# Compositional Failure in Audio-Visual LLMs: Late-Layer Prior Dominance Under Cross-modal Conflict

arXiv: https://arxiv.org/abs/2608.27785 · HTML: https://mlsanity.com/paper/2608.27785/

## Key facts

- No citing papers in the map yet (data through 14 Oct 2026).
- Builds on 7 of its 8 references that are in the map.
- First posted on arXiv 27 Aug 2026.

## Builds on (most cited first, 7 of 7)

- [VideoLLaMA 2: Advancing Spatial-Temporal Modeling and Audio Understanding in Video-LLMs](https://mlsanity.com/paper/2406.07476/) (2024)
- [AVHBench: A Cross-Modal Hallucination Benchmark for Audio-Visual Large Language Models](https://mlsanity.com/paper/2410.18325/) (2024)
- [The Curse of Multi-Modalities: Evaluating Hallucinations of Large Multimodal Models across Language, Visual, and Audio](https://mlsanity.com/paper/2410.12787/) (2024)
- [AVCD: Mitigating Hallucinations in Audio-Visual Large Language Models through Contrastive Decoding](https://mlsanity.com/paper/2505.20862/) (2025)
- [Aligned Better, Listen Better for Audio-Visual Large Language Models](https://mlsanity.com/paper/2504.02061/) (2025)
- [Aurelia: Test-time Reasoning Distillation in Audio-Visual LLMs](https://mlsanity.com/paper/2503.23219/) (2025)
- [Fork-Merge Decoding: Enhancing Multimodal Understanding in Audio-Visual Large Language Models](https://mlsanity.com/paper/2505.20873/) (2025)

## Abstract (arXiv)

We study audio-visual conflict as a compositional generalization test for AV-LLMs: the model must combine synchronized but semantically incompatible audio and video evidence and decide whether the pair matches. On VideoLLaMA 2-7B-AV, three alignment configurations remain nearchance on the scored exact-string Yes/No subset of AVHBench, even though their output priors shift substantially. Similarly, off-the-shelf InternVideo2 experienced a 32.3% accuracy decrease specifically under cross-modal conflict, accompanied by a 17.3% instruction-following failure. We call this failure mode prior dominance: late-layer commitment to an internally preferred answer pattern that is weakly grounded in the conflicting inputs. To explain this behavior, we conduct a mechanistic interpretability analysis and find that commitment remains concentrated at 25.5 ± 1 layers. We show that stronger temporal alignment changes answer bias, but do not improve compositional conflict resolution. Code and data to reproduce our mechanistic audit and behavioral evaluations are available at https://github.com/AdarshSudheer09/AVHBench-dmai.