AI
Devils in Question Relay: Source-Conditioned Relay Steering to Mitigate Hallucinations in Audio-visual Large Language Models
Publication: Hugging Face Daily Papers
Author: Not available in collected metadata
Source published: 28/09/2026, 20:00:00 UTC
Briefing created: 03/10/2026, 03:07:53 UTC
Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 02/10/2026, 08:18:25 UTC.
Briefings
Audio-visual large language models (AVLLMs) struggle with source-confused grounding hallucinations where unused modality cues interfere with required modalities. This study introduces SECRET, a training-free method using question-relay steering to mitigate cross-modal interference. Experiments on CMM and AVHBench benchmarks across three AVLLMs show SECRET outperforms prior methods, reducing hallucinations by up to 18.0 percentage points on CMM and 7.1 percentage points on AVHBench. The method generalizes to open-ended generation via modality-specific captioning.
Key points
- SECRET is a training-free method that steers question states toward required-source evidence using contrasting representations elicited through different modality-pathway interventions.
- Cutting pathways from interfering modality to question states yields greater correct-answer logit recovery than cutting those to the generation position.
Context and limitations
Results are limited to two benchmarks (CMM and AVHBench) and three specific AVLLMs; generalizability to other models or modalities remains unverified.
Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 03/10/2026, 03:07:53 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.
View archived briefing on IPFS
bafybeiavwvzwu4ewwsniwz27twmvjy5hd2ycmqcnoy2m3kmm4ycf3mkp5y