{"id":"4b925dbe-6d38-43a7-8878-305b4d579acb","arxiv_id":"2603.25702","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.5,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Training-free self-speculation reuses a block-diffusion model’s block-size-1 mode as a local AR verifier, improving accuracy–speed tradeoffs over confidence-threshold decoding.","lead":"S2D2 speeds up block-diffusion language models without retraining by using the same model as both parallel drafter and autoregressive verifier. It improves the accuracy–speed tradeoff over confidence-threshold decoding on several public diffusion LLM families.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.5","headline":"Headline multipliers rest on selected configs and an uncalibrated B=1 critic; residual resampling may not systematically improve committed tokens enough to justify the extra pass.","rationale":"The reader correctly isolates the load-bearing assumption: that B=1 AR mode is a sufficiently calibrated sequence-level critic for rejection sampling on diffusion drafts. That is exactly where the argument is least secure—training never couples the two modes as drafter/verifier, and the energy interpretation in Remark 1 is heuristic. I agree the paper is a solid systems contribution with multi-family evidence and linked code, so the verdict should stay CONDITIONAL rather than move to REJECT or ACCEPT. The additional concrete risk is config selection and missing uncertainty: the strongest multipliers are taken from two hand-chosen points per model, while ablations show high sensitivity. A single fixed-policy multi-seed comparison against dynamic BD3 (and an oracle without residual resampling) would settle whether the critic actually improves committed tokens enough to amortize the extra pass, or whether the reported frontier is an artifact of post-hoc operating-point choice. No stronger internal inconsistency was found; the concern is empirical robustness of the calibration assumption, not a formal flaw.","tokens_in":32081,"tokens_out":757,"duration_ms":9043,"concrete_test":"On SDAR-1.7B/8B GSM8K+MBPP (same 200-sample ablation subset), freeze one routing policy a priori (e.g., min-span τ_span=2 or entropy score τ_score=0, c=1) and report mean±std over ≥3 seeds for: (i) S2D2, (ii) dynamic BD3 at the same B, (iii) an oracle that accepts only when q_i ≥ p_i without residual resampling. If under the fixed policy the accuracy–speed gain vs dynamic BD3 disappears or residual resampling underperforms pure rejection (no residual), the headline 1.57×/+4.5 claim does not hold beyond selected configs.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim (Abstract; §5.1 Tables 1–3) is that training-free S2D2 improves the accuracy–speed frontier over strong dynamic confidence baselines, with headline numbers up to 4.7× vs AR and 1.57× vs dynamic with +4.5 avg points. That claim load-bears on Remark 1 / §4.1–4.2: that the same model’s block-size-1 mode supplies a calibrated local critic so that min(1, q_i/p_i) plus residual resampling (P_ℓver − P_ℓ)+ yields better committed tokens than confidence thresholding alone, enough to amortize one extra forward. The model was never trained as a joint drafter–verifier pair; B=1 is only an extreme conditional of the same ELBO family (§A.3), not a matched residual energy model. Empirically the paper reports only two hand-picked operating points per model (config-A/B), no error bars, and non-ideal AR baselines for Fast-dLLM (B=4, SB=1). Ablations (Tables 6–10) show large sensitivity of both accuracy and speedup to τ_span, score thresholds, and estimator choice, so the frontier improvement is not shown to be robust under a fixed, pre-specified policy. If the AR critic is miscalibrated on diffusion drafts, residual resampling can inject worse tokens while still paying the verifier cost—undermining both the quality and the amortization half of the claim.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The paper proposes S2D2, a training-free self-speculative decoding method for block-diffusion language models. The core idea is that reducing block size to one turns the same pretrained model into an autoregressive verifier, so standard block-diffusion decoding can draft tokens in parallel and the B=1 mode can accept or residual-resample them via rejection sampling (Alg. 3; §4.1–4.2). Lightweight routing policies (minimum-span, score-threshold, hysteresis, UCB) decide when the extra verifier pass is worth its cost (§4.3). Empirically, across SDAR (1.7B/4B/8B), Fast-dLLM v2, and LLaDA2.1-Mini on GSM8K, MBPP, HumanEval, and IFEval, S2D2 improves the accuracy–speed frontier relative to static/dynamic confidence-threshold baselines (Tables 1–3; Fig. 3), with reported highlights of up to 4.7× over AR and up to 1.57× over a tuned dynamic baseline with +4.5 average accuracy points on SDAR. Analysis frames verification as local AR-guided residual-energy correction (Remark 1; §4.4, App. A.3).","tokens_in":32508,"tokens_out":1513,"duration_ms":22329,"significance":"If the empirical claims hold under fairer reporting, this is a practically useful contribution: a plug-and-play, training-free acceleration recipe for an emerging class of hybrid AR–diffusion LMs, without auxiliary models or distillation. The observation that the same block-diffusion checkpoint supplies both drafter and verifier is simple and transferable, and the multi-family evaluation (position-aligned and right-shifted architectures; SDAR, Fast-dLLM v2, LLaDA2.1) plus public code strengthen credibility. The complementarity result with LLaDA’s token-editing self-correction (Table 3) is also valuable. The residual-energy interpretation is secondary; the main value is an inference-time method that often improves both speed and accuracy over strong dynamic confidence baselines in the few-step regime.","major_comments":[{"comment":"Tables 1–3 report only two hand-picked operating points (config-A/B, or A/B/C) per model, while Appendix Tables 6–10 show large sensitivity of both accuracy and speedup to τ_span, score thresholds, estimator choice, and hysteresis settings. The Abstract and §5.1 headline claims (e.g., 4.7× vs AR and 1.57× vs dynamic with +4.5 points on SDAR-1.7B config-B) therefore rest on selected configurations rather than a fixed, pre-specified policy. Please either (i) designate one policy and hyperparameter rule a priori and report it across all models/tasks, or (ii) report full Pareto frontiers / multi-seed summaries so the claimed frontier improvement is not confounded by post-hoc config selection.","section":null},{"comment":"The load-bearing assumption in §4.1–4.2 and Remark 1 is that the pretrained B=1 mode is a sufficiently calibrated local critic for min(1, q_i/p_i) acceptance and residual resampling (P_ℓver − P_ℓ)+ on diffusion drafts. App. A.3 correctly notes this is only an in-family AR conditional, not a matched residual energy model. The paper does not quantify calibration (e.g., acceptance-rate vs. actual quality of residual resamples, or q/p reliability by position/progress). Without this, it is hard to know when residual resampling systematically improves committed tokens versus paying for a miscalibrated extra pass. A short diagnostic—acceptance rates, residual-resample win rates, or quality of accepted vs. rejected prefixes—would substantially strengthen the central mechanism claim.","section":null},{"comment":"Main results (Tables 1–3) report point estimates with no error bars, seeds, or variance over sampling/decoding stochasticity. Given that ablations already show nontrivial accuracy swings (e.g., Table 6, B=32: Avg from ~67.5 to ~76.5 depending on τ_span), statistical uncertainty is material to the “consistently improves” claim in the Abstract and §5.1. Please add multi-seed or bootstrap intervals on at least the primary SDAR and LLaDA settings, or otherwise justify that single-run differences of a few points are stable.","section":null},{"comment":"For Fast-dLLM v2 (Table 2), speedups are relative to B=4, SB=1 rather than true B=1 AR, because B=1 is “unreliable.” That is understandable, but it weakens cross-family comparability of the “× over AR” narrative used in the Abstract. Please state this limitation more prominently in §5.1 and avoid aggregating Fast-dLLM multipliers with SDAR’s true-AR baselines without qualification.","section":null}],"minor_comments":[{"comment":"Figure 1 is dense; the routing-policy panel (d) and dual attention-mask layouts (c) would benefit from a short caption walkthrough of one full step (draft → route → verify/fallback).","section":null},{"comment":"Notation for draft vs. verifier logits (ℓ vs. ℓ_ver) and residual (P_ℓver − P_ℓ)+ is clear in Alg. 3 but less so in the main text of §4.1; a one-line definition of residual resampling there would help.","section":null},{"comment":"Table 3 is labeled “preliminary” for LLaDA2.1-Mini and covers only GSM8K/MBPP; either expand to HumanEval/IFEval for parity with Tables 1–2 or soften the Abstract’s LLaDA claim to match the limited scope.","section":null},{"comment":"App. A.5 notes that better acceptance-length estimators (hard margin) do not yield better task accuracy than soft entropy; this is an interesting finding and deserves a sentence in the main §5.2 rather than only the appendix.","section":null},{"comment":"Minor polish: “aspeed-firstquestion” spacing in §1; consistent use of “block-size-1” vs. “block size 1”; ensure all appendix table references in §5.2 match the final numbering.","section":null}],"recommendation":"minor_revision","confidential_remarks":"Solid engineering paper with a clean idea and unusually broad multi-family evaluation for this subarea. I would not reject on novelty grounds: the B=1 self-verifier reuse is a genuine, practical contribution relative to ASSD/EDLM/SSD. The main risk is overstated robustness of headline multipliers; if the authors fix reporting (fixed policy or full frontiers + uncertainty), this is a clear accept-level methods contribution for a systems/ML venue. Scope fits cs.CL / inference-acceleration well."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The useful takeaway is simple: for existing block-diffusion LMs, you can treat the same checkpoint’s B=1 mode as a local AR verifier, run standard speculative acceptance on the first contiguous mask span, and route so you only pay the extra forward when it is likely to pay off. That is genuinely new as a finished, plug-and-play procedure. Prior speculative work either needs a separate draft model, any-subset architecture (ASSD), hierarchical batching, or extra training (EDLM). S2D2 reuses what the model already has.\n\nWhat they do well: clear algorithms (1–4), the 2L verifier mask for position-aligned models, lightweight routing (min-span, score, hysteresis, bandit), and evaluation across three families (SDAR 1.7/4/8B, Fast-dLLM v2, LLaDA2.1-Mini) on GSM8K/MBPP/HumanEval/IFEval. Code is linked. The residual-energy remark is just a rewrite of the usual acceptance ratio—not a new theorem—but it is honest about hybrid, local correction rather than global AR equivalence. Ablations on estimators and policies are extensive. On LLaDA they show complementarity to token editing, which is the right check.\n\nSoft spots, in proportion: the stress-test is half right. Headline numbers (4.7× vs AR, 1.57× vs dynamic with +4.5 points) come from two hand-picked configs per model, without error bars, and Fast-dLLM’s AR baseline is B=4/SB=1 because B=1 is unreliable. Routing is sensitive (Tables 6–10). The B=1 critic is not a matched residual model—only an extreme conditional of the same family—so calibration is an assumption, not a guarantee. That said, the paper does not hide this; it reports both accuracy- and speed-oriented points, shows flatter ITS curves at large B, and still beats strong dynamic confidence baselines in most settings. The central engineering claim (better accuracy–speed frontier without retraining) is supported even if the exact multipliers are not universal.\n\nThis is for people shipping or studying diffusion-LM inference, not for foundational LM theory. Math is light but appropriate; citations are fair; circularity is low. I would send it to peer review. Engage if you care about practical decoding for this model class; skip if you only want new training objectives or distribution-preserving theory.","headline":"Training-free self-speculation for block-diffusion LMs is a real, usable systems idea; the headline multipliers are config-selected but the multi-family evidence still holds up.","tokens_in":33211,"tokens_out":615,"would_cite":true,"duration_ms":7625,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"The same block-diffusion language model can draft tokens in parallel and verify them in autoregressive mode, yielding faster and often more accurate decoding with no extra training.","keywords":["block diffusion","diffusion language models","speculative decoding","self-speculation","training-free acceleration","confidence thresholding","rejection sampling","hybrid decoding"],"falsifier":"On the same models and benchmarks, run always-on and routed S2D2 against carefully tuned dynamic confidence baselines at matched wall-clock budget: if average accuracy never rises (or falls) while speed stays flat or worse—especially at large block sizes—the claim that AR self-verification is a useful local critic collapses.","tokens_in":32913,"feed_emoji":"⚡","tokens_out":934,"duration_ms":19000,"temperature":0.7,"pith_summary":"Block-diffusion language models try to beat pure left-to-right generation by finishing whole blocks with parallel denoising, but in the few-step regime needed for real speedups, simple confidence thresholds are brittle: set them high and quality falls, set them low and you waste steps. This paper argues that you do not need a second model or more training to fix that. Because shrinking the block size to one turns the same network into an autoregressive model, that network can draft a block with diffusion and then act as its own local sequence-level critic via standard speculative rejection sampling. Lightweight routers decide when the extra verifier pass is worth the cost. Across three major model families, the resulting hybrid trajectory improves the accuracy–speed frontier over strong dynamic confidence baselines, with large speedups over pure autoregressive decoding and, in several settings, higher accuracy as well.","feed_headline":"Same model drafts and verifies for faster diffusion LLMs","feed_subtitle":"Training-free self-speculation beats confidence thresholds on speed and often accuracy","key_machinery":"S2D2 (training-free self-speculative decoding): diffusion drafts tokens and confidences; the same model under block-size-1 masking supplies verifier probabilities; tokens in the first contiguous masked span are accepted left-to-right with probability min(1, q_i/p_i), with residual resampling on the first rejection; routing policies (minimum-span, score-threshold, hysteresis, optional bandit) decide when to pay for the verifier pass.","core_discovery":"A pretrained block-diffusion language model already contains both a parallel drafter (standard block decoding) and an autoregressive verifier (block size one). Inserting speculative acceptance of the first contiguous masked span, gated by cheap routing policies, produces a training-free hybrid decoder that is often both faster and more accurate than tuned confidence-threshold diffusion decoding.","pith_inferences":["If block-size-one scores are only weakly calibrated, soft tempering of the acceptance ratio or better draft-probability estimators may matter more than fancier routers.","The same self-verification idea may transfer to other hybrid AR–diffusion or any-order generators that expose a cheap causal mode of the same weights.","Routing policies that track acceptance reward online could eventually replace fixed thresholds for long generations where draft quality drifts.","When draft and verifier disagree systematically on certain token classes (code vs math), task-conditioned routing may be a natural next control knob."],"forward_implications":["Existing block-diffusion checkpoints can be accelerated at inference without distillation, auxiliary verifiers, or architecture changes.","Large-block few-step regimes that were previously unstable under pure confidence thresholding become usable operating points.","Built-in token-editing self-correction (as in some diffusion LMs) can be stacked with S2D2 rather than replaced by it.","Verification cost can be amortized by simple length- or score-based routers, so the method need not verify every step.","Speculative acceptance supplies a local AR-vs-diffusion residual-energy check that steers drafts without global multi-sample reweighting."],"fun_headline_variants":["Same model drafts parallel tokens and verifies as AR critic","Training-free self-speculation for faster block-diffusion LLMs","S2D2: model as both drafter and verifier in hybrid decoding","Block-diffusion self-speculation beats confidence thresholds","One pretrained model drafts and verifies for speed gains"],"cache_read_input_tokens":16256,"weakest_assumption_plain":"The model’s block-size-one mode is a good enough local critic of its own diffusion drafts that rejection sampling plus residual correction improves committed tokens enough to repay the extra forward pass.","fun_headline_variants_meta":{"raw":{"variants":["Same model drafts parallel tokens and verifies as AR critic","Training-free self-speculation for faster block-diffusion LLMs","S2D2: model as both drafter and verifier in hybrid decoding","Block-diffusion self-speculation beats confidence thresholds","One pretrained model drafts and verifies for speed gains"]},"model":"grok-4.5","effort":"low","cost_usd":0.00458,"raw_usage":{"total_tokens":1390,"prompt_tokens":844,"num_sources_used":0,"completion_tokens":65,"cost_in_usd_ticks":45800000,"prompt_tokens_details":{"text_tokens":844,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":481,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":844,"tokens_out":65,"duration_ms":4324,"temperature":1.0,"reasoning_tokens":481,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-13T18:03:25.948148+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"On the same models and benchmarks, run always-on and routed S2D2 against carefully tuned dynamic confidence baselines at matched wall-clock budget: if average accuracy never rises (or falls) while speed stays flat or worse—especially at large block sizes—the claim that AR self-verification is a useful local critic collapses.","supporting_citations":[],"review_version":1}