Audio visual language models reduce hallucinations by steering question interpretation
Devils in Question Relay: Source-Conditioned Relay Steering to Mitigate Hallucinations in Audio-visual Large Language Models
Computation and Language
Summary
Audio-visual large language models combine sight, sound, and language to answer questions, but sometimes they get confused by signals from a wrong source and give wrong answers. The authors found that this confusion happens because the model mixes up question details with information from an unused sense. They fixed this by gently guiding the model to focus only on the right signals when interpreting questions, without needing to retrain it. This approach helps the models give more accurate responses in tests involving images and sounds.
What this means in practice
- •For multimedia application developers: Improve the accuracy of apps that process combined audio and visual data by reducing source-confused errors in language models.
- •For voice assistant teams: Enhance multimodal voice assistants so their answers rely correctly on spoken or visual inputs without confusing one for the other.
Authors
Yu Zhang, Pingrui Zhang, Xuefeng Bai, Pengfei Zhang, Yang Xiang, Kehai Chen
Abstract
Audio-visual large language models (AVLLMs) have made remarkable progress in multimodal understanding and reasoning through interactions among visual, auditory, and linguistic information. However, recent studies show that AVLLMs face a critical challenge: $\textbf{source-confused grounding hallucination}$, where cues from the unused modality induce responses that the required modality does not support, undermining reliability in real-world applications. Existing methods have made progress in mitigating this failure, yet how it arises from internal cross-modal interactions remains insufficiently understood. To address this gap, we conduct path-intervention and representation analyses, revealing a $\textbf{question-relay}$ mechanism: question states carry interfering cues alongside required-source evidence, undermining grounding in required-modality evidence. Cutting pathways from interfering modality to question states yields greater correct-answer logit recovery than cutting those to the generation position. Motivated by these findings, we propose $\textbf{SECRET}$ ($\textbf{S}$ourc$\textbf{E}$-$\textbf{C}$onditioned $\textbf{RE}$lay s$\textbf{T}$eering), a training-free method that mitigates cross-modal interference at the question relay. Using contrasting question representations elicited through different modality-pathway interventions, SECRET steers the original question states toward required-source evidence. Experiments on two widely adopted benchmarks CMM and AVHBench across three AVLLMs show that SECRET consistently outperforms prior training-free methods, substantially mitigating source-confused grounding hallucinations (e.g., up to +18.0 and +7.1 percentage points over base models). Modality-specific captioning further demonstrates its generalizability to open-ended generation.