The source code is available at https://github.com/Alex036225/PhysLLM.	Accep... Accepted by International Conference on Learning Representations (ICLR) 2026
Cache What Lasts: Token Retention for Memory-Bounded KV Cache in LLMs	2026-03-01	Show Memory and computation remain core bottlenecks in long-horizon LLM inference due to the quadratic cost of self-attention and the ever-growing key-value (KV) cache. Existing strategies for memory-bounded inference, such as quantization, offloading, or heuristic KV eviction, either incur high orchestration costs or rely on unreliable attention-based proxies of importance. We propose TRIM-KV, a novel approach that learns each token's intrinsic importance at creation time via a lightweight retention gate. Each gate predicts a scalar retention score that decays over time, reflecting the long-term utility of the token for a specific layer and head. Tokens with low scores are evicted when the memory budget is exceeded, ensuring that the cache always contains the most critical tokens. TRIM-KV is trained efficiently through distillation from a frozen LLM combined with a capacity loss, requiring only gate fine-tuning and adding negligible inference overhead. Across mathematical reasoning (GSM8K, MATH-500, AIME24), procedural generation (LongProc), conversational long-memory benchmarks (LongMemEval), and long-context understanding (LongBenchV2 and SCBench), TRIM-KV consistently outperforms strong eviction and learnable retrieval baselines, especially in low-memory regimes. Remarkably, it even surpasses full-cache models in some settings, showing that selective retention can serve as a form of regularization, suppressing noise from uninformative tokens. Qualitative analyses further reveal that learned retention scores align with human intuition, naturally recovering heuristics such as sink tokens, sliding windows, and gist compression without explicit design. Beyond efficiency, retention scores provide insights into layer- and head-specific roles, suggesting a new path toward LLM interpretability.
Sparse Shift Autoencoders for Identifying Concepts from Large Language Model Activations	2026-02-27	Show Unsupervised approaches to large language model (LLM) interpretability, such as sparse autoencoders (SAEs), offer a way to decode LLM activations into interpretable and, ideally, controllable concepts. On the one hand, these approaches alleviate the need for supervision from concept labels, paired prompts, or explicit causal models. On the other hand, without additional assumptions, SAEs are not guaranteed to be identifiable. In practice, they may learn latent dimensions that entangle multiple underlying concepts. If we use these dimensions to extract vectors for steering specific LLM behaviours, this non-identifiability might result in interventions that inadvertently affect unrelated properties. In this paper, we bring the question of identifiability to the forefront of LLM interpretability research. Specifically, we introduce Sparse Shift Autoencoders (SSAEs) which learn sparse representations of differences between embeddings rather than the embeddings themselves. Crucially, we show that SSAEs are identifiable from paired observations which differ in multiple unknown concepts, but not all. With this key identifiability result, we show that we can steer single concepts with only this weak form of supervision. Finally, we empirically demonstrate identifiable concept recovery across multiple real-world language datasets by disentangling activations from different LLMs.	27 pages, 9 figures
Neuro-Symbolic Control with Large Language Models for Language-Guided Spatial Tasks	2026-02-21	Show Although large language models (LLMs) have recently become effective tools for language-conditioned control in embodied systems, instability, slow convergence, and hallucinated actions continue to limit their direct application to continuous control. A modular neuro-symbolic control framework that clearly distinguishes between low-level motion execution and high-level semantic reasoning is proposed in this work. While a lightweight neural delta controller performs bounded, incremental actions in continuous space, a locally deployed LLM interprets symbolic tasks. We assess the suggested method in a planar manipulation setting with spatial relations between objects specified by language. Numerous tasks and local language models, such as Mistral, Phi, and LLaMA-3.2, are used in extensive experiments to compare LLM-only control, neural-only control, and the suggested LLM+DL framework. In comparison to LLM-only baselines, the results show that the neuro-symbolic integration consistently increases both success rate and efficiency, achieving average step reductions exceeding 70% and speedups of up to 8.83x while remaining robust to language model quality. The suggested framework enhances interpretability, stability, and generalization without any need of reinforcement learning or costly rollouts by controlling the LLM to symbolic outputs and allocating uninterpreted execution to a neural controller trained on artificial geometric data. These outputs show empirically that neuro-symbolic decomposition offers a scalable and principled way to integrate language understanding with ongoing control, this approach promotes the creation of dependable and effective language-guided embodied systems.
Stroke Lesions as a Rosetta Stone for Language Model Interpretability	2026-02-03	Show Large language models (LLMs) have achieved remarkable capabilities, yet methods to verify which model components are truly necessary for language function remain limited. Current interpretability approaches rely on internal metrics and lack external validation. Here we present the Brain-LLM Unified Model (BLUM), a framework that leverages lesion-symptom mapping, the gold standard for establishing causal brain-behavior relationships for over a century, as an external reference structure for evaluating LLM perturbation effects. Using data from individuals with chronic post-stroke aphasia (N = 410), we trained symptom-to-lesion models that predict brain damage location from behavioral error profiles, applied systematic perturbations to transformer layers, administered identical clinical assessments to perturbed LLMs and human patients, and projected LLM error profiles into human lesion space. LLM error profiles were sufficiently similar to human error profiles that predicted lesions corresponded to actual lesions in error-matched humans above chance in 67% of picture naming conditions (p < 10^{-23}) and 68.3% of sentence completion conditions (p < 10^{-61}), with semantic-dominant errors mapping onto ventral-stream lesion patterns and phonemic-dominant errors onto dorsal-stream patterns. These findings open a new methodological avenue for LLM interpretability in which clinical neuroscience provides external validation, establishing human lesion-symptom mapping as a reference framework for evaluating artificial language systems and motivating direct investigation of whether behavioral alignment reflects shared computational principles.	45 pages, 17 figures
Uncertainty-Aware Gradient Signal-to-Noise Data Selection for Instruction Tuning	2026-01-20	Show Instruction tuning is a standard paradigm for adapting large language models (LLMs), but modern instruction datasets are large, noisy, and redundant, making full-data fine-tuning costly and often unnecessary. Existing data selection methods either build expensive gradient datastores or assign static scores from a weak proxy, largely ignoring evolving uncertainty, and thus missing a key source of LLM interpretability. We propose GRADFILTERING, an objective-agnostic, uncertainty-aware data selection framework that utilizes a small GPT-2 proxy with a LoRA ensemble and aggregates per-example gradients into a Gradient Signal-to-Noise Ratio (G-SNR) utility. Our method matches or surpasses random subsets and strong baselines in most LLM-as-a-judge evaluations as well as in human assessment. Moreover, GRADFILTERING-selected subsets converge faster than competitive filters under the same compute budget, reflecting the benefit of uncertainty-aware scoring.	Preprint
A Shared Geometry of Difficulty in Multilingual Language Models	2026-01-19	Show Predicting problem-difficulty in large language models (LLMs) refers to estimating how difficult a task is according to the model itself, typically by training linear probes on its internal representations. In this work, we study the multilingual geometry of problem-difficulty in LLMs by training linear probes using the AMC subset of the Easy2Hard benchmark, translated into 21 languages. We found that difficulty-related signals emerge at two distinct stages of the model internals, corresponding to shallow (early-layers) and deep (later-layers) internal representations, that exhibit functionally different behaviors. Probes trained on deep representations achieve high accuracy when evaluated on the same language but exhibit poor cross-lingual generalization. In contrast, probes trained on shallow representations generalize substantially better across languages, despite achieving lower within-language performance. Together, these results suggest that LLMs first form a language-agnostic representation of problem difficulty, which subsequently becomes language-specific. This closely aligns with existing findings in LLM interpretability showing that models tend to operate in an abstract conceptual space before producing language-specific outputs. We demonstrate that this two-stage representational process extends beyond semantic content to high-level meta-cognitive properties such as problem-difficulty estimation.
Causal-SAM-LLM: Large Language Models as Causal Reasoners for Robust Medical Segmentation	2026-01-16	Show The clinical utility of deep learning models for medical image segmentation is severely constrained by their inability to generalize to unseen domains. This failure is often rooted in the models learning spurious correlations between anatomical content and domain-specific imaging styles. To overcome this fundamental challenge, we introduce Causal-SAM-LLM, a novel framework that elevates Large Language Models (LLMs) to the role of causal reasoners. Our framework, built upon a frozen Segment Anything Model (SAM) encoder, incorporates two synergistic innovations. First, Linguistic Adversarial Disentanglement (LAD) employs a Vision-Language Model to generate rich, textual descriptions of confounding image styles. By training the segmentation model's features to be contrastively dissimilar to these style descriptions, it learns a representation robustly purged of non-causal information. Second, Test-Time Causal Intervention (TCI) provides an interactive mechanism where an LLM interprets a clinician's natural language command to modulate the segmentation decoder's features in real-time, enabling targeted error correction. We conduct an extensive empirical evaluation on a composite benchmark from four public datasets (BTCV, CHAOS, AMOS, BraTS), assessing generalization under cross-scanner, cross-modality, and cross-anatomy settings. Causal-SAM-LLM establishes a new state of the art in out-of-distribution (OOD) robustness, improving the average Dice score by up to 6.2 points and reducing the Hausdorff Distance by 15.8 mm over the strongest baseline, all while using less than 9% of the full model's trainable parameters. Our work charts a new course for building robust, efficient, and interactively controllable medical AI systems.	Accep... Accepted by IEEE ICASSP 2026
Self-reflection in Automated Qualitative Coding: Improving Text Annotation through Secondary LLM Critique	2026-01-14	Show Large language models (LLMs) allow for sophisticated qualitative coding of large datasets, but zero- and few-shot classifiers can produce an intolerable number of errors, even with careful, validated prompting. We present a simple, generalizable two-stage workflow: an LLM applies a human-designed, LLM-adapted codebook; a secondary LLM critic performs self-reflection on each positive label by re-reading the source text alongside the first model's rationale and issuing a final decision. We evaluate this approach on six qualitative codes over 3,000 high-content emails from Apache Software Foundation project evaluation discussions. Our human-derived audit of 360 positive annotations (60 passages by six codes) found that the first-line LLM had a false-positive rate of 8% to 54%, despite F1 scores of 0.74 and 1.00 in testing. Subsequent recoding of all stage-one annotations via a second self-reflection stage improved F1 by 0.04 to 0.25, bringing two especially poor performing codes up to 0.69 and 0.79 from 0.52 and 0.55 respectively. Our manual evaluation identified two recurrent error classes: misinterpretation (violations of code definitions) and meta-discussion (debate about a project evaluation criterion mistaken for its use as a decision justification). Code-specific critic clauses addressing observed failure modes were especially effective with testing and refinement, replicating the codebook-adaption process for LLM interpretation in stage-one. We explain how favoring recall in first-line LLM annotation combined with secondary critique delivers precision-first, compute-light control. With human guidance and validation, self-reflection slots into existing LLM-assisted annotation pipelines to reduce noise and potentially salvage unusable classifiers.
Word Synchronization Challenge: A Benchmark for Word Association Responses for Large Language Models	2026-01-14	Show This paper introduces the Word Synchronization Challenge, a novel benchmark to evaluate large language models (LLMs) in Human-Computer Interaction (HCI). This benchmark uses a dynamic game-like framework to test LLMs ability to mimic human cognitive processes through word associations. By simulating complex human interactions, it assesses how LLMs interpret and align with human thought patterns during conversational exchanges, which are essential for effective social partnerships in HCI. Initial findings highlight the influence of model sophistication on performance, offering insights into the models capabilities to engage in meaningful social interactions and adapt behaviors in human-like ways. This research advances the understanding of LLMs potential to replicate or diverge from human cognitive functions, paving the way for more nuanced and empathetic human-machine collaborations.
Can LLMs interpret figurative language as humans do?: surface-level vs representational similarity	2026-01-14	Show Large language models generate judgments that resemble those of humans. Yet the extent to which these models align with human judgments in interpreting figurative and socially grounded language remains uncertain. To investigate this, human participants and four instruction-tuned LLMs of different sizes (GPT-4, Gemma-2-9B, Llama-3.2, and Mistral-7B) rated 240 dialogue-based sentences representing six linguistic traits: conventionality, sarcasm, funny, emotional, idiomacy, and slang. Each of the 240 sentences was paired with 40 interpretive questions, and both humans and LLMs rated these sentences on a 10-point Likert scale. Results indicated that humans and LLMs aligned at the surface level with humans, but diverged significantly at the representational level, especially in interpreting figurative sentences involving idioms and Gen Z slang. GPT-4 most closely approximates human representational patterns, while all models struggle with context-dependent and socio-pragmatic expressions like sarcasm, slang, and idiomacy.	17 pages, 5 figures
Neuro-Symbolic Compliance: Integrating LLMs and SMT Solvers for Automated Financial Legal Analysis	2026-01-07	Show Financial regulations are increasingly complex, hindering automated compliance-especially the maintenance of logical consistency with minimal human oversight. We introduce a Neuro-Symbolic Compliance Framework that integrates Large Language Models (LLMs) with Satisfiability Modulo Theories (SMT) solvers to enable formal verifiability and optimization-based compliance correction. The LLM interprets statutes and enforcement cases to generate SMT constraints, while the solver enforces consistency and computes the minimal factual modification required to restore legality when penalties arise. Unlike transparency-oriented methods, our approach emphasizes logic-driven optimization, delivering verifiable, legally consistent reasoning rather than post-hoc explanation. Evaluated on 87 enforcement cases from Taiwan's Financial Supervisory Commission (FSC), the system attains 86.2% correctness in SMT code generation, improves reasoning efficiency by over 100x, and consistently corrects violations-establishing a preliminary foundation for optimization-based compliance applications.	10 pa... 10 pages, 6 tables, 3 figures, accepted by the 2nd ACM AIware Conference
LLM Interpretability with Identifiable Temporal-Instantaneous Representation	2026-01-02	Show Despite Large Language Models' remarkable capabilities, understanding their internal representations remains challenging. Mechanistic interpretability tools such as sparse autoencoders (SAEs) were developed to extract interpretable features from LLMs but lack temporal dependency modeling, instantaneous relation representation, and more importantly theoretical guarantees, undermining both the theoretical foundations and the practical confidence necessary for subsequent analyses. While causal representation learning (CRL) offers theoretically grounded approaches for uncovering latent concepts, existing methods cannot scale to LLMs' rich conceptual space due to inefficient computation. To bridge the gap, we introduce an identifiable temporal causal representation learning framework specifically designed for LLMs' high-dimensional concept space, capturing both time-delayed and instantaneous causal relations. Our approach provides theoretical guarantees and demonstrates efficacy on synthetic datasets scaled to match real-world complexity. By extending SAE techniques with our temporal causal framework, we successfully discover meaningful concept relationships in LLM activations. Our findings show that modeling both temporal and instantaneous conceptual relationships advances the interpretability of LLMs.	NeurIPS 2025
Trust Me, I Know This Function: Hijacking LLM Static Analysis using Bias	2025-12-18	Show Large Language Models (LLMs) are increasingly trusted to perform automated code review and static analysis at scale, supporting tasks such as vulnerability detection, summarization, and refactoring. In this paper, we identify and exploit a critical vulnerability in LLM-based code analysis: an abstraction bias that causes models to overgeneralize familiar programming patterns and overlook small, meaningful bugs. Adversaries can exploit this blind spot to hijack the control flow of the LLM's interpretation with minimal edits and without affecting actual runtime behavior. We refer to this attack as a Familiar Pattern Attack (FPA). We develop a fully automated, black-box algorithm that discovers and injects FPAs into target code. Our evaluation shows that FPAs are not only effective against basic and reasoning models, but are also transferable across model families (OpenAI, Anthropic, Google), and universal across programming languages (Python, C, Rust, Go). Moreover, FPAs remain effective even when models are explicitly warned about the attack via robust system prompts. Finally, we explore positive, defensive uses of FPAs and discuss their broader implications for the reliability and safety of code-oriented LLMs.
Verification-Guided Context Optimization for Tool Calling via Hierarchical LLMs-as-Editors	2025-12-15	Show Tool calling enables large language models (LLMs) to interact with external environments through tool invocation, providing a practical way to overcome the limitations of pretraining. However, the effectiveness of tool use depends heavily on the quality of the associated documentation and knowledge base context. These materials are usually written for human users and are often misaligned with how LLMs interpret information. This problem is even more pronounced in industrial settings, where hundreds of tools with overlapping functionality create challenges in scalability, variability, and ambiguity. We propose Verification-Guided Context Optimization (VGCO), a framework that uses LLMs as editors to automatically refine tool-related documentation and knowledge base context. VGCO works in two stages. First, Evaluation collects real-world failure cases and identifies mismatches between tools and their context. Second, Optimization performs hierarchical editing through offline learning with structure-aware, in-context optimization. The novelty of our LLM editors has three main aspects. First, they use a hierarchical structure that naturally integrates into the tool-calling workflow. Second, they are state-aware, action-specific, and verification-guided, which constrains the search space and enables efficient, targeted improvements. Third, they enable cost-efficient sub-task specialization, either by prompt engineering large editor models or by post-training smaller editor models. Unlike prior work that emphasizes multi-turn reasoning, VGCO focuses on the single-turn, large-scale tool-calling problem and achieves significant improvements in accuracy, robustness, and generalization across LLMs.	Accep... Accepted by AAAI 2026 Workshop on Agentic AI Benchmarks and Applications for Enterprise Tasks
Visualizing token importance for black-box language models	2025-12-12	Show We consider the problem of auditing black-box large language models (LLMs) to ensure they behave reliably when deployed in production settings, particularly in high-stakes domains such as legal, medical, and regulatory compliance. Existing approaches for LLM auditing often focus on isolated aspects of model behavior, such as detecting specific biases or evaluating fairness. We are interested in a more general question -- can we understand how the outputs of black-box LLMs depend on each input token? There is a critical need to have such tools in real-world applications that rely on inaccessible API endpoints to language models. However, this is a highly non-trivial problem, as LLMs are stochastic functions (i.e. two outputs will be different by chance), while computing prompt-level gradients to approximate input sensitivity is infeasible. To address this, we propose Distribution-Based Sensitivity Analysis (DBSA), a lightweight model-agnostic procedure to evaluate the sensitivity of the output of a language model for each input token, without making any distributional assumptions about the LLM. DBSA is developed as a practical tool for practitioners, enabling quick, plug-and-play visual exploration of LLMs reliance on specific input tokens. Through illustrative examples, we demonstrate how DBSA can enable users to inspect LLM inputs and find sensitivities that may be overlooked by existing LLM interpretability methods.
SpeechIQ: Speech-Agentic Intelligence Quotient Across Cognitive Levels in Voice Understanding by Large Language Models	2025-12-01	Show We introduce Speech-based Intelligence Quotient (SIQ) as a new form of human cognition-inspired evaluation pipeline for voice understanding large language models, LLM Voice, designed to assess their voice understanding ability. Moving beyond popular voice understanding metrics such as word error rate (WER), SIQ examines LLM Voice across three cognitive levels motivated by Bloom's Taxonomy: (1) Remembering (i.e., WER for verbatim accuracy); (2) Understanding (i.e., similarity of LLM's interpretations); and (3) Application (i.e., QA accuracy for simulating downstream tasks). We demonstrate that SIQ not only quantifies voice understanding abilities but also provides unified comparisons between cascaded methods (e.g., ASR LLM) and end-to-end models, identifies annotation errors in existing benchmarks, and detects hallucinations in LLM Voice. Our framework represents a first-of-its-kind intelligence examination that bridges cognitive principles with voice-oriented benchmarks, while exposing overlooked challenges in multi-modal training. Our code and data will be open source to encourage future studies.	ACL 2... ACL 2025 main. Our Speech-IQ leaderboard is hosted at huggingface.co/spaces/nvidia/Speech-IQ-leaderboard. Speech-IQ Calculator: https://github.com/YukinoWan/SpeechIQ

Explainable AI

Title	Date	Abstract	Comment
Explainable AI for Blind and Low-Vision Users: Navigating Trust, Modality, and Interpretability in the Agentic Era	2026-03-31	Show Explainable Artificial Intelligence (XAI) is critical for ensuring trust and accountability, yet its development remains predominantly visual. For blind and low-vision (BLV) users, the lack of accessible explanations creates a fundamental barrier to the independent use of AI-driven assistive technologies. This problem intensifies as AI systems shift from single-query tools into autonomous agents that take multi-step actions and make consequential decisions across extended task horizons, where a single undetected error can propagate irreversibly before any feedback is available. This paper investigates the unique XAI requirements of the BLV community through a comprehensive analysis of user interviews and contemporary research. By examining usage patterns across environmental perception and decision support, we identify a significant modality gap. Empirical evidence suggests that while BLV users highly value conversational explanations, they frequently experience "self-blame" for AI failures. The paper concludes with a research agenda for accessible Explainable AI in agentic systems, advocating for multimodal interfaces, blame-aware explanation design, and participatory development.	Human... Human-centered Explainable AI Workshop (HCXAI) @ CHI 2026, Barcelona, Spain, 2026
From Patterns to Policy: A Scoping Review Based on Bibliometric Analysis (ScoRBA) of Intelligent and Secure Smart Hospital Ecosystems	2026-03-31	Show This study examines the evolution of Intelligent and Secure Smart Hospital Ecosystems using a Scoping Review with Bibliometric Analysis (ScoRBA) to map research patterns, identify gaps, and derive policy implications. Analyzing 891 journal articles from Scopus (2006-2025) through co-occurrence analysis, network visualization, overlay analysis, and the Enhanced Strategic Diagram (ESD), the study applies the PAGER framework to link Patterns, Advances, Gaps, Research directions, and Evidence-based policy implications. Findings reveal three interrelated clusters: AI-driven intelligent healthcare systems, decentralized privacy-preserving digital health ecosystems, and scalable cloud-edge infrastructures, showing a convergence toward integrated ecosystem architectures where intelligence, trust, and infrastructure reinforce each other. Despite progress in AI, blockchain, and cloud computing, gaps remain in interoperability, real-world implementation, governance, and cross-layer integration. Emerging themes such as explainable AI, federated learning, and privacy mechanisms highlight areas needing further research. Policy-relevant recommendations focus on coordinated governance, scalable infrastructure, and secure data ecosystems, particularly for developing country contexts. The study bridges bibliometric evidence with actionable policies, supporting informed decision-making in smart hospital development.	28 pa... 28 pages, 8 figures, 3 tables
Explainable histomorphology-based survival prediction of glioblastoma, IDH-wildtype	2026-03-31	Show Glioblastoma, IDH-wildtype (GBM-IDHwt) is the most common malignant brain tumor. While histomorphology is a crucial component of GBM-IDHwt diagnosis, it is not further considered for prognosis. Here, we present an explainable artificial intelligence (AI) framework to identify and interpret histomorphological features associated with patient survival. The framework combines an explainable multiple instance learning (MIL) architecture that directly identifies prognostically relevant image tiles with a sparse autoencoder (SAE) that maps these tiles to interpretable visual patterns. The MIL model was trained and evaluated on a new real-world dataset of 720 GBM-IDHwt cases from three hospitals and four cancer registries across Germany. The SAE was trained on 1,878 whole-slide images from five independent public glioblastoma collections. Despite the many factors influencing survival time, our method showed some ability to discriminate between patients living less than 180 days or more than 360 days solely based on histomorphology (AUC: 0.67; 95% CI: 0.63-0.72). Cox proportional hazards regression confirmed a significant survival difference between predicted groups after adjustment for established prognostic factors (hazard ratio: 1.47; 95% CI: 1.26-1.72). Three neuropathologists categorized the identified visual patterns into seven distinct histomorphological groups, revealing both established prognostic features and unexpected associations, the latter being potentially attributable to surgery-related confounders. The presented explainable AI framework facilitates prognostic biomarker discovery in GBM-IDHwt and beyond, highlighting promising histomorphological features for further analysis and exposing potential confounders that would be hidden in black-box models.
Interpretable Ensemble Learning for Network Traffic Anomaly Detection: A SHAP-based Explainable AI Framework for Embedded Systems Security	2026-03-30	Show Network security threats in embedded systems pose significant challenges to critical infrastructure protection. This paper presents a comprehensive framework combining ensemble learning methods with explainable artificial intelligence (XAI) techniques for robust anomaly detection in network traffic. We evaluate multiple machine learning models including Random Forest, Gradient Boosting, Support Vector Machines, and ensemble methods on a real-world network traffic dataset containing 19 features derived from packet-level and frequency domain characteristics. Our experimental results demonstrate that ensemble methods achieve superior performance, with Random Forest attaining 90% accuracy and an AUC of 0.617 on validation data. Furthermore, we employ SHAP (SHapley Additive exPlanations) analysis to provide interpretable insights into model predictions, revealing that packet_count_5s,inter_arrival_time, and spectral_entropy are the most influential features for anomaly detection. The integration of XAI techniques enhances model trustworthiness and facilitates deployment in security-critical embedded systems where interpretability is paramount.
Position: Explainable AI is Causality in Disguise	2026-03-30	Show The demand for Explainable AI (XAI) has triggered an explosion of methods, producing a landscape so fragmented that we now rely on surveys of surveys. Yet, fundamental challenges persist: conflicting metrics, failed sanity checks, and unresolved debates over robustness and fairness. The only consensus on how to achieve explainability is a lack of one. This has led many to point to the absence of a ground truth for defining ``the'' correct explanation as the main culprit. This position paper posits that the persistent discord in XAI arises not from an absent ground truth but from a ground truth that exists, albeit as an elusive and challenging target: the causal model that governs the relevant system. By reframing XAI queries about data, models, or decisions as causal inquiries, we prove the necessity and sufficiency of causal models for XAI. We contend that without this causal grounding, XAI remains unmoored. Ultimately, we encourage the community to converge around advanced concept and causal discovery to escape this entrenched uncertainty.
Explainable AI needs formalization	2026-03-30	Show The field of "explainable artificial intelligence" (XAI) seemingly addresses the desire that decisions of machine learning systems should be human-understandable. However, in its current state, XAI itself needs scrutiny. Popular methods cannot reliably answer relevant questions about ML models, their training data, or test inputs, because they systematically attribute importance to input features that are independent of the prediction target. This limits the utility of XAI for diagnosing and correcting data and models, for scientific discovery, and for identifying intervention targets. The fundamental reason for this is that current XAI methods do not address well-defined problems and are not evaluated against targeted criteria of explanation correctness. Researchers should formally define the problems they intend to solve and design methods accordingly. This will lead to diverse use-case-dependent notions of explanation correctness and objective metrics of explanation performance that can be used to validate XAI algorithms.
What-If Explanations Over Time: Counterfactuals for Time Series Classification	2026-03-29	Show Counterfactual explanations emerge as a powerful approach in explainable AI, providing what-if scenarios that reveal how minimal changes to an input time series can alter the model's prediction. This work presents a survey of recent algorithms for counterfactual explanations for time series classification. We review state-of-the-art methods, spanning instance-based nearest-neighbor techniques, pattern-driven algorithms, gradient-based optimization, and generative models. For each, we discuss the underlying methodology, the models and classifiers they target, and the datasets on which they are evaluated. We highlight unique challenges in generating counterfactuals for temporal data, such as maintaining temporal coherence, plausibility, and actionable interpretability, which distinguish the temporal from tabular or image domains. We analyze the strengths and limitations of existing approaches and compare their effectiveness along key dimensions (validity, proximity, sparsity, plausibility, etc.). In addition, we implemented an open-source implementation library, Counterfactual Explanations for Time Series (CFTS), as a reference framework that includes many algorithms and evaluation metrics. We discuss this library's contributions in standardizing evaluation and enabling practical adoption of explainable time series techniques. Finally, based on the literature and identified gaps, we propose future research directions, including improved user-centered design, integration of domain knowledge, and counterfactuals for time series forecasting.	24 pa... 24 pages, 1 figure, 3 tables, accepted at the XAI 2026
Culturally Adaptive Explainable LLM Assessment for Multilingual Information Disorder: A Human-in-the-Loop Approach	2026-03-28	Show Recognizing information disorder is difficult because judgments about manipulation depend on cultural and linguistic context. Yet current Large Language Models (LLMs) often behave as monocultural, English-centric "black boxes," producing fluent rationales that overlook localized framing. Preliminary evidence from the multilingual Information Disorder (InDor) corpus suggests that existing models struggle to explain manipulated news consistently across communities. To address this gap, this ongoing study proposes a Hybrid Intelligence Loop, a human-in-the-loop (HITL) framework that grounds model assessment in human-written rationales from native-speaking annotators. The approach moves beyond static target-language few-shot prompting by pairing English task instructions with dynamically retrieved target-language exemplars drawn from filtered InDor annotations through In-Context Learning (ICL). In the initial pilot, the Exemplar Bank is seeded from these filtered annotations and used to compare static and adaptive prompting on Farsi and Italian news. The study evaluates span and severity prediction, the quality and cultural appropriateness of generated rationales, and model alignment across evaluator groups, providing a testbed for culturally grounded explainable AI.	9 pag... 9 pages, 3 figures, 1 table. Accepted to the Information Disorder Workshop at LREC 2026
XMENTOR: A Rank-Aware Aggregation Approach for Human-Centered Explainable AI in Just-in-Time Software Defect Prediction	2026-03-26	Show Machine learning (ML)-based defect prediction models can improve software quality. However, their opaque reasoning creates an HCI challenge because developers struggle to trust models they cannot interpret. Explainable AI (XAI) methods such as LIME, SHAP, and BreakDown aim to provide transparency, but when used together, they often produce conflicting explanations that increase confusion, frustration, and cognitive load. To address this usability challenge, we introduce XMENTOR, a human-centered, rank-aware aggregation method implemented as a VS Code plugin. XMENTOR unifies multiple post-hoc explanations into a single, coherent view by applying adaptive thresholding, rank and sign agreement, and fallback strategies to preserve clarity without overwhelming users. In a user study, nearly 90% of the participants preferred aggregated explanations, citing reduced confusion and stronger support for daily tasks of debugging and review of defects. Our findings show how combining explanations and embedding them into developer workflows can enhance interpretability, usability, and trust.	10 pa... 10 pages, 14 figures, conference (FORGE '26: 2026 IEEE/ACM Third International Conference on AI Foundation Models and Software Engineering, Rio de Janeiro, Brazil, April 2026)
From What to Why: A Multi-Agent System for Evidence-based Chemical Reaction Condition Reasoning	2026-03-26	Show The chemical reaction recommendation is to select proper reaction condition parameters for chemical reactions, which is pivotal to accelerating chemical science. With the rapid development of large language models (LLMs), there is growing interest in leveraging their reasoning and planning capabilities for reaction condition recommendation. Despite their success, existing methods rarely explain the rationale behind the recommended reaction conditions, limiting their utility in high-stakes scientific workflows. In this work, we propose ChemMAS, a multi-agent system that reframes condition prediction as an evidence-based reasoning task. ChemMAS decomposes the task into mechanistic grounding, multi-channel recall, constraint-aware agentic debate, and rationale aggregation. Each decision is backed by interpretable justifications grounded in chemical knowledge and retrieved precedents. Experiments show that ChemMAS achieves 20-35% gains over domain-specific baselines and outperforms general-purpose LLMs by 10-15% in Top-1 accuracy, while offering falsifiable, human-trustable rationales, which establishes a new paradigm for explainable AI in scientific discovery.	Accep... Accepted by ICLR 2026
On the Vulnerability of Deep Automatic Modulation Classifiers to Explainable Backdoor Threats	2026-03-26	Show Deep learning (DL) has been widely studied for assisting applications of modern wireless communications. One of the applications is automatic modulation classification (AMC). However, DL models are found to be vulnerable to adversarial machine learning (AML) threats. One of the most persistent and stealthy threats is the backdoor (Trojan) attack. Nevertheless, most studied threats originate from other AI domains, such as computer vision (CV). Therefore, in this paper, a physical backdoor attack targeting the wireless signal before transmission is studied. The adversary is considered to be using explainable AI (XAI) to guide the placement of the trigger in the most vulnerable parts of the signal. Then, a class prototype combined with principal components is used to generate the trigger. The studied threat was found to be efficient in breaching multiple DL-based AMC models. The attack achieves high success rates for a wide range of SNR values and a small poisoning ratio.
Does Explanation Correctness Matter? Linking Computational XAI Evaluation to Human Understanding	2026-03-26	Show Explainable AI (XAI) methods are commonly evaluated with functional metrics such as correctness, which computationally estimate how accurately an explanation reflects the model's reasoning. Higher correctness is assumed to produce better human understanding, but this link has not been tested experimentally with controlled levels. We conducted a user study (N=200) that manipulated explanation correctness at four levels (100%, 85%, 70%, 55%) in a time series classification task where participants could not rely on domain knowledge or visual intuition and instead predicted the AI's decisions based on explanations (forward simulation). Correctness affected understanding, but not at every level: performance dropped at 70% and 55% correctness relative to fully correct explanations, while further degradation below 70% produced no additional loss. Rather than shifting performance uniformly, lower correctness decreased the proportion of participants who learned the decision pattern. At the same time, even fully correct explanations did not guarantee understanding, as only a subset of participants achieved high accuracy. Exploratory analyses showed that self-reported ratings correlated with demonstrated performance only when explanations were fully correct and participants had learned the pattern. These findings show that not all differences in functional correctness translate to differences in human understanding, underscoring the need to validate functional metrics against human outcomes.	24 pa... 24 pages, 9 figures, 2 tables
An Explainable Ensemble Learning Framework for Crop Classification with Optimized Feature Pyramids and Deep Networks	2026-03-26	Show Agriculture is increasingly challenged by climate change, soil degradation, and resource depletion, and hence requires advanced data-driven crop classification and recommendation solutions. This work presents an explainable ensemble learning paradigm that fuses optimized feature pyramids, deep networks, self-attention mechanisms, and residual networks for bolstering crop suitability predictions based on soil characteristics (e.g., pH, nitrogen, potassium) and climatic conditions (e.g., temperature, rainfall). With a dataset comprising 3,867 instances and 29 features from the Ethiopian Agricultural Transformation Agency and NASA, the paradigm leverages preprocessing methods such as label encoding, outlier removal using IQR, normalization through StandardScaler, and SMOTE for balancing classes. A range of machine learning models such as Logistic Regression, K-Nearest Neighbors, Support Vector Machines, Decision Trees, Random Forest, Gradient Boosting, and a new Relative Error Support Vector Machine are compared, with hyperparameter tuning through Grid Search and cross-validation. The suggested "Final Ensemble" meta-ensemble design outperforms with 98.80% accuracy, precision, recall, and F1-score, compared to individual models such as K-Nearest Neighbors (95.56% accuracy). Explainable AI methods, such as SHAP and permutation importance, offer actionable insights, highlighting critical features such as soil pH, nitrogen, and zinc. The paradigm addresses the gap between intricate ML models and actionable agricultural decision-making, fostering sustainability and trust in AI-powered recommendations
Dissecting Model Failures in Abdominal Aortic Aneurysm Segmentation through Explainability-Driven Analysis	2026-03-25	Show Computed tomography image segmentation of complex abdominal aortic aneurysms (AAA) often fails because the models assign internal focus to irrelevant structures or do not focus on thin, low-contrast targets. Where the model looks is the primary training signal, and thus we propose an Explainable AI (XAI) guided encoder shaping framework. Our method computes a dense, attribution-based encoder focus map ("XAI field") from the final encoder block and uses it in two complementary ways: (i) we align the predicted probability mass to the XAI field to promote agreement between focus and output; and (ii) we route the field into a lightweight refinement pathway and a confidence prior that modulates logits at inference, suppressing distractors while preserving subtle structures. The objective terms serve only as control signals; the contribution is the integration of attribution guidance into representation and decoding. We evaluate clinically validated challenging cases curated for failure-prone scenarios. Compared to a base SAM setup, our implementation yields substantial improvements. The observed gains suggest that explicitly optimizing encoder focus via XAI guidance is a practical and effective principle for reliable segmentation in complex scenarios.
No Single Metric Tells the Whole Story: A Multi-Dimensional Evaluation Framework for Uncertainty Attributions	2026-03-25	Show Research on explainable AI (XAI) has frequently focused on explaining model predictions. More recently, methods have been proposed to explain prediction uncertainty by attributing it to input features (uncertainty attributions). However, the evaluation of these methods remains inconsistent as studies rely on heterogeneous proxy tasks and metrics, hindering comparability. We address this by aligning uncertainty attributions with the well-established Co-12 framework for XAI evaluation. We propose concrete implementations for the correctness, consistency, continuity, and compactness properties. Additionally, we introduce conveyance, a property tailored to uncertainty attributions that evaluates whether controlled increases in epistemic uncertainty reliably propagate to feature-level attributions. We demonstrate our evaluation framework with eight metrics across combinations of uncertainty quantification and feature attribution methods on tabular and image data. Our experiments show that gradient-based methods consistently outperform perturbation-based approaches in consistency and conveyance, while Monte-Carlo dropconnect outperforms Monte-Carlo dropout in most metrics. Although most metrics rank the methods consistently across samples, inter-method agreement remains low. This suggests no single metric sufficiently evaluates uncertainty attribution quality. The proposed evaluation framework contributes to the body of knowledge by establishing a foundation for systematic comparison and development of uncertainty attribution methods.	Accep... Accepted at the Fourth World Conference on Explainable Artificial Intelligence, xAI 2026, Fortaleza, Brazil, July 1-3, 2026
ShapBPT: Image Feature Attributions Using Data-Aware Binary Partition Trees	2026-03-25	Show Pixel-level feature attributions are an important tool in eXplainable AI for Computer Vision (XCV), providing visual insights into how image features influence model predictions. The Owen formula for hierarchical Shapley values has been widely used to interpret machine learning (ML) models and their learned representations. However, existing hierarchical Shapley approaches do not exploit the multiscale structure of image data, leading to slow convergence and weak alignment with the actual morphological features. Moreover, no prior Shapley method has leveraged data-aware hierarchies for Computer Vision tasks, leaving a gap in model interpretability of structured visual data. To address this, this paper introduces ShapBPT, a novel data-aware XCV method based on the hierarchical Shapley formula. ShapBPT assigns Shapley coefficients to a multiscale hierarchical structure tailored for images, the Binary Partition Tree (BPT). By using this data-aware hierarchical partitioning, ShapBPT ensures that feature attributions align with intrinsic image morphology, effectively prioritizing relevant regions while reducing computational overhead. This advancement connects hierarchical Shapley methods with image data, providing a more efficient and semantically meaningful approach to visual interpretability. Experimental results confirm ShapBPT's effectiveness, demonstrating superior alignment with image structures and improved efficiency over existing XCV methods, and a 20-subject user study confirming that ShapBPT explanations are preferred by humans.	Prese... Presented at AAAI-26 conference and published in Proceedings of the The Fortieth AAAI Conference on Artificial Intelligence (AAAI-26)
Bridging the Dual Nature: How Integrated Explanations Enhance Understanding of Technical Artifacts	2026-03-25	Show Purpose: Understanding a technical artifact requires grasping both its internal structure (Architecture) and its purpose and significance (Relevance), as formalized by Dual Nature Theory. This controlled experimental study investigates whether how explainers address these perspectives affects explainees' understanding. Methods: In a between-subjects experiment, 104 participants received explanations of the board game Quarto! from trained confederates in one of three conditions: Architecture-focused (A), Relevance-focused (R), or Integrated (AR). Understanding was assessed on comprehension (knowing that) and enabledness (knowing how). Results: The A and R conditions produced equivalent understanding despite different explanation content. The AR condition yielded significantly higher enabledness than the focused conditions combined $\mathrm{F}(1, 102) = 4.83$, $p = .030$, $η^2_p = .045$}, while no differences emerged for comprehension. Conclusion: Integrating Architecture and Relevance specifically enhances explainees' ability to apply their understanding in practice, suggesting that fostering agency with technical artifacts requires bridging both perspectives. This has implications for technology education and explainable AI design.
Analyzing animal movement using deep learning	2026-03-25	Show Understanding how animals move through heterogeneous landscapes is central to ecology and conservation. In this context, step selection functions (SSFs) have emerged as the main statistical framework to analyze how biotic and abiotic predictors influence movement paths observed by radio tracking, GPS tags, or similar sensors. A traditional SSF consists of a generalized linear model (GLM) that infers the animal's habitat preferences (selection coefficients) by comparing each observed movement step to random steps. Such GLM-SSFs, however, cannot flexibly consider non-linear or interacting effects, unless those have been specified a priori. To address this problem, generalized additive models have been integrated in the SSF framework, but those GAM-SSFs are still limited in their ability to represent complex habitat preferences and inter-individual variability. Here we explore the utility of deep neural networks (DNNs) to overcome these limitations. We find that DNN-SSFs, coupled with explainable AI to extract selection coefficients, offer many advantages for analyzing movement data. In the case of linear effects, they effectively retrieve the same effect sizes and p-values as conventional GLMs. At the same time, however, they can automatically detect complex interaction effects, nonlinear responses, and inter-individual variability if those are present in the data. We conclude that DNN-SSFs are a promising extension of traditional SSF. Our analysis extends previous research on DNN-SSF by exploring differences and similarities of GLM, GAM and DNN-based SSF models in more depth, in particular regarding the validity of statistical indicators that are derived from the DNN. We also propose new DNN structures to capture inter-individual effects that can be viewed as a nonlinear random effect. All methods used in this paper are available via the 'citoMove' R package.	34 pages, 7 figures
Explainable embeddings with Distance Explainer	2026-03-24	Show While eXplainable AI (XAI) has advanced significantly, few methods address interpretability in embedded vector spaces where dimensions represent complex abstractions. We introduce Distance Explainer, a novel method for generating local, post-hoc explanations of embedded spaces in machine learning models. Our approach adapts saliency-based techniques from RISE to explain the distance between two embedded data points by assigning attribution values through selective masking and distance-ranked mask filtering. We evaluate Distance Explainer on cross-modal embeddings (image-image and image-caption pairs) using established XAI metrics including Faithfulness, Sensitivity/Robustness, and Randomization. Experiments with ImageNet and CLIP models demonstrate that our method effectively identifies features contributing to similarity or dissimilarity between embedded data points while maintaining high robustness and consistency. We also explore how parameter tuning, particularly mask quantity and selection strategy, affects explanation quality. This work addresses a critical gap in XAI research and enhances transparency and trustworthiness in deep learning applications utilizing embedded spaces.	20 pa... 20 pages, 12 figures. Accepted to the 4th World Conference on eXplainable Artificial Intelligence. Method implementation: https://research-software-directory.org/software/distance-explainer
An Explainable AI-Driven Framework for Automated Brain Tumor Segmentation Using an Attention-Enhanced U-Net	2026-03-24	Show Computer-aided segmentation of brain tumors from MRI data is of crucial significance to clinical decision-making in diagnosis, treatment planning, and follow-up disease monitoring. Gliomas, owing to their high malignancy and heterogeneity, represent a very challenging task for accurate and reliable segmentation into intra-tumoral sub-regions. Manual segmentation is typically time-consuming and not reliable, which justifies the need for robust automated techniques.This research resolves this problem by leveraging the BraTS 2020 dataset, where we have labeled MRI scans of glioma patients with four significant classes: background/healthy tissue, necrotic/non-enhancing core, edema, and enhancing tumor. In this work, we present a new segmentation technique based on a U-Net model augmented with executed attention gates to focus on the most significant regions of images. To counter class imbalance, we employ manually designed loss functions like Dice Loss and Categorical Dice Loss, in conjunction with standard categorical cross-entropy. Other evaluation metrics, like sensitivity and specificity, were used to measure discriminability of the model between tumor classes. Besides, we introduce Grad-CAM-based explainable AI to enable visualizing attention regions and improve model interpretability, together with a smooth heatmap generation technique through Gaussian filtering. Our approach achieved superior performance with accuracy of 0.9919, Dice coefficient of 0.9901, mean IoU of 0.9873, sensitivity of 0.9908, and specificity of 0.9974. This study demonstrates that the use of attention mechanisms, personalized loss functions, and explainable AI significantly improves highly complex tumor structure segmentation precision in MRI scans, providing a reliable and explainable method for clinical applications.

Mechanistic Interpretability

Title	Date	Abstract	Comment
Thinking Wrong in Silence: Backdoor Attacks on Continuous Latent Reasoning	2026-04-01	Show A new generation of language models reasons entirely in continuous hidden states, producing no tokens and leaving no audit trail. We show that this silence creates a fundamentally new attack surface. ThoughtSteer perturbs a single embedding vector at the input layer; the model's own multi-pass reasoning amplifies this perturbation into a hijacked latent trajectory that reliably produces the attacker's chosen answer, while remaining structurally invisible to every token-level defense. Across two architectures (Coconut and SimCoT), three reasoning benchmarks, and model scales from 124M to 3B parameters, ThoughtSteer achieves >=99% attack success rate with near-baseline clean accuracy, transfers to held-out benchmarks without retraining (94-100%), evades all five evaluated active defenses, and survives 25 epochs of clean fine-tuning. We trace these results to a unifying mechanism: Neural Collapse in the latent space pulls triggered representations onto a tight geometric attractor, explaining both why defenses fail and why any effective backdoor must leave a linearly separable signature (probe AUC>=0.999). Yet a striking paradox emerges: individual latent vectors still encode the correct answer even as the model outputs the wrong one. The adversarial information is not in any single vector but in the collective trajectory, establishing backdoor perturbations as a new lens for mechanistic interpretability of continuous reasoning. Code and checkpoints are available.
Closing the Confidence-Faithfulness Gap in Large Language Models	2026-04-01	Show Large language models (LLMs) tend to verbalize confidence scores that are largely detached from their actual accuracy, yet the geometric relationship governing this behavior remain poorly understood. In this work, we present a mechanistic interpretability analysis of verbalized confidence, using linear probes and contrastive activation addition (CAA) steering to show that calibration and verbalized confidence signals are encoded linearly but are orthogonal to one another -- a finding consistent across three open-weight models and four datasets. Interestingly, when models are prompted to simultaneously reason through a problem and verbalize a confidence score, the reasoning process disrupts the verbalized confidence direction, exacerbating miscalibration. We term this the "Reasoning Contamination Effect." Leveraging this insight, we introduce a two-stage adaptive steering pipeline that reads the model's internal accuracy estimate and steers verbalized output to match it, substantially improving calibration alignment across all evaluated models.
Spatiotemporal Characterization of Overdose Mortality in Georgia, USA Using Spectral and Nonlinear Interaction Analysis, 2003-2021	2026-04-01	Show Drug overdose mortality in the United States exhibits strong geographic heterogeneity and complex temporal evolution, yet most spatiotemporal studies focus on trends and risks without explicitly characterizing the underlying dynamical structure of overdose trajectories. We develop a nonlinear spectral-spatiotemporal framework to analyze county-level overdose mortality in the state of Georgia from 2003 to 2021. Annual mortality rates are decomposed into low- and high-frequency components to distinguish long-term epidemic pressure from short-term variability, and nonlinear cross-frequency interaction is quantified using bispectral intensity. Counties are grouped into spectral phenotypes using unsupervised clustering, and single-breakpoint change-point models are used to identify regime shifts and quantify post-break acceleration across phenotypes. We find that overdose dynamics across Georgia are dominated by persistent low-frequency growth with limited independent short-term volatility. Nonlinear amplification is spatially concentrated and co-occurs with strong long-term epidemic pressure. Despite synchronous statewide breakpoints around 2014, post-break growth accelerates most sharply in counties exhibiting high low-frequency power and elevated nonlinear interaction. Together, these results provide a mechanistically interpretable framework for identifying dynamical risk phenotypes and structural transitions in spatial overdose epidemics.
Tracking Equivalent Mechanistic Interpretations Across Neural Networks	2026-03-31	Show Mechanistic interpretability (MI) is an emerging framework for interpreting neural networks. Given a task and model, MI aims to discover a succinct algorithmic process, an interpretation, that explains the model's decision process on that task. However, MI is difficult to scale and generalize. This stems in part from two key challenges: there is no precise notion of a valid interpretation; and, generating interpretations is often an ad hoc process. In this paper, we address these challenges by defining and studying the problem of interpretive equivalence: determining whether two different models share a common interpretation, without requiring an explicit description of what that interpretation is. At the core of our approach, we propose and formalize the principle that two interpretations of a model are equivalent if all of their possible implementations are also equivalent. We develop an algorithm to estimate interpretive equivalence and case study its use on Transformer-based models. To analyze our algorithm, we introduce necessary and sufficient conditions for interpretive equivalence based on models' representation similarity. We provide guarantees that simultaneously relate a model's algorithmic interpretations, circuits, and representations. Our framework lays a foundation for the development of more rigorous evaluation methods of MI and automated, generalizable interpretation discovery methods.	32 pa... 32 pages, 5 figures, ICLR 2026
Do Language Models Encode Semantic Relations? Probing and Sparse Feature Analysis	2026-03-31	Show Understanding whether large language models (LLMs) capture structured meaning requires examining how they represent concept relationships. In this work, we study three models of increasing scale: Pythia-70M, GPT-2, and Llama 3.1 8B, focusing on four semantic relations: synonymy, antonymy, hypernymy, and hyponymy. We combine linear probing with mechanistic interpretability techniques, including sparse autoencoders (SAE) and activation patching, to identify where these relations are encoded and how specific features contribute to their representation. Our results reveal a directional asymmetry in hierarchical relations: hypernymy is encoded redundantly and resists suppression, while hyponymy relies on compact features that are more easily disrupted by ablation. More broadly, relation signals are diffuse but exhibit stable profiles: they peak in the mid-layers and are stronger in post-residual/MLP pathways than in attention. Difficulty is consistent across models (antonymy easiest, synonymy hardest). Probe-level causality is capacity-dependent: on Llama 3.1, SAE-guided patching reliably shifts these signals, whereas on smaller models the shifts are weak or unstable. Our results clarify where and how reliably semantic relations are represented inside LLMs, and provide a reproducible framework for relating sparse features to probe-level causal evidence.	accep... accepted at LREC 2026
Interpretable and Steerable Concept Bottleneck Sparse Autoencoders	2026-03-30	Show Sparse autoencoders (SAEs) promise a unified approach for mechanistic interpretability, concept discovery, and model steering in LLMs and LVLMs. However, realizing this potential requires learned features to be both interpretable and steerable. To that end, we introduce two new computationally inexpensive interpretability and steerability metrics for a systematic analysis of LVLM SAEs. This uncovers two observations; (i) a majority of SAE neurons exhibit either low interpretability or low steerability or both, rendering them ineffective for downstream use; and (ii) user-desired concepts are often absent in the SAE, thus limiting their practical utility. To address these limitations, we propose Concept Bottleneck Sparse Autoencoders (CB-SAE) - a novel post-hoc framework that prunes low-utility neurons and augments the latent space with a lightweight concept bottleneck aligned to a user-defined concept set. The resulting CB-SAE improves interpretability by +32.1% and steerability by +14.5% across LVLMs and image generation tasks.	CVPR 2026
Omni-Weather: A Unified Multimodal Model for Weather Radar Understanding and Generation	2026-03-30	Show Weather modeling requires both accurate prediction and mechanistic interpretation, yet existing methods treat these goals in isolation, separating generation from understanding. To address this gap, we present Omni-Weather, the first multimodal foundation model that unifies weather generation and understanding within a single architecture. Omni-Weather integrates a radar encoder for weather generation tasks, followed by unified processing using a shared self-attention mechanism. Moreover, we construct a Chain-of-Thought dataset for causal reasoning in weather generation, enabling interpretable outputs and improved perceptual quality. Extensive experiments show Omni-Weather achieves state-of-the-art performance in both weather generation and understanding. Our findings further indicate that generative and understanding tasks in the weather domain can mutually enhance each other. Omni-Weather also demonstrates the feasibility and value of unifying weather generation and understanding.
Low-Rank Adaptation Reduces Catastrophic Forgetting in Sequential Transformer Encoder Fine-Tuning: Controlled Empirical Evidence and Frozen-Backbone Representation Probes	2026-03-29	Show Sequential fine-tuning of pretrained language encoders often overwrites previously acquired capabilities, but the forgetting behavior of parameter-efficient updates remains under-characterized. We present a controlled empirical study of Low-Rank Adaptation (LoRA) in sequential transformer encoder fine-tuning with companion representation probes that test a frozen-backbone explanation of its robustness. In five full-validation BERT-base reruns on an RTE->MRPC->CoLA->SST-2 sequence, full fine-tuning yields 19.9%+/-4.8% average forgetting, whereas standard LoRA (r=8, query/value modules) yields 0.6%+/-1.4% (paired t-test, p=0.002, Cohen's d_s=3.12). Task-level analyses confirm this reduction is not merely an aggregate effect. Secondary experiments on RoBERTa-base show the same pattern, and the strongest EWC baseline remains at 15.5%+/-1.4% forgetting. A six-task extension reveals that low average forgetting can hide strong task-level heterogeneity. Fine-grained freezing ablations show a marked forgetting drop once frozen parameters exceed roughly 95%, with classifier-only and shallow-adapter baselines approaching LoRA. Companion task-similarity probes in GPT-2 and RoBERTa show the same directional story: frozen-backbone regimes preserve higher inter-task similarity than full fine-tuning, gradual unfreezing weakens stability, and full fine-tuning exhibits its clearest divergence at the final transformer layer. These results support a restrained mechanistic interpretation: LoRA helps largely because backbone freezing preserves a more stable shared feature scaffold. We position standard LoRA as both a strong empirical baseline for sequential encoder adaptation and a useful probe of how selective plasticity shapes interference in transformer continual learning.	14 pa... 14 pages, 11 figures, 4 tables. 234 experiments across BERT-base, RoBERTa-base, GPT-2. Submitted to TMLR
Identifying Connectivity Distributions from Neural Dynamics Using Flows	2026-03-27	Show Connectivity structure shapes neural computation, but inferring this structure from population recordings is degenerate: multiple connectivity structures can generate identical dynamics. Recent work uses low-rank recurrent neural networks (lrRNNs) to infer low-dimensional latent dynamics and connectivity structure from observed activity, enabling a mechanistic interpretation of the dynamics. However, standard approaches for training lrRNNs can recover spurious structures irrelevant to the underlying dynamics. We first characterize the identifiability of connectivity structures in lrRNNs and determine conditions under which a unique solution exists. Then, to find such solutions, we develop an inference framework based on maximum entropy and continuous normalizing flows (CNFs), trained via flow matching. Instead of estimating a single connectivity matrix, our method learns the maximally unbiased distribution over connection weights consistent with observed dynamics. This approach captures complex yet necessary distributions such as heavy-tailed connectivity found in empirical data. We validate our method on synthetic datasets with connectivity structures that generate multistable attractors, limit cycles, and ring attractors, and demonstrate its applicability in recordings from rat frontal cortex during decision-making. Our framework shifts circuit inference from recovering connectivity to identifying which connectivity structures are computationally required, and which are artifacts of underconstrained inference.
Entanglement as Memory: Mechanistic Interpretability of Quantum Language Models	2026-03-27	Show Quantum language models have shown competitive performance on sequential tasks, yet whether trained quantum circuits exploit genuinely quantum resources -- or merely embed classical computation in quantum hardware -- remains unknown. Prior work has evaluated these models through endpoint metrics alone, without examining the memory strategies they actually learn internally. We introduce the first mechanistic interpretability study of quantum language models, combining causal gate ablation, entanglement tracking, and density-matrix interchange interventions on a controlled long-range dependency task. We find that single-qubit models are exactly classically simulable and converge to the same geometric strategy as matched classical baselines, while two-qubit models with entangling gates learn a representationally distinct strategy that encodes context in inter-qubit entanglement -- confirmed by three independent causal tests (p < 0.0001, d = 0.89). On real quantum hardware, only the classical geometric strategy survives device noise; the entanglement strategy degrades to chance. These findings open mechanistic interpretability as a tool for the science of quantum language models and reveal a noise-expressivity tradeoff governing which learned strategies survive deployment.	9 pag... 9 pages, 5 figures, 7 tables
Sparse Auto-Encoders and Holism about Large Language Models	2026-03-27	Show Does Large Language Model (LLM) technology suggest a meta-semantic picture i.e. a picture of how words and complex expressions come to have the meaning that they do? One modest approach explores the assumptions that seem to be built into how LLMs capture the meanings of linguistic expressions as a way of considering their plausibility (Grindrod, 2026a, 2026b). It has previously been argued that LLMs, in employing a form of distributional semantics, adopt a form of holism about meaning (Grindrod, 2023; Grindrod et al., forthcoming). However, recent work in mechanistic interpretability presents a challenge to these arguments. Specifically, the discovery of a vast array of interpretable latent features within the high dimensional spaces used by LLMs potentially challenges the holistic interpretation. In this paper, I will present the original reasons for thinking that LLMs embody a form of holism (section 1), before introducing recent work on features generated through sparse auto-encoders, and explaining how the discovery of such features suggests an alternative decompositional picture of meaning (section 2). I will then respond to this challenge by considering in greater detail the nature of such features (section 3). Finally, I will return to the holistic picture defended by Grindrod et al. and argue that the picture still stands provided that the features are countable (section 4).
Beyond Content Safety: Real-Time Monitoring for Reasoning Vulnerabilities in Large Language Models	2026-03-26	Show Large language models (LLMs) increasingly rely on explicit chain-of-thought (CoT) reasoning to solve complex tasks, yet the safety of the reasoning process itself remains largely unaddressed. Existing work on LLM safety focuses on content safety--detecting harmful, biased, or factually incorrect outputs -- and treats the reasoning chain as an opaque intermediate artifact. We identify reasoning safety as an orthogonal and equally critical security dimension: the requirement that a model's reasoning trajectory be logically consistent, computationally efficient, and resistant to adversarial manipulation. We make three contributions. First, we formally define reasoning safety and introduce a nine-category taxonomy of unsafe reasoning behaviors, covering input parsing errors, reasoning execution errors, and process management errors. Second, we conduct a large-scale prevalence study annotating 4111 reasoning chains from both natural reasoning benchmarks and four adversarial attack methods (reasoning hijacking and denial-of-service), confirming that all nine error types occur in practice and that each attack induces a mechanistically interpretable signature. Third, we propose a Reasoning Safety Monitor: an external LLM-based component that runs in parallel with the target model, inspects each reasoning step in real time via a taxonomy-embedded prompt, and dispatches an interrupt signal upon detecting unsafe behavior. Evaluation on a 450-chain static benchmark shows that our monitor achieves up to 84.88% step-level localization accuracy and 85.37% error-type classification accuracy, outperforming hallucination detectors and process reward model baselines by substantial margins. These results demonstrate that reasoning-level monitoring is both necessary and practically achievable, and establish reasoning safety as a foundational concern for the secure deployment of large reasoning models.
Mechanistically Interpreting Compression in Vision-Language Models	2026-03-26	Show Compressed vision-language models (VLMs) are widely used to reduce memory and compute costs, making them a suitable choice for real-world deployment. However, compressing these models raises concerns about whether internal computations and safety behaviors are preserved. In this work, we use causal circuit analysis and crosscoder-based feature comparisons to examine how pruning and quantization fundamentally change the internals across representative VLMs. We observe that pruning generally keeps circuit structure intact but rotates and attenuates internal features, while quantization modifies the circuits at a higher level yet leaves the surviving features better aligned. Leveraging this insight, we also introduce VLMSafe-420, a novel benchmark that pairs harmful inputs with matched benign counterfactuals across various safety categories. Our findings show that pruning causes a sharp drop in genuine refusal behavior, suggesting that the choice of compression has safety implications.	15 pa... 15 pages, 7 figures, 12 tables
Steering LLMs for Culturally Localized Generation	2026-03-24	Show LLMs are deployed globally, yet produce responses biased towards cultures with abundant training data. Existing cultural localization approaches such as prompting or post-training alignment are black-box, hard to control, and do not reveal whether failures reflect missing knowledge or poor elicitation. In this paper, we address these gaps using mechanistic interpretability to uncover and manipulate cultural representations in LLMs. Leveraging sparse autoencoders, we identify interpretable features that encode culturally salient information and aggregate them into Cultural Embeddings (CuE). We use CuE both to analyze implicit cultural biases under underspecified prompts and to construct white-box steering interventions. Across multiple models, we show that CuE-based steering increases cultural faithfulness and elicits significantly rarer, long-tail cultural concepts than prompting alone. Notably, CuE-based steering is complementary to black-box localization methods, offering gains when applied on top of prompt-augmented inputs. This also suggests that models do benefit from better elicitation strategies, and don't necessarily lack long-tail knowledge representation, though this varies across cultures. Our results provide both diagnostic insight into cultural representations in LLMs and a controllable method to steer towards desired cultures.	preprint
SafeSeek: Universal Attribution of Safety Circuits in Language Models	2026-03-24	Show Mechanistic interpretability reveals that safety-critical behaviors (e.g., alignment, jailbreak, backdoor) in Large Language Models (LLMs) are grounded in specialized functional components. However, existing safety attribution methods struggle with generalization and reliability due to their reliance on heuristic, domain-specific metrics and search algorithms. To address this, we propose \ourmethod, a unified safety interpretability framework that identifies functionally complete safety circuits in LLMs via optimization. Unlike methods focusing on isolated heads or neurons, \ourmethod introduces differentiable binary masks to extract multi-granular circuits through gradient descent on safety datasets, while integrates Safety Circuit Tuning to utilize these sparse circuits for efficient safety fine-tuning. We validate \ourmethod in two key scenarios in LLM safety: \textbf{(1) backdoor attacks}, identifying a backdoor circuit with 0.42% sparsity, whose ablation eradicates the Attack Success Rate (ASR) from 100% $\to$ 0.4% while retaining over 99% general utility; \textbf{(2) safety alignment}, localizing an alignment circuit with 3.03% heads and 0.79% neurons, whose removal spikes ASR from 0.8% $\to$ 96.9%, whereas excluding this circuit during helpfulness fine-tuning maintains 96.5% safety retention.
LLM-Powered Flood Depth Estimation from Social Media Imagery: A Vision-Language Model Framework with Mechanistic Interpretability for Transportation Resilience	2026-03-23	Show Urban flooding poses an escalating threat to transportation network continuity, yet no operational system currently provides real-time, street-level flood depth information at the centimeter resolution required for dynamic routing, electric vehicle (EV) safety, and autonomous vehicle (AV) operations. This study presents FloodLlama, a fine-tuned open-source vision-language model (VLM) for continuous flood depth estimation from single street-level images, supported by a multimodal sensing pipeline using TikTok data. A synthetic dataset of approximately 190000 images was generated, covering seven vehicle types, four weather conditions, and 41 depth levels (0-40 cm at 1 cm resolution). Progressive curriculum training enabled coarse-to-fine learning, while LLaMA 3.2-11B Vision was fine-tuned using QLoRA. Evaluation across 34797 trials reveals a depth-dependent prompt effect: simple prompts perform better for shallow flooding, whereas chain-of-thought (CoT) reasoning improves performance at greater depths. FloodLlama achieves a mean absolute error (MAE) below 0.97 cm and Acc@5cm above 93.7% for deep flooding, exceeding 96.8% for shallow depths. A five-phase mechanistic interpretability framework identifies layer L23 as the critical depth-encoding transition and enables selective fine-tuning that reduces trainable parameters by 76-80% while maintaining accuracy. The Tier 3 configuration achieves 98.62% accuracy on real-world data and shows strong robustness under visual occlusion. A TikTok-based data pipeline, validated on 676 annotated flood frames from Detroit, demonstrates the feasibility of real-time, crowd-sourced flood sensing. The proposed framework provides a scalable, infrastructure-free solution with direct implications for EV safety, AV deployment, and resilient transportation management.	There... There is a update in result, which is needed to be addressed
BioBO: Biology-informed Bayesian Optimization for Perturbation Design	2026-03-23	Show Efficient design of genomic perturbation experiments is crucial for accelerating drug discovery and therapeutic target identification, yet exhaustive perturbation of the human genome remains infeasible due to the vast search space of potential genetic interactions and experimental constraints. Bayesian optimization (BO) has emerged as a powerful framework for selecting informative interventions, but existing approaches often fail to exploit domain-specific biological prior knowledge. We propose Biology-Informed Bayesian Optimization (BioBO), a method that integrates Bayesian optimization with multimodal gene embeddings and enrichment analysis, a widely used tool for gene prioritization in biology, to enhance surrogate modeling and acquisition strategies. BioBO combines biologically grounded priors with acquisition functions in a principled framework, which biases the search toward promising genes while maintaining the ability to explore uncertain regions. Through experiments on established public benchmarks and datasets, we demonstrate that BioBO improves labeling efficiency by 25-40%, and consistently outperforms conventional BO by identifying top-performing perturbations more effectively. Moreover, by incorporating enrichment analysis, BioBO yields pathway-level explanations for selected perturbations, offering mechanistic interpretability that links designs to biologically coherent regulatory circuits.	ICLR 2026
Steering Sparse Autoencoder Latents to Control Dynamic Head Pruning in Vision Transformers (Student Abstract)	2026-03-23	Show Dynamic head pruning in Vision Transformers (ViTs) improves efficiency by removing redundant attention heads, but existing pruning policies are often difficult to interpret and control. In this work, we propose a novel framework by integrating Sparse Autoencoders (SAEs) with dynamic pruning, leveraging their ability to disentangle dense embeddings into interpretable and controllable sparse latents. Specifically, we train an SAE on the final-layer residual embedding of the ViT and amplify the sparse latents with different strategies to alter pruning decisions. Among them, per-class steering reveals compact, class-specific head subsets that preserve accuracy. For example, bowl improves accuracy (76% to 82%) while reducing head usage (0.72 to 0.33) via heads h2 and h5. These results show that sparse latent features enable class-specific control of dynamic pruning, effectively bridging pruning efficiency and mechanistic interpretability in ViTs.	3 pag... 3 pages, 5 figures. Accepted as AAAI 2026 Student Abstract. Includes additional appendix with extended analysis
MCLR: Improving Conditional Modeling in Visual Generative Models via Inter-Class Likelihood-Ratio Maximization and Establishing the Equivalence between Classifier-Free Guidance and Alignment Objectives	2026-03-23	Show Diffusion models have achieved state-of-the-art performance in generative modeling, but their success often relies heavily on classifier-free guidance (CFG), an inference-time heuristic that modifies the sampling trajectory. From a theoretical perspective, diffusion models trained with standard denoising score matching (DSM) are expected to recover the target data distribution, raising the question of why inference-time guidance is necessary in practice. In this work, we ask whether the DSM training objective can be modified in a principled manner such that standard reverse-time sampling, without inference-time guidance, yields effects comparable to CFG. We identify insufficient inter-class separation as a key limitation of standard diffusion models. To address this, we propose MCLR, a principled alignment objective that explicitly maximizes inter-class likelihood-ratios during training. Models fine-tuned with MCLR exhibit CFG-like improvements under standard sampling, achieving comparable qualitative and quantitative gains without requiring inference-time guidance. Beyond empirical benefits, we provide a theoretical result showing that the CFG-guided score is exactly the optimal solution to a weighted MCLR objective. This establishes a formal equivalence between classifier-free guidance and alignment-based objectives, offering a mechanistic interpretation of CFG.
Does Mechanistic Interpretability Transfer Across Data Modalities? A Cross-Domain Causal Circuit Analysis of Variational Autoencoders	2026-03-22	Show Although mechanism-based interpretability has generated an abundance of insight for discriminative network analysis, generative models are less understood -- particularly outside of image-related applications. We investigate how much of the causal circuitry found within image-related variational autoencoders (VAEs) will generalize to tabular data, as VAEs are increasingly used for imputation, anomaly detection, and synthetic data generation. In addition to extending a four-level causal intervention framework to four tabular and one image benchmark across five different VAE architectures (with 75 individual training runs per architecture and three random seed values for each run), this paper introduces three new techniques: posterior-calibration of Causal Effect Strength (CES), path-specific activation patching, and Feature-Group Disentanglement (FGD). The results from our experiments demonstrate that: (i) Tabular VAEs have circuits with modularity that is approximately 50% lower than their image counterparts. (ii) $β$-VAE experiences nearly complete collapse in CES scores when applied to heterogeneous tabular features (0.043 CES score for tabular data compared to 0.133 CES score for images), which can be directly attributed to reconstruction quality degradation (r = -0.886 correlation coefficient between CES and MSE). (iii) CES successfully captures nine of eleven statistically significant architecture differences using Holm--Šidák corrections. (iv) Interventions with high specificity predict the highest downstream AUC values (r = 0.460, p < .001). This study challenges the common assumption that architectural guidance from image-related studies can be transferred to tabular datasets.