{"id":"c21c7562-e51c-4698-a0d3-6e88b167dcfe","arxiv_id":"2608.12703","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Speculative ASR drafts lose accuracy between verification steps because their audio attention drifts from the correct position, and runtime or training-time repositioning recovers part of the lost speed.","lead":"This paper finds that a small auxiliary model used to speed up speech recognition loses track of where it is in the audio as it guesses several words in a row, and this drift explains a large part of why its later guesses get rejected. The authors test two ways to restore the audio position, and they show the speed gain is real only when the recovered guesses pay for the correction.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 'correct' window is anchored solely by MMS-FA; if MMS-FA is systematically biased for Qwen, the causal position contrast and the drift metric lose their stated meaning. An independent aligner cross-check would settle it.","rationale":"The reader's weakest_assumption identifies exactly the same load-bearing concern: MMS-FA is treated as ground truth for audio position, and the Voxtral +205 ms offset shows that attention-based position measures can disagree systematically with MMS-FA. The paper's own boundary case therefore demonstrates that MMS-FA is not self-evidently the true anchor; it is an aligner whose errors are unmeasured for Qwen. This concern is load-bearing because the causal window intervention defines 'correct' and 'wrong' centers relative to MMS-FA, and the acceptance recovery (+0.254 correct-versus-shifted at depth 2, Table 12) is the primary evidence for the mechanism. If an independent aligner disagrees with MMS-FA by a large constant, the 'recovery' could be an artifact of the aligner's bias rather than a recovery of genuine audio position. The runtime correction and AnchorDraft also depend on MMS-FA (layer selection and training supervision), so the concern propagates beyond the offline intervention. However, the paper has strong internal controls: matched window width, prespecified shift, official-split replication, bootstrap intervals, and the full five-arm matrix; these rule out many confounds but not a constant aligner bias. This is a standard validation gap in speech alignment research, not an internal inconsistency. The concrete test -- comparing MMS-FA to a second independent aligner and re-running the key contrast -- is straightforward and decisive. Since the reader already gave CONDITIONAL based on this and two other secondary concerns, my analysis does not move the verdict; it strengthens the same condition. No significant new objection beyond the reader's weakest assumption was identified after carefully reading the mechanism, corrections, and limitations sections.","tokens_in":27670,"tokens_out":6744,"duration_ms":71983,"concrete_test":"On the same 120-utterance official-split evaluation sets used in Supp. C.1 and Table 12, produce an independent forced alignment with a second aligner (e.g., wav2vec2-CTC or the target Qwen model's own greedy non-speculative attention). Compute the median absolute difference between this second aligner's per-token frames and MMS-FA. If the median difference exceeds roughly 5 frames (one post-stride frame is ~80 ms), recompute the depth-2 correct-versus-shifted continuation acceptance contrast in Table 12 using the second aligner's frames as window centers. If the +0.254 contrast does not persist with the alternative centers, the stated mechanism is MMS-FA-specific rather than a property of true audio position.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central causal claim in Section 4 ('Matched Effect of Window Position') rests on defining the 'correct' audio position as the MMS-FA forced-alignment frame. The paper validates the draft attention peak by correlation with MMS-FA (Supp. table 7), but MMS-FA is itself a model-based aligner with no independent ground-truth check for Qwen. Supp. B.2 reports Voxtral attention is systematically +205 ms late relative to MMS-FA, demonstrating that attention-based and MMS-FA positions can differ by a large constant offset. For Qwen, the signed offset is not reported; if a similar bias exists, both the 'correct' and 'shifted' window centers in the prespecified intervention (Supp. A.2) are displaced from the true audio anchor. The depth-2 correct-versus-shifted continuation contrast of +0.254 (Table 12) would then show sensitivity to some reference position, but the specific mechanism -- the draft losing the true audio anchor -- would not be established. AnchorDraft's training supervision (Eq. 6) and the runtime layer-21 selection (Supp. A.3) also use MMS-FA as the reference, so the bias would propagate to both corrections. The paper's matched controls, width sweep, and official-split replication are strong internal evidence, but they cannot rule out a constant aligner bias because every 'correct' position is defined by the same aligner. This is a validation gap, not an observed contradiction; an external alignment reference is the missing check.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies single-model speculative decoding for ASR, where a lightweight draft attached to the target model proposes tokens between target verifications. It identifies a phenomenon it calls 'alignment drift': the draft's audio cross-attention moves away from a reference audio frame (the 'anchor') during continuation, and this drift is associated with lower acceptance of later draft proposals. The authors support this with a matched intervention that fixes window width and changes only the window center: a window centered on forced-alignment frames recovers continuation acceptance, while an equally narrow wrong-position window reduces it. They propose two corrections: a runtime correction that reads audio position from verification attention, and a training correction (AnchorDraft) that adds a guided-attention loss using forced-alignment supervision. Both improve end-to-end speed when the recovered accepted tokens cover the correction cost, with the cost condition expressed by a screening rule. The paper includes extensive supplementary material with official-split replications, paired bootstrap intervals, cross-architecture comparisons, and deployment-scope measurements.","tokens_in":27976,"tokens_out":5300,"duration_ms":53692,"significance":"If the central claim holds, the paper makes a valuable contribution by showing that in single-model speculative ASR, continuation quality depends not only on next-token prediction but also on audio-position tracking and on the cost of the draft. The study is unusually careful in several respects: the main causal estimate uses matched official-split arms with a fixed window width and paired bootstrap confidence intervals; the draft attention peak is validated against an external forced aligner (MMS-FA); the cost screening rule is tested across multiple settings; and the authors explicitly separate association from causation, including diagnostic boundaries and a limitations section. The artifact includes bootstrap code, run metadata, and mapping rules, which supports reproducibility. The main weakness is that MMS-FA, the sole external reference used to define 'correct' audio position, is itself a model-based aligner and is not independently validated for the Qwen checkpoints in this paper; this gap affects the interpretation of the causal contrast and should be addressed before publication.","major_comments":[{"comment":"The manuscript defines the 'correct' audio position using MMS-FA forced-alignment frames and uses that reference both to validate the draft attention peak and to center the window intervention. MMS-FA is itself a model-based aligner with no independent ground-truth check for Qwen in this paper. Supp. B.2 reports that Voxtral attention is systematically +205 ms late relative to MMS-FA, showing that attention-based and MMS-FA positions can differ by a large constant; for Qwen, Supp. table 7 reports only mean absolute errors (136–155 ms) and no signed offset. If Qwen has a similar systematic bias, both the 'correct' and 'shifted' window centers in the matched intervention are displaced from the actual acoustic anchor, and the depth-2 correct-versus-shifted contrast of +0.254 (fig. 2a, Supp. table 12) would demonstrate sensitivity to some reference position but would not specifically establish the mechanism of losing the true audio anchor. The 2-frame verification-attention error is also measured against MMS-FA, so it does not break the circularity. I recommend reporting the signed median offset for Qwen, performing a sensitivity analysis that shifts all window centers by that constant, and/or validating the anchor on a small set with an independent alignment source (for example, manually checked phoneme boundaries or a second aligner). This is a validation-gap concern rather than an observed contradiction, but it is load-bearing because the labels 'alignment drift' and 'the draft loses the true audio anchor' depend on the reference being accurate.","section":"Section 4, Supp. A.2, Supp. B.2, Supp. table 7"},{"comment":"The matched 'with audio' vs. 'no audio' drafts differ not only in whether per-step audio access exists but also in architecture: the audio-conditioned draft adds cross-attention parameters, so the roughly doubled continuation acceptance could partly reflect additional capacity rather than audio access per se. The paper states that both drafts share the same layer budget, but it does not control for parameter count. This does not invalidate the window-intervention causal estimate, but it weakens the motivating claim that 'per-step audio access changes the first proposal modestly but roughly doubles later-proposal acceptance.' I suggest either adding a control with a non-audio cross-attention module of comparable parameter count or explicitly stating in the main text that this comparison confounds access with added capacity.","section":"Section 3, Table 1, Supp. table 6"},{"comment":"The cost screening rule in Eq. (1) is used to claim that the measured result becomes positive only after the recovered accepted length clears the cost, with a reported 10/10 direction match across deployed evaluation-set point estimates. However, the rule's components—the draft-step cost ratio ρ, the correction cost ΔC_corr, and the accepted-length gain ΔL—are all measured on the same implementation and possibly on the same settings used to form the rule. The paper does not state whether the 19 settings in fig. 3c were used to choose the rule or whether the rule was specified before inspecting the timing results. As written, the sign-matching test has limited confirmatory value. I recommend clarifying the timeline and, if possible, reporting a leave-one-setting-out check on the direction match.","section":"Section 5, Eq. (1), fig. 3c, Supp. D.1"}],"minor_comments":[{"comment":"The abstract and main text report '21 frames' and '2-frame median' without stating in the same sentence that these are distances to MMS-FA reference frames; please state the reference explicitly to avoid implying a ground-truth acoustic position.","section":"Abstract, Section 4"},{"comment":"Please add the signed median offset for Qwen, alongside the Voxtral +205 ms offset, so readers can assess whether a constant-bias correction would change the window-intervention conclusions.","section":"Supp. table 7"},{"comment":"The legend in the figure uses 'wrong' while other parts of the paper use 'shifted'; please standardize the terminology across the figure and text.","section":"Figure 2c"},{"comment":"The window-width robustness result in Supp. Section A is mentioned in the main text, but the actual tested half-width values are not given; please add the range (e.g., ±200 ms to ±800 ms) to the main text for completeness.","section":"Section 4, 'Position Interventions'"},{"comment":"The main text says the probe layer and ridge strength are fixed on development data but does not give the selected layer; Supp. table 9 reports layer 20, so please include this value in the main text.","section":"Section 4, 'Probe protocol'"}],"recommendation":"major_revision","confidential_remarks":"The MMS-FA validation gap is the main risk to the paper's central mechanism claim. The authors should be asked to report signed offsets for Qwen and a sensitivity analysis shifting the intervention centers by that offset, or to provide an independent alignment check on a small sample. The paper is otherwise thorough, with strong internal controls and a useful artifact, and it fits the journal's scope."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nRead this one. The headline is that the paper finds a real, previously undifferentiated failure mode in single-model speculative ASR: the draft degrades between target verifications because it loses the audio position, not just because it makes worse token predictions. The restart/continuation split is the key move. The matched audio/no-audio drafts show the loss is concentrated in continuation, and the prespecified window-centering intervention—same width, only center changed—makes a decent causal case: correct center recovers continuation, wrong center lowers it, with bootstrap intervals that don't cross zero through depth three. That is the strongest part of the paper. The cost condition (Eq. 1) is also a nice discipline: they only claim speedups when recovered accepted length covers readout/draft cost, and they measure in the real cached loop rather than offline.\n\nThe paper is honest about its limits. It flags that low-precision verification is not token-exact, that the position intervention explains only part of the rejection, and that batch-one short-form does not generalize. That transparency is real.\n\nSoft spots, in order. First, the 'correct' audio position is defined by MMS-FA forced alignment, which is itself a model. The paper validates the attention peak against MMS-FA, but never checks MMS-FA against anything independent for Qwen. The Voxtral +205 ms offset shows attention and MMS-FA can disagree by a large constant. If MMS-FA is similarly biased on Qwen, both the 'correct' and 'shifted' windows are displaced and the causal contrast shows sensitivity to a reference position rather than to the true anchor. That is a validation gap, not a contradiction. An independent aligner or a direct manual time-stamp check would settle it. Second, the headline runtime correction at 1.7B has a clean interval that includes zero: five-set mean +4.5%, but the interval crosses zero. The AnchorDraft training gains are cleaner. Third, the artifact is described but there is no URL or commit hash in the preprint, so I could not verify the reproducibility claim. Minor: depth-four estimates are based on tiny survival fractions and the interval includes zero; the paper says so, so I don't hold it against them.\n\nWho this is for: ASR folks who care about speculative decoding and anyone who thinks about self-speculation in continuous-input seq2seq. It deserves a serious referee. Send it to review and ask the authors to (1) cross-check the aligner on Qwen, (2) release the artifact, and (3) separate the AnchorDraft result from the runtime correction in the abstract framing. I would cite it.","headline":"A genuinely useful empirical paper: the restart/continuation split and matched window intervention make a real case for alignment drift, but the anchor rests on MMS-FA and the artifact is not yet checkable.","tokens_in":28551,"tokens_out":2064,"would_cite":true,"duration_ms":21279,"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":"In single-model speculative ASR, the draft's audio attention drifts between verifications; recentering it on the audio anchor recovers continuation acceptance.","keywords":["single-model speculative decoding","automatic speech recognition","alignment drift","audio anchor","continuation acceptance","verification attention","AnchorDraft","draft-verify loop"],"falsifier":"Run the fixed-width window intervention on utterances whose token-to-audio positions come from a gold-standard source (e.g., manually transcribed time boundaries or a forced aligner with independently verified accuracy on the target models). If windows centered on those frames no longer recover continuation acceptance relative to equally narrow wrong-position windows, or if the correct-versus-wrong contrast reverses, the alignment-drift explanation is refuted. A complementary test: use a synthetic audio condition where token durations are constant, so no drift should accumulate, and show that continuation acceptance under draft-only steps does not decay.","tokens_in":27422,"feed_emoji":"🎙️","tokens_out":14475,"duration_ms":116341,"temperature":0.7,"pith_summary":"The paper tries to establish that in single-model speculative decoding for speech recognition, the draft's proposals degrade between target verifications because its audio attention drifts away from the audio position of the next token; the authors call this alignment drift. They claim this drift is a recoverable cause of lost continuation acceptance, distinguishable from weak next-token prediction or limited draft capacity. The key evidence is a window intervention that keeps the visible audio width fixed and moves only the center: a window centered on the forced-alignment frame recovers continuation acceptance, while an equally narrow window at a wrong position reduces it. The paper also shows that the position can be reacquired cheaply at runtime from verification attention, or taught during training with a guided-attention loss called AnchorDraft, and that either correction speeds up end-to-end decoding only when the recovered accepted tokens cover its cost. If true, the result makes audio-position tracking a first-class requirement for self-speculative ASR, alongside token prediction and draft cost.","feed_headline":"Alignment drift slows ASR speculation; recentering restores speed","feed_subtitle":"Draft attention drifts up to 21 frames; verification attention recenters it to 2, restoring speed.","key_machinery":"The central object is the audio anchor: the encoded-audio frame that should feed the next draft token, operationally measured as the peak of the draft's audio cross-attention and validated against MMS-FA forced alignment. The argument's load-bearing tool is the window intervention, which fixes the attention mask width (half-width ±400 ms) and changes only its center, isolating audio position from token information and width effects. Two corrections attach to this machinery: a runtime correction that reads the target's verification attention (layer 21 selected on development data) as a fresh anchor, and AnchorDraft, a training-only loss that matches the draft's first-layer attention to a Gaussian centered on the forced-alignment frame for the next token while leaving the inference graph unchanged. A cost-screening inequality, $\\frac{\\Delta L}{1+L} \\gtrsim \\frac{\\Delta C_{\\mathrm{corr}}}{K\\rho+1}$, with $\\rho$ the draft-to-verification cost ratio, predicts which corrections will improve latency.","core_discovery":"In single-model speculative decoding for ASR, per-step audio access changes the first proposal after verification only modestly, yet it roughly doubles the acceptance of later draft-only proposals. The paper's central claim is that this continuation gap is caused by alignment drift: the draft's audio cross-attention peak moves away from the audio anchor by a growing number of frames as the draft runs without the target, with late-draft median error reaching 21 frames in the hardest condition, while target attention during verification stays within a 2-frame median. Because the accepted transcript records only token position, not audio position, the draft loses track of where the next token should be read from. The causal test is a fixed-width attention window whose center is the only thing that changes: a correct-position window recovers continuation (depth-two conditional acceptance rises by +0.144 over unrestricted attention), a wrong-position window lowers it (−0.111), and the matched correct-versus-wrong contrast is +0.254 with 95% CI [+0.241,+0.268] through depth three. The paper further argues that this position is recoverable: verification attention reduces late-draft error to a 2-frame median and gives a mean +3.4% speed gain at 1.7B after readout cost, and the AnchorDraft training objective (a Gaussian guided-attention loss on the first draft layer) improves five-set mean speed by +3.9% to +5.0% at 1.7B and +6.4% to +7.1% at 0.6B, with WER changes at most 0.01 percentage points. The paper concludes that ASR self-speculation depends on token prediction, audio-position tracking, and draft cost, with the cost condition $\\frac{\\Delta L}{1+L} \\gtrsim \\frac{\\Delta C_{\\mathrm{corr}}}{K\\rho+1}$ determining when a correction helps.","pith_inferences":["Beyond the paper's tested ASR models, the same restart-continuation asymmetry should appear in any single-model speculative decoder that consumes a continuous input (video frames, raw audio for music, etc.) because the draft must track a moving anchor between verifications; measuring that anchor error would transfer the result directly.","The cost condition implies specific deployment thresholds: when batch size grows (the draft-to-verification cost ratio $\\rho$ rises from 0.0275 at batch one to 0.1274 at batch 64), the runtime correction should become less beneficial; this is a quantitative prediction that the paper's own measurement setup could test.","Since correct positioning recovers only part of the continuation loss at depth four, a residual source of failure remains; combining anchor correction with a draft that exposes next-token features earlier (the token probe shows they peak late in the decoder) is a plausible next step to close the remaining gap."],"forward_implications":["Audio access without audio localization is insufficient: the first proposal after verification barely changes, so any single-model ASR draft must maintain position across draft-only steps.","A correct-centered window improves cumulative survival at depth two by +0.105 and accepted length by +0.262, defining a quantitative recovery target for any position-correcting proposal.","Verification attention is a cost-effective runtime position source on the tested Qwen checkpoints, reducing late-draft median error from up to 21 frames to 2 frames, with a five-set mean speed gain of +3.4% at 1.7B.","AnchorDraft training delivers speed gains at both tested scales without altering the decoding graph or degrading WER (max absolute change 0.01 percentage points).","The two corrections are complementary: after AnchorDraft training, adding the runtime readout gives a further +1.8% (95% CI [+0.9%,+2.7%]) speed gain at 1.7B."],"supporting_citations":[{"why":"defines the speculative decoding verification rule that the whole paper assumes","marker":"Leviathan, Kalman, and Matias 2023"},{"why":"establishes the draft-then-verify framework for autoregressive generation","marker":"Chen et al. 2023"},{"why":"provides the single-model draft architecture (direct token prediction plus multi-layer feature input) that the shallow ASR draft is built on","marker":"Li et al. 2025a"},{"why":"supplies the forced-alignment frames that define the 'correct' window center and validate the attention-peak anchor measure","marker":"Pratap et al. 2024"},{"why":"gives the guided-attention objective that AnchorDraft adds to the draft training loss","marker":"Tachibana, Uenoyama, and Aihara 2018"},{"why":"are the target ASR models whose checkpoints (0.6B and 1.7B) carry the main mechanism and correction results","marker":"Shi et al. 2026"},{"why":"is the other-family ASR model used as a boundary case, with its attention peak systematically late by about +205 ms relative to MMS-FA","marker":"Liu et al. 2025"},{"why":"provides the LibriSpeech corpus from which the matched feasibility drafts are trained and on which acceptance and speed are measured","marker":"Panayotov et al. 2015"}],"fun_headline_variants":["21-frame drift fixed by 2-frame verification for faster ASR","Draft drift fixed: verification recenters audio to 2 frames","AnchorDraft trains draft to track audio, speeding ASR speculation","Only fix draft drift when acceptance gains offset readout cost","ASR speculation: correct audio window improves acceptance, wrong one hurts"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The causal claim that centering the window on the audio position recovers acceptance assumes the MMS-FA forced-alignment frame is the true per-token audio position; if MMS-FA is systematically biased for the target model, both the 'correct' and 'wrong' centers shift together and the measured correct-versus-wrong contrast could be distorted.","fun_headline_variants_meta":{"raw":{"variants":["21-frame drift fixed by 2-frame verification for faster ASR","Draft drift fixed: verification recenters audio to 2 frames","AnchorDraft trains draft to track audio, speeding ASR speculation","Only fix draft drift when acceptance gains offset readout cost","ASR speculation: correct audio window improves acceptance, wrong one hurts"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001839,"raw_usage":{"total_tokens":7364,"prompt_tokens":1219,"completion_tokens":6145,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":835,"completion_tokens_details":{"reasoning_tokens":6056}},"tokens_in":835,"tokens_out":6145,"duration_ms":41633,"temperature":1.0,"reasoning_tokens":6056,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T04:12:59.623150+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the fixed-width window intervention on utterances whose token-to-audio positions come from a gold-standard source (e.g., manually transcribed time boundaries or a forced aligner with independently verified accuracy on the target models). If windows centered on those frames no longer recover continuation acceptance relative to equally narrow wrong-position windows, or if the correct-versus-wrong contrast reverses, the alignment-drift explanation is refuted. A complementary test: use a synthetic audio condition where token durations are constant, so no drift should accumulate, and show that continuation acceptance under draft-only steps does not decay.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"supplies the forced-alignment frames that define the 'correct' window center and validate the attention-peak anchor measure"},{"cited_title":"2026 , eprint=","cited_arxiv_id":null,"evidence_quote":"are the target ASR models whose checkpoints (0.6B and 1.7B) carry the main mechanism and correction results"}],"review_version":1}