CPMT: Cross-Path Mutual Teaching for Audio-Visual Online Test-Time Adaptation

This repository is now organized around the CPMT paper line.

Current Review Entry Point

If you are reviewing the current KS32 / VGG method line, start here first:

• docs/KS32_METHOD_HANDOFF_2026-03-22.md
• docs/GPT_PRO_STRICT_METHOD_REVIEW_PROMPT.md

Important:

• EXPERIMENT_HISTORY.md is a full archive, not the best first-stop summary.
• Some early KS32 interpretations in the archive were later corrected by targeted diagnostics.

The active line is:

• paper-facing method name: CPMT
• final retained configuration: CPMT-Lite (Anchor Teacher + Conformal + CompLoss + ModGate + REM)
• backbone/runtime: CAV-MAE with TSA-compatible adaptors
• main experiment assets:
  • results/paper_experiments/
  • results/overnight_ablation/
  • results/cpmt_lite_followup/
  • results/cpmt_lite_shuffle_ablation/
  • results/table6_completion/

Older archived paper assets, protocol-alignment packs, reviewer packs, and exploratory branches are archived under past/. The active runtime now lives under runtime/ rather than under an abandoned method name.

Active Layout

• runtime/
  • neutral runtime and baseline code used by the current experiments
• external/
  • local third-party dependencies such as TSA and AVRobustBench
• paper/
  • current paper working directory
• results/paper_experiments/
  • main CPMT paper runs and summary tables
• results/overnight_ablation/
  • full-to-lite ablations and multiseed stability runs
• results/cpmt_lite_followup/
  • corrected shuffled CPMT-Lite multiseed and teacher decomposition
• results/cpmt_lite_shuffle_ablation/
  • corrected shuffled five-component ablation
• results/table6_completion/
  • active-pipeline READ/SAR completion for the missing Table 6 baselines
• results/submission_readiness/
  • current benchmark-asset and runtime-surface audits for the next submission-strengthening tasks
• scripts/
  • current entrypoints only
• tools/
  • current CPMT and AV2C runners only
• past/
  • archived non-mainline assets

Current Entry Points

Setup:

bash scripts/setup_external_tsa.sh <TSA_REPO_URL> [TSA_REF]
bash scripts/setup_external_avrobustbench.sh <AVROBUSTBENCH_REPO_URL> [REF]

Run the main paper pack:

bash scripts/run_paper_experiments.sh

Run the overnight ablation pack:

bash scripts/run_overnight_ablation_multiseed.sh

Run the missing Table 6 completion pack:

bash scripts/run_table6_completion.sh

Progress helpers:

bash scripts/check_paper_progress.sh
bash scripts/check_overnight_progress.sh
bash scripts/check_table6_completion.sh

Audit the next-step experiment surface:

python3 scripts/audit_submission_readiness.py

Download the official KINETICS-2C clean asset pack from Hugging Face:

python3 scripts/download_avrobust_kinetics_clean.py

Current Result Snapshot

From results/paper_experiments/tables/:

• VGGSound-2C:
  • Source: 41.35
  • READ: 39.08
  • SAR: 30.88
  • AV2C: 47.45
  • CPMT-REM (full): 48.82
  • CPMT-Lite: 49.27
• KS50:
  • Source: 61.85
  • SAR: 63.39
  • READ: 65.87
  • AV2C: 61.87
  • CPMT-REM (full): 65.06
  • CPMT-Lite: 65.01

From the corrected shuffled follow-up:

• CPMT-Lite multiseed:
  • VGGSound-2C: 49.29 +- 0.02
  • KS50: 65.17 +- 0.13
• baseline multiseed follow-up:
  • VGG READ: 39.31 +- 0.23
  • VGG SAR: 30.82 +- 0.08
  • KS50 READ: 65.76 +- 0.11
  • KS50 SAR: 63.25 +- 0.26
• teacher decomposition:
  • source-anchor-only: 48.02
  • cross-path-only: 49.31
  • both: 49.31

Archive Policy

past/ is intentional. It keeps previous method lines and exploratory assets available without letting them clutter the current repository surface.

If you are working on the active paper line, start from the root-level scripts/, tools/, results/, and paper/ directories and ignore past/ unless you need historical context.