SECRET: training-free relay steering cuts cross-modal hallucinations in AVLLMs by up to 18 points
Yu Zhang · hf · 2026-10-02
This paper dissects source-confused grounding hallucinations in audio-visual LLMs via path-intervention analysis, revealing a question-relay mechanism where interfering modality cues contaminate question states. It proposes SECRET, a training-free steering method that outperforms prior training-free baselines on CMM and AVHBench across three AVLLMs, with gains up to +18.0 percentage points.
More from Research
- AMap open-sources ABot-Recon: streaming 3D reconstruction from video with a 12-frame local context — rsasaki0109 · 2026-10-02
- Neuralink pretrains decoders on 50,000+ hours of neural data—dataset may be the real moat — CurieuxExplorer · 2026-10-02
- CyberGym cybersecurity benchmark effectively saturated on verified task subset — aryaman2020 · 2026-10-02
- SemEval-2027 Task 9 calls for teams on 4-language multimodal news framing analysis — preslav_nakov · 2026-10-02
- Alternating prompt and model upgrades lift science agent from 42% to 73% — rohanpaul_ai · 2026-10-02
- ISMIR paper teaches a transformer to play in 12 jazz piano legends' styles — umpedronosapato · 2026-10-02