Source-linked AI summary
Compositional Failure in Audio-Visual LLMs: Late-Layer Prior Dominance Under Cross-modal Conflict
Adarsh Sudheer, David Li, Omar Elbanna, Ishaan Kodarapu, Arjun Bahuguna, Vasu Sharma
TL;DR
This paper asks whether audio-visual models compose evidence when synchronized modalities conflict. It reframes conflict as a compositional learning problem, audits model behavior and late-layer commitment, and finds that stronger alignment changes answer bias without improving conflict resolution.
Problem
The paper asks whether audio-visual models compare conflicting audio and video evidence rather than defaulting to a high-probability prior.
Method
The paper evaluates conflict behavior and uses logit-lens analysis, attention-head ablations, and activation patching to investigate prior dominance and late-layer commitment.
Results
Stronger alignment changes answer bias without improving conflict-set accuracy, while InternVideo2 accuracy drops from 60.1% on full AVHBench to 27.8% on the conflict set.
Takeaways & Limitations
Conflict composition should be evaluated directly, with alignment improvements paired with tests and controls for late-layer commitment in safety-relevant settings.
Takeaways & Limitations
Mechanistic evidence is limited to the VideoLLaMA 2 architecture, and Table 1 is restricted to exact Yes/No completions.
Abstract
from arXiv · showhide
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 $\pm$ 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.
1. Introduction
The paper frames cross-modal conflict as a test of whether AV-LLMs compose audio and video evidence rather than rely on a unimodal shortcut or preferred answer prior. Across alignment interventions, answer bias changes without resolving compositional conflict.
- Cross-modal conflict tests whether models compare semantically incompatible audio and video streams instead of using shallow single-modality co-occurrence.
- Agreement examples may be solved by one modality, whereas conflict examples expose reliance on a high-probability prior.
- ACTC, TATI, and AMD change answer bias but do not improve compositional conflict resolution.
- Prior dominance describes final answers increasingly explained by an internally preferred response pattern rather than composed auditory and visual evidence.
2. Conflict-Based Evaluation and Notation
The evaluation defines compositional success as judging compatibility after extracting and binding evidence from both streams. A logit-lens probe tracks layer-wise predictions to identify when the answer stabilizes into a prior.
- A conflict example has individually coherent audio and video streams that are semantically incompatible as a pair.
- The logit lens projects residual-stream states through the model’s output mapping to track predicted-token probabilities across layers.
- The snap layer is the earliest layer after which the top prediction remains unchanged through the final layer, serving as a probe of prior commitment.
3. Methodology
The methodology combines adversarial conflict evaluation, alignment ablations, behavioral measurement, and mechanistic audits of VideoLLaMA2-7B. It examines both modality-integration choices and the layers and heads associated with answer behavior.
- Evaluation and models: InternVideo2 and VideoLLaMA2-7B-AV are evaluated to measure behavioral collapse and alignment effects under cross-modal conflict.
- Evaluation and models: The adversarial conflict split contains 1,281 filtered examples with severe semantic contradictions, while final pipeline evaluation uses the balanced full AVHBench dataset.
- Evaluation and models: 92% of the curated conflict split has No ground truths, creating a methodological trap because rejection bias can produce illusory accuracy gains.
- Alignment pipeline: ACTC concatenates audio and visual tokens, TATI synchronously interleaves corresponding tokens, and AMD stochastically drops one modality during training.
- Alignment pipeline: TATI adds shared temporal embeddings so each visual frame is processed alongside its corresponding audio token.
- Mechanistic audit: The logit lens tracks P(Yes), P(No), and predicted tokens across all layers to locate the snap layer where the top prediction stabilizes.
- Mechanistic audit: The mechanistic audit ablates attention heads in layers 14–27 using 100 conflict samples and defines hallucination or conflict-resolution heads by Δ thresholds.
- Mechanistic audit: The audit identifies 21 conflict-resolution heads clustered in layers 15–18, with no heads passing the hallucination threshold.
4. Results & Discussion
Across AVHBench conflict evaluations, alignment interventions shifted answer priors without improving compositional accuracy. Layer-wise and open-ended analyses support prior dominance: internally preferred outputs override conflicting multimodal evidence.
- 4.1. Baseline Compositional Failure Under Conflict: 32.3% accuracy decrease on InternVideo2 under cross-modal conflict coincided with a 17.3% instruction failure rate.Accuracy fell from 60.1% on full AVHBench to 27.8% on the conflict set; 1,108 of 6,408 inferences violated the Yes/No format.
- 4.2. Bias shifts without compositional gains: All three alignment configurations remained near chance on exact-string AVHBench results and failed to exceed the VideoLLaMA 2-7B-AV base accuracy of 51.7%.The table separately shows substantial shifts in Yes/No output priors despite stable accuracy.
- 4.2. Bias shifts without compositional gains: The full pipeline produced 3,446 No versus 1,853 Yes predictions, indicating a strong No bias without better conflict resolution.The authors associate this pattern with failure in binding and compatibility judgment rather than learning to compare modalities.
- 4.3. Prior dominance appears in both captions and layer-wise traces: 25.5 ± 1 layers marked the snap point across configurations and the baseline, with no significant target-token probability mass before layer 18.Conflict-resolution heads cluster between layers 15 and 18, while later commitment increasingly favors No regardless of ground truth.
- 4.3. Prior dominance appears in both captions and layer-wise traces: Logit-lens analysis found that alignment fine-tuning changes the direction of the prior rather than its timing, while corrected audio-visual survival remains identical across configurations.The reported mechanism is a surface-level bias shift without structural change in deep semantic routing.
5. Limitations
The mechanistic evidence is architecture-limited, and the evaluation does not fully separate temporal alignment from sequence-length effects.
- 5. Limitations: The mechanistic analysis is limited to VideoLLaMA 2, although behavioral baseline evaluation includes InternVideo2.Table 1 is also restricted to exact Yes/No completions, making it a strict probe.
- 5. Limitations: The evaluation does not isolate increased sequence length from temporal alignment in the current TATI implementation.The interleaving process produces longer sequences than the ACTC baseline.
6. Conclusion
Neither off-the-shelf InternVideo2 nor aligned VideoLLaMA 2-7B-AV shows strong compositional generalization under cross-modal conflict. The findings identify late-layer prior commitment as a practical obstacle and motivate direct conflict evaluation with interventions targeting that commitment.
- Neither off-the-shelf InternVideo2 nor aligned VideoLLaMA 2-7B-AV shows strong compositional generalization under cross-modal conflict.
- Stronger alignment does not improve VideoLLaMA 2-7B-AV conflict resolution, instead mainly reshaping answer bias.
- Late-layer commitment to an internally preferred response pattern is identified as a practical obstacle across plain and aligned AVLLMs.
- Future work should evaluate conflict composition directly and test or control late-layer commitment in safety-relevant settings.
A. Appendix
The appendix presents cosine similarity between late-layer audio and visual token states as an exploratory representation diagnostic. Because residual states already mix information, the proxy is suggestive rather than an exact measure of modality attribution and is not used for the core claim.
- Figure 4 reports an exploratory cosine-similarity proxy s(ℓ) for audio-visual representations across layers.
- The proxy measures cosine similarity between late-layer audio and visual token states after removing the shared timestamp embedding.
- Because deep residual states mix information through attention and MLP updates, timestamp subtraction is not an exact decomposition of modality content.
- The paper includes this plot as suggestive context and does not rely on the proxy for its core claim.
B. Training Setup
The training setup fine-tunes a frozen VideoLLaMA 2-7B-AV backbone with LoRA while jointly training custom vision and audio projectors. Supervised fine-tuning uses eight A100 GPUs and batch size 64.
- All three pipeline stages use Low-Rank Adaptation on top of the frozen VideoLLaMA 2-7B-AV backbone.
- Supervised fine-tuning is conducted on 8 A100 GPUs with a batch size of 64.
- Custom vision and audio projectors are trained alongside LoRA adapters to map modality-specific features into the shared language-model embedding space.
C. Expanded Related Works
Prior work emphasizes grounded audio-visual reasoning and alignment through temporal interleaving and synchronization. This paper complements that work by asking whether such alignment improves composition under contradiction.
- AVHBench, AVCD, and broader multimodal hallucination evaluations establish the importance of testing grounded audio-visual reasoning.
- Temporal interleaving and synchronization are widely used design choices in AV-LLMs.
- The paper complements alignment-module research by testing whether alignment pressures improve composition specifically under contradiction.
D. Discussion & Future Mitigation Strategies
The mechanistic audit finds a temporal gap between conflict detection and prior commitment, but this observation is correlational and requires causal testing.
- The observed separation between conflict-detection at layers 15–18 and prior commitment at layer 25.5 is currently correlational.The authors propose causal interventions, including steering vectors or residual-stream dampening at layers 20–24, to distinguish dedicated prior dominance from downstream information propagation.
- An operational threshold of Δ = ±0.01 motivates future statistical calibration using bootstrapping or randomized ablation baselines.
- Testing structurally distinct architectures is needed to assess whether the snap layer generalizes across autoregressive multimodal systems.