{"id":"2379b1cb-dfed-4fcb-94d4-0ebb44275730","arxiv_id":"2507.20169","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"SI-SDA uses attention-saliency prompt-reliance as an unsupervised reward to reinforce a model's own beam hypotheses, improving audio LLM WER and BLEU across ASR, S2TT, and SQA without labeled data.","lead":"The paper proposes SI-SDA, a method that lets an audio language model improve itself on a new speech domain using only unlabeled audio: it scores its own beam-search outputs with an attention-saliency metric and uses reinforcement learning to reinforce the best-scoring hypotheses. If correct, it reduces the need for labeled target-domain data when deploying audio LLMs.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The load-bearing gap is that Q(Ŷ) in Eqs. (7)–(8) is either undefined or circular: Eq. (2) needs a loss target that unlabeled data cannot provide, and if the hypothesis itself is used as target, the motivating R_E>R_C evidence from Fig. 1 does not transfer.","rationale":"The reader's weakest_assumption is the validity of Q(Ŷ) as an unsupervised reward, and my concern is the same one, sharpened. The paper's only empirical support for the reward is Fig. 1/Sec. III-B, where correct and erroneous tokens are identified with ground-truth labels and the loss is computed against those labels. In the actual SI-SDA pipeline no labels exist, so either Eq. (2) is undefined or the loss is computed against the model's own generated hypothesis. Those two computations are not the same quantity; the R_E>R_C correlation is a property of gradients with respect to ground-truth targets and need not survive when targets are the hypothesis itself. This is why no amount of additional tables can establish the central claim unless the loss target is specified and the unsupervised Q is shown to rank hypotheses by WER/BLEU. The proposed correlation test is the minimal check that would settle this. I do not see a reason to move beyond the reader's CONDITIONAL verdict: the issue is addressable and the surrounding results could be correct, but as written the mechanism is unverified. I agree with the reader rather than adding a separate concern; other issues (ties in Tables I/III, missing error bars) are secondary to this one.","tokens_in":8832,"tokens_out":7123,"duration_ms":77130,"concrete_test":"On a held-out ASR dev set (CHiME-4 or LS-Freesound-babble), generate N-best lists with the same beam size and LoRA setup as §IV. For every hypothesis, compute Q(Ŷ) using Eqs. (7)–(8) with L in Eq. (2) set to cross-entropy of the hypothesis against itself (the only label-free target consistent with the text). Measure Spearman rank correlation between Q and the hypothesis's actual WER, and compare the WER of the lowest-Q hypothesis against the mean N-best WER. Repeat across the target domains and, for S2TT, with BLEU instead of WER. If the rank correlation is not consistently positive, R_E>R_C computed with ground-truth targets is not a valid basis for the unsupervised reward. An independent re-derivation of Eq. (8) from the authors' released code would also settle which L was actually used.","verdict_should_be":"UNCHANGED","load_bearing_attack":"SI-SDA's central claim is that Q(Ŷ) ranks hypotheses by quality without labels. But Eq. (2) computes saliency from ∂L(x)/∂A_{h,l}, and the paper never specifies L(x) when no ground-truth transcript exists. The natural label-free choice is cross-entropy with the hypothesis itself as target; then every position's target equals the token the model already generated, so Q measures self-agreement rather than fidelity to any external text. The motivating evidence in Sec. III-B and Fig. 1, however, uses ground-truth tokens to define C and E, and for erroneous tokens the loss is evaluated against the correct token, not the model's output. That is exactly the setting in which R_E>R_C is observed; it does not imply the same separation when gradients are taken with respect to the model's own output tokens. If the authors intend some other label-free loss, it is not stated, so Eqs. (7)–(8) cannot be implemented from the paper. This is a correctness risk, not a stylistic omission: if Q is uncorrelated with true hypothesis quality, Eq. (10) will reinforce or suppress hypotheses on a signal unrelated to WER/BLEU, and the gains in Tables I–III lack an identified mechanism. The method should be re-evaluated with a specified L and a direct validation of the unsupervised Q ordering.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes SI-SDA, an unsupervised domain adaptation method for audio large language models. The method computes saliency scores from attention matrices and loss gradients, defines a sequence-level prompt-reliance score Q(Ŷ) that is intended to indicate hypothesis quality without labels, and then uses Q as a reward in a policy-gradient-style objective over a beam set, reinforcing low-Q hypotheses. Experiments on Qwen2-Audio cover ASR (CHiME-4, LS-Freesound, CommonVoice, Switchboard, TED-3), speech-to-text translation (CoVoST2), and spoken question answering (LibriSQA), reporting WER and BLEU improvements over zero-shot and several unsupervised baselines, with up to 20.8% relative WER reduction on TED-3.","tokens_in":9075,"tokens_out":4100,"duration_ms":41916,"significance":"If the central claim holds, the method is practically valuable: it is source-free, uses only unlabeled target-domain speech, requires no recall of source data, and reports consistent improvements across several tasks and datasets. The paper has notable strengths: it validates the method against external metrics (WER, BLEU) rather than only against its own reward, introduces no fitted constants, and demonstrates data efficiency. However, the main claimed mechanism depends on an unsupervised saliency signal whose computation is not fully specified, and the empirical support lacks statistical rigor. The contribution is interesting and within scope for a speech/audio venue, but the central derivation needs clarification and the significance claims need stronger evidence before publication.","major_comments":[{"comment":"The saliency score in Eq. (2) is defined via ∂L(x)/∂A_{h,l}, but the paper never specifies what L(x) is when the input is unlabeled. Since the method explicitly avoids ground-truth transcripts, the loss used to compute the saliency map is undefined, so Eqs. (7)-(8) cannot be implemented as written. If the intended choice is cross-entropy with the model's own hypothesis as the target, then the saliency measures self-agreement rather than alignment with any external text, and the R_E > R_C evidence in Fig. 1 does not transfer directly because that analysis is computed against ground-truth tokens. Please state the exact loss used in the unsupervised setting and provide a direct validation that the resulting Q ordering correlates with WER/BLEU on held-out unlabeled data.","section":"III-A, Eq. (2)"},{"comment":"The claim that SI-SDA 'consistently and significantly improves' performance is not supported by the reported numbers, which lack error bars, confidence intervals, significance tests, and multiple seeds. Several differences are very small, e.g., CommonVoice Africa WER 4.9 vs 5.0, LS-Freesound car 2.5 vs 2.4 (STAR), and Table III it-en BLEU 36.1 vs 36.1. Please report means over at least three runs or a paired significance test across utterances, and temper the abstract's 'significantly' unless such evidence is provided.","section":"Tables I-III and abstract"},{"comment":"The optimization loop is under-specified: since Q and the saliency maps depend on the model parameters θ, the paper should state whether the beam search, Q computation, and policy-gradient update are performed once per epoch, per mini-batch, or with re-sampling at each step, and whether gradients flow through the saliency computation. This is needed for reproducibility and for assessing whether the observed gains come from the reward signal or from a specific implementation detail.","section":"III-D, Eq. (10)"},{"comment":"The motivating evidence that erroneous tokens rely more heavily on prompt tokens (R_E > R_C) is presented for only two ASR test sets and with no numerical values or error bars; the caption refers to 'clean and noise ASR datasets' while the text mentions 'LibriSpeech test set LS-Freesound.' Since this analysis is the sole justification for using Q as a reward in S2TT and SQA, the paper should provide a broader token-level analysis or, at minimum, an utterance-level correlation between Q and external quality scores for each task family.","section":"V-A and Fig. 1"}],"minor_comments":[{"comment":"The notation X = {x1, x2, ..., x3} appears to be a typo; it should be x_L or similar.","section":"III-A"},{"comment":"The model is called Qwen-Audio2 in the text but the reference [25] is the Qwen2-Audio technical report; please align the naming.","section":"IV-A"},{"comment":"The dataset name 'LirbriSQA' is a typo; it should be LibriSQA.","section":"V-C"},{"comment":"The caption contains the typo 'RESUTLS'; also, for the it-en row, SI-SDA ties with STAR, which is worth noting in the text.","section":"Table III"},{"comment":"The description of Figure 1 is confusing: the text says 'both clean LibriSpeech test set LS-Freesound' but the figure and caption suggest a comparison across multiple datasets; please clarify what is plotted and on which data the normalization is performed.","section":"V-A"}],"recommendation":"major_revision","confidential_remarks":"The paper is likely within scope for Interspeech or a similar audio/speech venue, and the empirical setting is interesting. The main risk is that Eq. (2) lacks a well-defined unsupervised loss, making the core reward signal unimplementable as written; asking for that specification and a direct correlation analysis is essential. I do not see a novelty-disclosure or scope problem. If the implementation details and statistical evidence are added, the paper could become acceptable, but the current version is not yet sufficiently precise."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nHere's the short version: this paper has a genuinely new idea — using attention-saliency-derived prompt reliance as an unsupervised reward for RL-based self-improvement of audio LLMs — and the broad experimental sweep across ASR, S2TT, and SQA gives the results some real plausibility. But the paper has a load-bearing under-specification in how it computes the saliency score when there are no labels, and without that the mechanism isn't fully established.\n\nWhat's good: The combination of saliency as a quality score with policy-gradient optimization is new relative to the STAR, DPO, and self-training baselines they compare against. The gains are consistent across tasks (up to ~20% WERR on TED-3), and the data efficiency — a few hours of unlabeled audio — is practically meaningful. The paper also runs a sensible ablation (Min-Q, Re-Atten, DPO) that supports the RL choice. That's real work and credit where it's due.\n\nThe soft spot: Eq. (2) defines saliency using the gradient of a loss L(x), and the paper never says what L is when the input is unlabeled. The only natural choice is the model's own decoded hypothesis as the target, but the motivating evidence in Fig. 1 is computed with ground-truth tokens for the erroneous tokens. Those are different gradients, and the R_E>R_C separation does not automatically transfer. This is not a cosmetic omission — if the unsupervised Q doesn't rank hypotheses by true quality, the RL objective can amplify the wrong hypotheses. The authors need to specify the loss and, ideally, validate Q's correlation with WER/BLEU on a held-out set without using labels for training.\n\nAlso, the abstract's 'consistently and significantly' is stronger than the tables support: there are no error bars or significance tests, and in the car subset SI-SDA (2.5) is worse than STAR (2.4), and en-de/it-en are ties with STAR. That overclaiming should be fixed.\n\nOverall, the core idea deserves serious attention and the empirical pattern is promising, but the paper as written is not fully reproducible because the unsupervised saliency computation is undefined. I'd recommend sending it to peer review with a request for major revision — the gap is addressable, and the mechanism is worth testing once the loss is specified and the Q-ordering is directly validated.\n\nFor a reading group, it's a good discussion paper on unsupervised rewards for speech models.","headline":"Novel saliency-based unsupervised reward for audio LLM self-improvement, but the label-free saliency computation is under-specified and the empirical claims are slightly overstated; the core idea deserves a careful major revision.","tokens_in":9642,"tokens_out":3627,"would_cite":false,"duration_ms":37254,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Attention saliency can serve as a label-free reward for improving audio LLMs, cutting word error rate by up to 20.8%.","keywords":["audio large language models","self-improvement","unsupervised domain adaptation","attention saliency","reinforcement learning","automatic speech recognition","speech-to-text translation","spoken question answering"],"falsifier":"On any target set with ground truth, rank the beam hypotheses by $Q(\\hat{Y})$ and measure the correlation with WER or BLEU; if low-Q hypotheses are not systematically better, the reward in Eq. (10) should push probability toward worse hypotheses. A cleaner falsifying experiment would run SI-SDA on a domain whose errors come from lexical confusions rather than attention fallback; if WER does not improve while Q is still computed, the central assumption fails.","tokens_in":8593,"feed_emoji":"🎙️","tokens_out":17537,"duration_ms":148145,"temperature":0.7,"pith_summary":"SI-SDA, the method this paper proposes, asks whether a pre-trained audio large language model can improve itself on speech recognition, translation, and spoken question answering using only hours of unlabeled audio from the target domain. The paper claims the answer is yes: while decoding, the model's attention to prompt tokens versus acoustic frames reveals how likely a generated hypothesis is to be wrong, and this signal can be used as a reward in reinforcement learning. The authors report consistent improvements over several unsupervised baselines, including a 20.8% relative word error rate (WER) reduction on TED-3 and BLEU gains on CoVoST2 and LibriSQA, using only 2–5 hours of unlabeled speech. The broader consequence is that audio LLMs may not need curated labels or source-domain data to adapt to noisy or accented conditions, which would make real-world deployment cheaper.","feed_headline":"Cut speech-recognition word errors by up to 20.8% with unlabeled audio","feed_subtitle":"Attention saliency becomes a label-free reward that also lifts BLEU in translation and spoken question answering.","key_machinery":"The central object is the saliency-based quality score $Q(\\hat{Y})$: for each decoded token, the ratio of attention saliency spent on prompt tokens to total attention saliency, averaged over the whole hypothesis. Saliency here is a gradient-based importance measure over attention weights, computed via Taylor expansion in Eq. (2). The score serves as a label-free reward in a reinforcement-learning loop: each hypothesis in the beam receives an advantage $-(Q-\\bar{Q})$, so training shifts probability mass toward hypotheses that align with acoustic content rather than falling back on the prompt. This is the mechanism that lets the model adapt itself without ground-truth transcripts or source-domain data.","core_discovery":"The central claim is that the internal saliency of an audio LLM's decoding carries a domain-general quality signal: tokens decoded incorrectly tend to rely more heavily on prompt tokens than on acoustic frames. SI-SDA turns this into an unsupervised reward by averaging, over all positions of a candidate hypothesis, the fraction of attention saliency that lands on prompt tokens, yielding the score $Q(\\hat{Y})$. The model generates an N-best beam set, assigns each hypothesis an advantage based on whether its Q lies below the beam average, and updates low-rank adapter parameters (LoRA) through the policy-gradient objective in Eq. (10) to raise the probability of low-Q hypotheses. With no labeled data and no need to recall source-domain data, the method reports consistent gains over zero-shot, self-training, filtering, confidence reweighting, and STAR baselines across CHiME-4, LibriSpeech with FreeSound noise, CommonVoice, Switchboard, and TED-3, plus BLEU gains on CoVoST2 and LibriSQA.","pith_inferences":["A natural extension of the same logic is to test whether prompt-reliance is a generic failure signature across modalities: the attention-saliency ratio should also rank the quality of generated text or image captions in vision-language models.","One testable prediction is that the method's gains should scale with the frequency of prompt-fallback errors; domains where errors come from systematic acoustic confusions that do not change attention patterns should show little or no improvement.","The data-efficiency result suggests an iterative recipe: run SI-SDA on a new unlabeled corpus, use the improved model to produce better pseudo-labels, and repeat, turning the method into a fully autonomous continual-learning loop.","If Q reliably identifies low-quality hypotheses, it could be combined with diversity-based filtering to select which utterances to train on, potentially improving data efficiency even further."],"forward_implications":["If SI-SDA works as described, audio LLMs can be adapted to a new acoustic domain with only a few hours of unlabeled audio, removing the annotation bottleneck that typically blocks deployment in noisy or accented environments.","The reward should transfer to tasks where labels are not unique, because the paper reports BLEU gains in both speech-to-text translation and spoken question answering.","Because the model relies only on its own decoding and unlabeled speech, it can be updated as new target-domain audio arrives without keeping source-domain data.","The reported performance converges in roughly 2–5 hours, so the method can act as a fast first adaptation step before deciding whether labeled data collection is worthwhile.","On several test sets the adapted model approaches or ties the supervised fine-tuning upper bound, while on others it remains short of it, so the method narrows but does not eliminate the gap to labeled adaptation."],"supporting_citations":[{"why":"Supplies Qwen2-Audio, the audio LLM whose decoder attention and saliency are analyzed and adapted.","marker":"[25]"},{"why":"Provides the Taylor-expansion method for computing attention saliency used in Eq. (2).","marker":"[26]"},{"why":"Establishes the saliency-map interpretation that motivates reading attention as evidence of quality.","marker":"[27]"},{"why":"Supplies the minimum word-error-rate / RL training formulation that the policy-gradient objective in Eq. (10) builds on.","marker":"[28]"},{"why":"Defines the STAR unsupervised adaptation setting and evaluation protocol that SI-SDA compares against.","marker":"[16]"},{"why":"Documents attention-based hallucination patterns in multimodal LLMs, supporting the premise that decoding reveals quality.","marker":"[13]"},{"why":"Provides DPO as the preference-optimization baseline in the RL ablation study.","marker":"[35]"},{"why":"Provides the CoVoST2 corpus used for the speech-to-text translation extension.","marker":"[33]"},{"why":"Provides the LibriSQA dataset used for the spoken question-answering extension.","marker":"[34]"},{"why":"Provides the CHiME-4 noisy speech corpus used for the main ASR evaluation.","marker":"[29]"}],"fun_headline_variants":["Unlabeled speech self-improves audio LLMs across tasks","Self-improvement for audio LLMs without labeled data","Audio LLM gets better with unlabeled audio via RL","SI-SDA: Unsupervised self-improvement for speech AI","Unlabeled audio boosts ASR, SQA, and translation BLEU"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method rests on a single bet: that an audio LLM's attention pattern marks its own mistakes, so a candidate hypothesis that leans more on the prompt than on the audio is likely wrong. The reward also depends on a saliency score computed from a loss gradient, yet in the unlabeled setting the paper does not specify what target that loss uses; if either link fails in a new domain, the reinforcement signal trains the model toward worse outputs.","fun_headline_variants_meta":{"raw":{"variants":["Unlabeled speech self-improves audio LLMs across tasks","Self-improvement for audio LLMs without labeled data","Audio LLM gets better with unlabeled audio via RL","SI-SDA: Unsupervised self-improvement for speech AI","Unlabeled audio boosts ASR, SQA, and translation BLEU"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.0002,"raw_usage":{"total_tokens":1352,"prompt_tokens":902,"completion_tokens":450,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":518,"completion_tokens_details":{"reasoning_tokens":364}},"tokens_in":518,"tokens_out":450,"duration_ms":5267,"temperature":1.0,"reasoning_tokens":364,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T17:47:10.587106+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On any target set with ground truth, rank the beam hypotheses by $Q(\\hat{Y})$ and measure the correlation with WER or BLEU; if low-Q hypotheses are not systematically better, the reward in Eq. (10) should push probability toward worse hypotheses. A cleaner falsifying experiment would run SI-SDA on a domain whose errors come from lexical confusions rather than attention fallback; if WER does not improve while Q is still computed, the central assumption fails.","supporting_citations":[{"cited_title":"Minimum word error rate training for attention-based sequence-to-sequence models,","cited_arxiv_id":null,"evidence_quote":"Supplies the minimum word-error-rate / RL training formulation that the policy-gradient objective in Eq. (10) builds on."},{"cited_title":"Self-Taught Recognizer: Toward Unsupervised Adaptation for Speech Foundation Models","cited_arxiv_id":"2405.14161","evidence_quote":"Defines the STAR unsupervised adaptation setting and evaluation protocol that SI-SDA compares against."},{"cited_title":"Opera: Alleviating hallucination in multi- modal large language models via over-trust penalty and retrospection- allocation,","cited_arxiv_id":null,"evidence_quote":"Documents attention-based hallucination patterns in multimodal LLMs, supporting the premise that decoding reveals quality."},{"cited_title":"Direct preference optimization: Your language model is secretly a reward model,","cited_arxiv_id":null,"evidence_quote":"Provides DPO as the preference-optimization baseline in the RL ablation study."},{"cited_title":"Librisqa: A novel dataset and framework for spoken question answering with large language models,","cited_arxiv_id":null,"evidence_quote":"Provides the LibriSQA dataset used for the spoken question-answering extension."},{"cited_title":"The 4th chime speech separation and recognition challenge,","cited_arxiv_id":null,"evidence_quote":"Provides the CHiME-4 noisy speech corpus used for the main ASR evaluation."}],"review_version":1}