{"id":"f187588c-8423-4183-b65e-fcc3b1ab5e1c","arxiv_id":"2607.13158","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Teacher-built bounded-waiting prefix targets let a chunked streaming speech LLM improve simultaneous translation quality by +1.54 COMETKiwi at +0.15 s latency.","lead":"A speech-translation model can be made to translate while speech is still arriving by training it on short, partial translations produced by a stronger teacher, instead of adding new policy heads. On an internal conversational test set, prefix-to-prefix fine-tuning improved streaming quality by +1.54 COMETKiwi while increasing average lagging by only 0.15 seconds.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Main +1.54 gain is confounded: CSSEL-P2P adds a 200h/pair fine-tuning stage that the CSSEL streaming baseline lacks; without an offline-reference same-data control, the improvement cannot be attributed to P2P labels.","rationale":"I chose the generic-fine-tuning confound over teacher-label quality because it is more directly load-bearing for the central attribution claim. Even if the teacher labels are perfect, the experiment as designed cannot distinguish 'P2P teaches commit-ready behavior' from 'extra fine-tuning on in-house conversational data helps streaming.' The reader's weakest assumption (label validity) is plausible and is admitted as a limitation, but it would only bite if labels are systematically wrong; the missing control bites regardless of label quality. The title's architectural claim is also overstated (the CSSEL encoder is causal/chunk-masked), but this affects framing more than the empirical comparison. A same-data offline-reference control would settle the attribution; until then, CONDITIONAL/UNCHANGED is appropriate: accept conditioned on this control. The paper has genuine positives: a clean rewind-based no-revision decoding formulation, a public CoVoST2 diagnostic showing the same trend, and explicit latency/quality tradeoff analyses. Those are preserved regardless of the control outcome. Hence no change to the reader's verdict; the condition should be made explicit.","tokens_in":10662,"tokens_out":3595,"duration_ms":35611,"concrete_test":"Fine-tune CSSEL on the exact same 200h/pair P2P data split, but replace the bounded-waiting teacher labels with standard offline full-sentence references (optionally truncated at chunk boundaries) while keeping the same training recipe, chunk size ∆=4s, rewind k=5, and decoding. If this offline-reference control reproduces most of the +1.54 COMETKiwi gain over CSSEL streaming, the claim that commit-ready P2P targets are the active ingredient fails; if the control gains only a small fraction, the P2P attribution is supported. Report per-direction gaps and AL.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that prefix-to-prefix supervision, not architectural/policy changes, drives SimulST gains. The evidence is Table I: CSSEL-P2P improves streaming COMETKiwi by +1.54 (68.96→70.50) at +0.15s AL, and Table II shows recovery of over half the streaming loss. But the comparison is not matched: CSSEL-P2P is obtained by fine-tuning CSSEL on an additional 200 hours per direction of teacher-labeled data (Sec. III-C, IV-A1), whereas the 'CSSEL streaming' baseline receives no additional fine-tuning. Any additional adaptation—to the conversational domain, to 4s chunked input, or to partial-audio conditions—could produce part or all of the gain. The paper's Sec. IV-C acknowledges teacher-label quality as a limitation, but the more immediate confound is the missing control. The reader's rationale already lists this; it should be the primary condition for acceptance. The title-level 'no architectural changes' issue is secondary: CSSEL itself uses a chunked streaming encoder with causal masking, so the architecture is not truly unchanged; but even granting that framing, the data-attribution problem remains.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper asks whether decoder-only LLMs need architectural/policy changes for simultaneous speech translation (SimulST) and proposes a data-driven alternative. It introduces CSSEL, a Phi-4-MM backbone with a chunked streaming speech encoder, and CSSEL-P2P, obtained by fine-tuning CSSEL on teacher-labeled prefix-to-prefix (P2P) targets. The P2P labels are produced by prompting GPT-o3-mini over the full chunk sequence with a bounded-waiting instruction, and inference uses cumulative chunk decoding with a rewind-based committed prefix. The main claim is in Table I: CSSEL-P2P improves streaming COMETKiwi from 68.96 to 70.50 at comparable average lagging (+0.15 s), and Table II on a CoVoST2 subset shows recovery of more than half of the offline-to-streaming degradation. The paper argues that P2P supervision, not architectural changes, is the main lever for SimulST.","tokens_in":10973,"tokens_out":3770,"duration_ms":43105,"significance":"If the central comparison is valid, the result is a useful contribution: it provides a fixed-chunk streaming recipe with explicit quality-latency knobs (chunk size Δ, rewind length k) and a concrete teacher-labeling protocol for commit-ready prefix targets, and it shows that such targets can be used to fine-tune a decoder-only LLM without adding policy heads. The paper includes a public diagnostic evaluation, a trade-off analysis over Δ and k, and an honest discussion of limitations (teacher-label quality, tuning knobs). These are strengths. The main weakness is that the headline gain is not established as caused by the P2P labels, because the comparison is confounded by an additional fine-tuning stage, and no statistical significance is reported. The paper's contribution would be solid if the missing control is supplied and the result persists.","major_comments":[{"comment":"The headline +1.54 COMETKiwi gain is not a matched comparison. CSSEL-P2P is obtained by fine-tuning CSSEL on an additional 200 hours per direction of teacher-labeled P2P data (Sec. III-C, IV-A1), whereas the CSSEL streaming baseline receives no additional fine-tuning. The paper does not report a control in which CSSEL is fine-tuned on the same 200 hours with ordinary offline full-sentence references and then evaluated under the same streaming decoder. Without such a control, the gain could reflect generic adaptation to the conversational domain or to 4-s chunked inputs. The absence of an offline CSSEL-P2P row in Table I also leaves open the possibility that P2P fine-tuning improves translation quality in general rather than streaming-specific behavior. This control is load-bearing for the central claim and must be added.","section":"§IV-A2, Table I; §III-C"},{"comment":"All reported quality scores are point estimates with no confidence intervals or significance tests. The per-direction differences in Table I are often small (e.g., 0.19-0.34 COMETKiwi), and the overall +1.54 average could be driven by a few large gains on reordering-heavy pairs. The paper should report bootstrap confidence intervals or at least a paired significance test for the CSSEL vs. CSSEL-P2P comparison, and ideally for the diagnostic Table II as well. Without this, the 'consistent improvements' claim is not quantitatively supported.","section":"Table I; Table II"},{"comment":"The teacher labels are the sole source of P2P supervision, but their quality is not independently verified. The teacher is prompted to simulate a causal translator while being given the full chunk sequence, and the paper itself lists 'dependence on teacher-label quality' as a limitation. Since all P2P training data depend on these labels, the paper should provide some evidence that the labels are actually commit-ready (e.g., human judgments on a sample, comparison with an alternative labeling strategy such as wait-k, or analysis of label waiting behavior). This is not a fatal flaw, but it is necessary to validate the central mechanism.","section":"§III-B.2; §IV-C"}],"minor_comments":[{"comment":"Typo: 'BELU' should be 'BLEU' in the text of the diagnostic evaluation section.","section":"§IV-B"},{"comment":"The right panel is labeled 'CSSEL DE→EN' while the left panel uses 'CSSEL-P2P EN→DE'; please clarify in the caption or the text whether the Δ sweep is performed on the base CSSEL or on CSSEL-P2P, as this affects the interpretation of the trade-off.","section":"Fig. 2"},{"comment":"The phrase 'without architectural changes' is overstated: CSSEL itself introduces a chunked streaming speech encoder with causal convolutions and a chunk-wise attention mask (Sec. IV-A1). The claim that is actually supported is that, within a fixed streaming architecture, P2P supervision is sufficient and no policy heads or explicit read/write tokens are needed. Recommend rewording to avoid confusion.","section":"Title and Abstract"},{"comment":"The description of the main in-house evaluation is thin: the test set sizes, the number of utterances per direction, and the selection/domain composition are not given. This is important for interpreting the average scores and for assessing whether the 'in-house conversational speech' claim generalizes.","section":"§IV-A1"}],"recommendation":"major_revision","confidential_remarks":"The central idea is promising and the experiments show a positive direction, but the missing matched control is a serious gap for a paper whose thesis is that P2P data, not architecture or policy, drives the gain. If the authors add the offline-reference same-data control and it shows the effect is specific to P2P labels, the paper would be acceptable. Also, given that all primary results are on inaccessible in-house data, the diagnostic public-set evidence should be reported with the same rigor. The lack of significance testing is an additional concern for a journal-level acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The method is worth reading: fixed-chunk decoding with a rewind buffer plus teacher-LLM bounded-waiting prefix labels. But the main result — P2P fine-tuning buying +1.54 COMETKiwi over the streaming baseline — is not yet convincing, because the comparison is not matched. CSSEL-P2P gets an extra 200 hours per direction of fine-tuning that the CSSEL streaming baseline doesn't get. Any additional in-domain adaptation could explain part or all of the gain, so the improvement cannot be attributed to P2P labels on this evidence.\n\nWhat is genuinely good here: the rewind-based anchoring is a clean trick. Commit only the prefix minus the last k tokens, so the model can adjust an editable suffix without revising user-visible output. The teacher-LLM label generation with a bounded waiting budget is a reasonable way to construct commit-ready targets, and the paper is honest about its dependence on teacher quality. The write-up is clear, and the small CoVoST2 diagnostic is a sensible check, even if it only covers three directions.\n\nThe load-bearing soft spot is the missing control. The authors need to fine-tune on the same 200h per pair with offline full-sentence references, or some other non-P2P target, to isolate the effect of the prefix supervision. Without that control, the central claim doesn't hold up. The title-level \"no architectural changes\" claim is also overstated: CSSEL itself introduces a chunked streaming encoder with causal masking, so the architecture is changed relative to the base LLM. What they likely mean is \"no policy heads,\" which is fair, but the title overreaches. The in-house-only evaluation with no error bars or significance tests is a secondary weakness, not fatal.\n\nWho this is for: people working on LLM-based SimulST, especially those deciding between adding policy heads and shaping behavior through data. I'd take the method idea seriously and would like to see the experiment re-run with the proper control. If the gain survives a same-data offline-reference fine-tuning control, this becomes a solid contribution. As it stands, it deserves a serious referee, but the referee should ask for that control before accepting. My recommendation: engage with it, send it out, and set the bar at the missing control.","headline":"A promising streaming recipe and an interesting label-generation idea, but the headline +1.54 COMETKiwi gain is confounded by an extra fine-tuning stage the baseline lacks; the paper needs a same-data offline-reference control before the claim can be trusted.","tokens_in":11471,"tokens_out":2772,"would_cite":false,"duration_ms":40948,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Fine-tuning on teacher-built prefix-to-prefix targets makes a decoder-only speech LLM translate simultaneously without architectural changes, lifting streaming quality by +1.54 COMETKiwi at only +0.15s extra lag.","keywords":["simultaneous speech translation","SimulST","decoder-only LLM","prefix-to-prefix supervision","streaming decoding","chunk-based decoding","rewind-based committed prefix","bounded waiting"],"falsifier":"A controlled ablation replacing the teacher LLM with a naive monotone truncation of the offline reference (no bounded-waiting, no global planning) would decide the matter: if truncation labels recover similar gains, the paper's specific label construction is not the active ingredient; if they do not, the teacher's planning is. A complementary quantitative check is to re-score each committed prefix by asking the full-utterance model how much of the reference's continuation is already contradicted; high contradiction rates would signal premature commit-ready labels.","tokens_in":10570,"feed_emoji":"🎙️","tokens_out":3952,"duration_ms":37268,"temperature":0.7,"pith_summary":"This paper asks whether decoder-only LLMs really need architectural changes for simultaneous speech translation, and answers no: the bottleneck is data, not architecture. It proposes fixed-chunk cumulative decoding with a rewind-based committed prefix, and teaches the model what to commit at each prefix using prefix-to-prefix targets generated by a teacher LLM under bounded waiting. On in-house conversational speech, fine-tuning on these targets raises streaming quality by +1.54 COMETKiwi over the streaming baseline at nearly the same latency. The central claim is that emission timing can be encoded in training targets rather than in read/write policy modules or new model components.","feed_headline":"Prefix-to-prefix fine-tuning lifts streaming speech translation","feed_subtitle":"No new policy heads: teacher-built commit-ready prefixes add +1.54 COMETKiwi at only +0.15s lag.","key_machinery":"The load-bearing objects are: (1) CSSEL, a decoder-only speech LLM with a chunked streaming speech encoder whose self-attention sees only a bounded left context; (2) cumulative chunk decoding with rewind-based prefix anchoring, where at each step the decoder is forced to continue from the previously emitted prefix minus the last k tokens; and (3) teacher-labeled prefix-to-prefix supervision, in which a teacher LLM prompted with all chunks produces, for each chunk, a translation prefix or an empty wait, subject to a maximum waiting budget of K=3. The P2P targets are the mechanism that encodes emission timing directly into the model's learned behavior, aligning training with the no-revision st","core_discovery":"The paper's central claim is that the dominant bottleneck for decoder-only LLM-based simultaneous speech translation is the training data, not the architecture. With fixed-length chunks of speech fed cumulatively to the model, and a rewind boundary that lets the last k tokens stay hidden and editable, the model can be anchored to a monotonically growing committed prefix. The key discovery is that fine-tuning the model on teacher-labeled prefix-to-prefix targets—where a strong instruction-following LLM plans a bounded-waiting emission schedule over chunk-level transcriptions—teaches the model when to emit and when to wait, without adding policy heads or latency-prediction modules. The result","pith_inferences":["If the result holds, streaming capability can be decoupled from model architecture: any decoder-only speech LLM with causal inference could, in principle, be made simultaneous by re-labeling its training data this way, a cheaper path than redesigning with read/write heads.","A natural testable extension is to apply the same teacher-labeled bounded-waiting construction to text-only simultaneous machine translation; if the gains reproduce, it would suggest the P2P target form, not the speech interface, is the core mechanism.","Because the teacher LLM plans with full access to all chunks, replacing it with a strictly causal teacher that never sees future chunks would reveal how much of the benefit comes from offline planning versus the bounded-waiting constraint itself.","The labels come from a proprietary teacher; an open question is whether an open-weights or weaker teacher would yield similar gains, which would determine how cheaply this recipe scales to more languages and domains."],"forward_implications":["P2P fine-tuning improves streaming COMETKiwi by +1.54 over the streaming baseline (70.50 vs 68.96) at only +0.15s more Average Lagging.","Gains are largest for reordering-heavy directions: EN→JA +3.96, EN→ZH +2.74, and ZH→JA +3.94.","Improvements appear even on directions that were not directly supervised (e.g., ZH→JA), suggesting the method shapes general streaming behavior rather than performing language-pair-specific tuning.","Chunk size Δ and rewind length k act as controllable knobs on the quality–latency frontier, giving practitioners a simple way to trade latency for quality.","On a public CoVoST2 diagnostic set, P2P fine-tuning recovers more than half of the offline-to-streaming degradation in BLEU."],"fun_headline_variants":["Data tweak beats new architecture for streaming speech translation","Rewind-based prefix training adds +1.54 COMETKiwi with no new LLM heads","Teacher-labeled prefixes lift streaming speech output without redesign","No architecture change needed: prefix-to-prefix data boosts SimulST"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The whole approach rests on the teacher LLM's prefix-to-prefix labels being genuinely commit-ready and correctly timed; if the teacher commits prematurely or waits wrongly at chunk boundaries, the student will learn mis-timed emission behavior.","fun_headline_variants_meta":{"raw":{"variants":["Data tweak beats new architecture for streaming speech translation","Rewind-based prefix training adds +1.54 COMETKiwi with no new LLM heads","Teacher-labeled prefixes lift streaming speech output without redesign","No architecture change needed: prefix-to-prefix data boosts SimulST"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000194,"raw_usage":{"total_tokens":1168,"prompt_tokens":700,"completion_tokens":468,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":444,"completion_tokens_details":{"reasoning_tokens":391}},"tokens_in":444,"tokens_out":468,"duration_ms":5361,"temperature":1.0,"reasoning_tokens":391,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T06:01:56.244376+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A controlled ablation replacing the teacher LLM with a naive monotone truncation of the offline reference (no bounded-waiting, no global planning) would decide the matter: if truncation labels recover similar gains, the paper's specific label construction is not the active ingredient; if they do not, the teacher's planning is. A complementary quantitative check is to re-score each committed prefix by asking the full-utterance model how much of the reference's continuation is already contradicted; high contradiction rates would signal premature commit-ready labels.","supporting_citations":[],"review_version":1}