Tracing Audio Grounding and Answer Selection in Audio LLMs
Hyebin Cho, Suho Yoo, Jihoo Jung, Joon Son Chung
cs.CL · cs.AI · cs.LG · cs.SD
Quick verdict
Claims fine-tuning moves Audio LLMs' audio use from early-layer grounding to late-layer answer selection; solid causal evidence but two Qwen models, LoRA-only.
Summary
The paper asks how fine-tuning on audio-dependent multiple-choice QA changes the way Audio LLMs use acoustic evidence internally. The authors fine-tune Qwen2-Audio and Qwen2.5-Omni with LoRA (r=16) on the AudioMCQ-StrongAC-GeminiCoT training set (19,480 samples), then compare zero-shot vs fine-tuned checkpoints on four unseen benchmarks: ADQA-Bench, MMAU, MMAR, and MMSU. Behaviorally, they define an acoustic-sensitivity score from accuracy drops under silence and audio-mismatch perturbations; accuracy gains (+10.9 to +14.3 points on ADQA/MMAR/MMSU) co-occur with positive sensitivity gains, while MMAU shows negative transfer for both models. Internally, attention knockout shows that blocking audio→options attention hurts most in early-to-middle layers (≈8–16) in both zero-shot and fine-tuned models, whereas blocking options→answer attention becomes important in middle-to-late layers (≈18–27) only after training. LoRA-band ablations localize the most necessary updates to layers 16–23 (Qwen2-Audio) and 21–27 (Qwen2.5-Omni); band-restricted training reveals that Qwen2-Audio can acquire most gains from layers 8–15 alone, while Qwen2.5-Omni needs updates across multiple bands. Main limitations: only two Qwen-family models, LoRA-only adaptation from a single training configuration (seeds not reported), knockout and ablation results lack confidence intervals, and the claimed mechanistic account remains coarse layer-band localization rather than circuit-level analysis.
Review complete
Assessment at a glance
4
risk flags identified
● high 0
● medium 2
● low 2
5 strengths · 0 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Fine-tuning on audio-dependent MCQ data increases behavioral sensitivity to acoustic evidence on exactly the benchmarks where accuracy improves, with MMAU a negative-transfer exception (Tab. 1).
- Audio information shapes answer-option representations most strongly in early-to-middle layers (≈8–16), both before and after training (Fig. 2).
- Training's main internal change is strengthening option→answer readout in middle-to-late layers (≈18–27), not the earlier audio-option interaction (Fig. 2).
- Learned LoRA updates are functionally concentrated in layers 16–23 (Qwen2-Audio) / 21–27 (Qwen2.5-Omni), yet for Qwen2-Audio layers 8–15 alone suffice for most gains (Fig. 4, Tab. 2).
Strengths
Multi-level causal triangulation
Behavioral perturbation (Saudio, Eq. 1–2), causal attention knockout (Eq. 3–4, Audio↛Options vs Options↛Answer), and weight-level LoRA band ablation (Sec. 2.4) converge on one consistent story across Secs. 3.1–3.3.
Honest negative-transfer case
MMAU (ΔA −13.5±3.2 for Qwen2.5-Omni, ΔSaudio −7.1±3.1) is retained and used diagnostically in Fig. 4 (disabling some bands improves MMAU accuracy) instead of being dropped.
Necessity vs sufficiency dissociation
Statistical care on the behavioral headline
Probability-based intervention metric
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
This four-page ICASSP-style study asks where and how fine-tuning changes the internal use of acoustic evidence in Audio LLMs. It advances three claims: (1) models trained on audio-dependent multiple-choice QA degrade substantially more when the audio is silenced or replaced (Sec. 3.1, Tab. 1); (2) audio shapes answer-option representations in early-to-middle layers both before and after training, while training specifically strengthens the option→answer readout in middle-to-late layers (Sec. 3.2, Fig. 2); (3) the most functionally important learned LoRA updates concentrate in specific layer bands — 16–23 for Qwen2-Audio and 21–27 for Qwen2.5-Omni (Sec. 3.3, Fig. 4). Scope is deliberately narrow: two Qwen-family models, LoRA r=16 adaptation on one curated training set (AudioMCQ-StrongAC-GeminiCoT, 19,480 samples, CoT discarded), evaluated on four held-out MCQ benchmarks (ADQA-Bench, MMAU-test-mini, MMAR, MMSU).
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 0 weaknesses, analyzed in detail
- 4 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is the main finding of this paper?
Fine-tuning Audio LLMs (Qwen2-Audio, Qwen2.5-Omni) on audio-dependent multiple-choice QA mainly strengthens how audio-informed answer-option representations feed the final prediction in middle-to-late layers (≈18–27), while audio-to-option grounding in early-to-middle layers (≈8–16) already exists before training.
Which layers should I fine-tune in Qwen2-Audio for audio QA?
The paper reports that training LoRA only on layers 8–15 nearly matches all-layer LoRA for Qwen2-Audio on ADQA-Bench, MMAR and MMSU (e.g., 51.13 vs 51.13 on ADQA), while layers 16–23 are the most functionally necessary band in the fully trained model. This is from one training set and one seed and needs replication.
What datasets and models were used?
Training: AudioMCQ-StrongAC-GeminiCoT (19,480 samples, DCASE 2026 Task 5, chain-of-thought discarded). Evaluation: ADQA-Bench (2,998), MMAU-test-mini (1,000), MMAR (1,000), MMSU (5,000). Models: Qwen2-Audio and Qwen2.5-Omni, adapted with LoRA r=16.
Is code available for this paper?
The paper text does not state a code release. All components are public, though: the two models, the four evaluation benchmarks, the DCASE 2026 training set, and standard LoRA and attention-knockout techniques.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures 2–4 are line plots; the text gives their qualitative layer ranges (e.g., 8–12, 18–22) and Table 2 values, but exact per-layer probability values cannot be extracted from the text.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.