Source-linked AI summary
Tracing Audio Grounding and Answer Selection in Audio LLMs
Hyebin Cho, Suho Yoo, Jihoo Jung, Joon Son Chung
TL;DR
Audio LLMs may answer from textual cues rather than acoustic evidence, and it is unclear how training changes internal audio use. The paper combines behavioral perturbations with layer-wise interventions across multiple-choice AudioQA benchmarks to trace these changes. Training increases sensitivity to the provided audio, with acoustic information shaping answer choices earlier and influencing final prediction later, while learned effects concentrate in model-specific layer bands.
Problem
Audio LLMs can achieve high AudioQA performance with little or no acoustic evidence, leaving how training changes internal audio use unclear.
Method
The study analyzes multiple-choice AudioQA across benchmarks using audio perturbations, attention interventions, and layer-band analyses of LoRA updates.
Results
Training increases acoustic-evidence use: audio replacement causes larger accuracy drops, acoustic information shapes answer choices earlier, and training strengthens their later influence on final prediction.
Takeaways & Limitations
Audio integration and final answer selection are distinct stages, with training mainly strengthening the later use of audio-informed representations and learned effects concentrated in model-specific LoRA bands.
Abstract
from arXiv · showhide
Audio Large Language Models (Audio LLMs) have advanced in audio understanding, yet they can still predict the answer by reasoning from textual cues or linguistic priors rather than the provided audio. A common remedy is to train models on data whose answers cannot be inferred from text alone. This approach can improve performance, but what changes within the model remains unclear. In this paper, we ask what must happen inside the model for the audio to actually determine the answer. Our findings are threefold. (1) Replacing the audio with silence or unrelated audio causes substantially larger performance degradation in the trained model than in the pretrained model. (2) Acoustic information most strongly shapes the model's representations of the answer choices in early-to-middle layers, while training mainly increases the influence of audio information on the final prediction in middle-to-late layers. (3) The weights learned during training have their largest impact in specific layer bands. Together, these results provide a mechanistic account of how training strengthens the use of acoustic evidence in Audio LLMs.
1. INTRODUCTION
Audio LLMs can achieve high AudioQA accuracy without relying on acoustic evidence, motivating training and mechanistic analysis of how audio supports final answers. The paper studies where training increases audio use and finds distinct early representation and later answer-selection stages.
- High AudioQA performance can persist with little or no acoustic evidence, exposing a gap between benchmark accuracy and actual audio use.
- Training data that require acoustic evidence address this limitation, but how training changes internal audio use remains unclear.
- The study asks whether training increases audio use, where acoustic information shapes answer-choice representations, and where learned changes are concentrated.
- Training increases acoustic-evidence use: replacing audio with silence or unrelated audio causes larger accuracy drops after training on benchmarks where training improves performance.
- Acoustic information shapes answer-choice representations in early-to-middle layers, while training strengthens their influence on final prediction in middle-to-late layers.
- The results provide a mechanistic account of how training strengthens acoustic evidence use by distinguishing audio integration from contribution to final answer selection.
2. METHOD
The method combines cross-benchmark multiple-choice AudioQA evaluation with audio perturbations, attention knockouts, and layer-band analyses of LoRA updates. These analyses measure behavioral audio sensitivity and identify where audio-informed representations and learned changes affect prediction.
- Analysis setup: Multiple-choice AudioQA examples contain audio, a question, candidate answers, and a correct option label, with models instructed to generate only the option letter.
- Behavioral sensitivity: Models are evaluated with original, silent, and mismatched audio conditions to test sensitivity to removing or replacing the paired acoustic evidence.
- Behavioral sensitivity: Behavioral acoustic-evidence sensitivity averages original accuracy minus silent and mismatched accuracies, and positive post-training change indicates increased sensitivity.
- Layer-wise interventions: Attention knockout blocks selected source-to-target connections within layer windows while keeping model parameters unchanged, measuring changes in probability assigned to the correct option.
- LoRA layer analysis: The study compares disabling one LoRA layer band at a time with training adapters restricted to single bands and full-depth LoRA, including a rank-matched r=4 baseline.
3. EXPERIMENTS
Training increases reliance on acoustic evidence when performance improves, primarily by strengthening later answer-selection pathways rather than early audio-to-option interactions. Learned LoRA effects are concentrated in model-specific layer bands, and restricted training reveals differences in how the two models distribute useful updates.
- 3.1. Behavioral Use of Acoustic Evidence: On all six improving model–dataset pairs, ∆Saudio is positive, whereas both models show negative ∆Saudio on MMAU.The results associate successful performance transfer with greater sensitivity to removing or replacing the corresponding audio.
- 3.2. From Acoustic Evidence to Answer Prediction: Acoustic information influences candidate-answer representations in early-to-middle layers before training mainly strengthens their later contribution to final prediction.Audio↛All Options is strongest around layers 8–12 or 10–16, while All Options↛Answer becomes prominent around layers 18–27 after training.
- 3.2. From Acoustic Evidence to Answer Prediction: Qwen2-Audio already distinguishes supportive correct-option information from competing wrong-option information, while Qwen2.5-Omni gains a stronger correct-option contribution after training.Blocking the correct option lowers correct-answer probability; blocking wrong options raises it in Qwen2-Audio, whereas wrong-option effects remain limited in Qwen2.5-Omni.
- 3.3. Where Learned LoRA Updates Matter: Removing layers 16–23 causes the largest performance drop in Qwen2-Audio, while removing layers 21–27 has the largest effect in Qwen2.5-Omni.These bands overlap with layers where Options↛Answer becomes more important after training; some MMAU bands instead improve accuracy when removed.
- 3.3. Where Learned LoRA Updates Matter: Training only layers 8–15 nearly matches or exceeds all-layer r=16 LoRA for Qwen2-Audio, but no single band recovers all-layer performance for Qwen2.5-Omni.The result suggests restricted layer ranges can capture much of Qwen2-Audio’s improvement, whereas Qwen2.5-Omni benefits from updates across multiple ranges.
- 3.3. Where Learned LoRA Updates Matter: The layer band most necessary after full-model training is not necessarily the band that performs best when trained alone.Band ablation measures reliance on learned updates, whereas single-band training measures what can be learned under restricted updates; the bands can work together.
4. CONCLUSION
The paper distinguishes two stages in acoustic grounding: acoustic evidence shapes candidate-answer representations early, while training strengthens their use for final answer selection later. Learned effects are concentrated in model-specific LoRA bands.
- 4. CONCLUSION: Acoustic evidence affects candidate-answer representations in early-to-middle layers, while training mainly strengthens their contribution to the final answer in middle-to-late layers.The conclusion treats making acoustic evidence available to answer representations and using those representations for the final decision as distinct stages.