{"id":"32868545-43d0-4d10-8816-ebab538bfcff","arxiv_id":"2608.02942","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":8,"one_line_summary":"OPTD selects the longest prefix of future token commits that preserves a frozen teacher's rollout outcome on on-policy student states, improving few-step diffusion LM speed-accuracy trade-offs.","lead":"OPTD speeds up diffusion language model decoding by training the student on its own visited states and using a frozen teacher to verify which future tokens can be committed together. It reports the best average quality-constrained AUP across four math and code benchmarks, improving efficiency with a modest accuracy trade-off.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The headline AUP advantage cannot be verified: the metric's endpoint-selection protocol and exact formula are deferred to a missing appendix, main results are single-seed, and no code is released. This leaves the central 313.18-vs-245.59 claim unsupported until the evaluation is reproducible.","rationale":"The reader's formal 'weakest_assumption' concerns the teacher-outcome proxy, but the reader's rationale also flags the endpoint-selection protocol and missing appendix/code. I take the AUP verification gap as the single most load-bearing concern because it is attached directly to the paper's headline quantitative claim: a 313.18 average AUP that is 27.5% above the best prior method. If the endpoint-selection rule is unfair, unreproducible, or post hoc, that claim fails regardless of whether the teacher-proxy assumption is sound. The teacher-proxy issue is real and honestly scoped by the paper—Lemma 1 covers training-target consistency, not inference correctness—and it would affect any gold-free distillation method, so it is less decisive for the specific AUP claim. In contrast, the AUP metric is nonstandard, the exact formula is deferred to an unavailable appendix, the main checkpoint is single-seed, and no code is released; these are exactly the conditions under which a curve-comparison headline cannot be checked. Therefore the verdict should remain CONDITIONAL pending release of the appendix/code and a fixed, pre-registered AUP recomputation, rather than being upgraded to ACCEPT or downgraded to REJECT on the current evidence.","tokens_in":13220,"tokens_out":8389,"duration_ms":84601,"concrete_test":"Release the technical appendix and AUP evaluation code; recompute Table 1 with a fixed, pre-registered endpoint rule: for every method, build the accuracy-vs-TPF curve by sweeping τ in {0.75,0.80,0.85,0.90} under the same decoder, anchor at TPF=1.00, and integrate up to each task's fastest endpoint whose accuracy is within 5 points of the anchor. If the reproducible OPTD average AUP falls below TAD-S or d3LLM, or if the gap is much smaller than 27.5%, the central claim must be revised. Also run the main checkpoint with 3-5 seeds and report AUP variance.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The strongest claim—OPTD's 313.18 average quality-constrained AUP, 27.5% above TAD-S—is a comparison of integrated accuracy-versus-TPF curves. The paper defines AUP in a nonstandard way: it anchors at TPF=1.00 accuracy, integrates trapezoidally beyond the anchor, and 'may use a quality-constrained threshold-sweep endpoint from the same decoder family,' with the exact formula, reference accuracies, and endpoint policy deferred to a missing technical appendix (Experiments, Main Results). The main checkpoint is single-seed, and no evaluation code is released. In such a curve comparison, the endpoint-selection rule is load-bearing: if OPTD's per-task endpoints were selected after inspecting results (the text only fixes the displayed operating point at τ=0.8, not the curve endpoints), the reported margin could shrink or reverse under a pre-registered rule. Lemma 1 does not mitigate this: it only guarantees training targets are teacher-outcome-preserving, and says nothing about the AUP evaluation. The concern is a reproducibility/verification gap, not an internal inconsistency in the method.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes OPTD (On-Policy Transition Distillation), a method for few-step distillation of masked diffusion language models. The method samples partial states from the evolving student's own rollout, uses a frozen, question-only teacher to generate a continuation, and then selects the longest prefix of confidence-ordered candidate tokens whose joint commitment preserves the teacher's rollout outcome. A set-bottleneck certainty loss promotes the verified candidates to the decoder's release threshold, while a frozen-teacher KL anchor regularizes the remaining active positions. The authors report results on GSM8K, MATH-500, HumanEval, and MBPP, claiming that OPTD achieves the strongest quality-constrained AUP (313.18 average), which is 27.5% above the best prior method TAD-S (245.59). They also provide controlled ablation studies isolating the effects of on-policy state sampling, joint verification, set reduction, and adaptive compression horizons.","tokens_in":13461,"tokens_out":5781,"duration_ms":55254,"significance":"If the results hold, the paper makes a useful contribution to few-step diffusion language model decoding: it explicitly addresses the state-distribution mismatch between off-policy training targets and student-visited states, and it introduces a state-dependent compression rule with an outcome-consistency check. The paper is careful in several respects: it provides a decoder-matched attribution study (Table 3), it reports controlled ablations with matched evaluation budgets, and it explicitly scopes the formal guarantee as applying to training-target construction rather than to inference-time actions. The paper also honestly notes its single-seed evidence and the task-dependence of the transfer results. However, the central quantitative claim depends on an AUP metric whose exact formula and endpoint-selection protocol are deferred to a technical appendix that is not included, and no evaluation code is released. These omissions prevent verification of the headline 27.5% margin as submitted.","major_comments":[{"comment":"The headline claim that OPTD achieves the strongest quality-constrained AUP (313.18 average, 27.5% above TAD-S) cannot be verified from the submitted manuscript. The text defines AUP only by saying it uses α=3, an unweighted TPF-1 accuracy anchor, trapezoidal integration beyond that anchor, and 'may use a quality-constrained threshold-sweep endpoint from the same decoder family'; the exact formula, reference accuracies, and endpoint-selection policy are deferred to a technical appendix that is not included. Since AUP is a curve comparison, the endpoint rule is load-bearing: under a different pre-registered endpoint rule the reported margin could change materially. Please include the complete metric definition and protocol in the paper, provide per-task AUP curves and reference accuracies, and release the evaluation code.","section":"Experiments, Main Results (Table 1)"},{"comment":"Lemma 1 (Fallback consistency and verification cost) is load-bearing for the training-target construction: it guarantees that the selected action preserves the teacher's rollout outcome. The proof and the formal statement of its scope are deferred to the technical appendix, which is not included. Please include a complete proof in the submission, including the definitions of OutcomeT, Commit, and the token-wise equality check, and state explicitly the determinism and tie-breaking assumptions used for Eq. (5).","section":"The OPTD Algorithm, Lemma 1"},{"comment":"The primary checkpoint, the state-source attribution in Figure 3, the matched loss ablation in Table 2, and the adaptive-vs-fixed comparison in Table 5 are all single-seed runs, and no confidence intervals are reported for the headline AUP comparison. The text itself acknowledges that Figure 3's curve is 'single-seed matched learning-curve evidence rather than a variance estimate.' Given that the central contribution is a quantitative AUP margin of 27.5%, please report at least three independent seeds (or example-level bootstrap intervals for the fixed checkpoint) for the main comparison and the key ablations.","section":"Experiments, Ablation Experiments; Analysis Experiments"},{"comment":"In Eq. (5), OPTD selects the longest candidate prefix that preserves the frozen teacher's rollout outcome. Because neither target construction nor training uses gold responses, a poor or systematically biased teacher outcome will be preserved and trained toward, and the paper does not analyze how often the teacher outcome agrees with the gold answer on the benchmark tasks. Please add a quantitative analysis of teacher-outcome correctness (e.g., agreement with gold on GSM8K/MATH-500/HumanEval/MBPP) and an error analysis showing that OPTD's accuracy losses do not come disproportionately from states where the teacher outcome is incorrect; if needed, add a gold-free calibration or filtering criterion.","section":"The OPTD Algorithm, Eq. (5)"}],"minor_comments":[{"comment":"The d3LLM MBPP entry '39.205.4682.20' appears to be a formatting error, with two numbers concatenated; please fix.","section":"Table 1"},{"comment":"The decoding configuration 'confidence Multi-Block decoding' and the block-add/decoded-token thresholds are not defined in the main text; define them or provide the appendix reference in the submission package.","section":"Experiments, Setup"},{"comment":"The phrase '512 two-rank updates' is unclear; if it means 512 rank-2 LoRA updates, please state that explicitly.","section":"Experiments, Ablation Experiments"},{"comment":"Table 5 reports single-seed results; the caption should state this clearly and qualify the claim of 'higher TPF at every matched horizon' accordingly.","section":"Analysis Experiments, Adaptive compression horizon"},{"comment":"Some notation is introduced before it is fully defined: for example, Aτθ and G(x) are used in Eq. (1) and the surrounding text before their components are explained. Please reorder or add definitions for readability.","section":"Preliminaries"}],"recommendation":"major_revision","confidential_remarks":"The most important editorial issue is the missing technical appendix and evaluation code; without them the headline AUP comparison is not checkable. I would advise requesting the appendix and code as part of the revision rather than relying on the main text alone."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. First, OPTD is a genuinely new combination: on-policy state sampling for dLLM transition distillation, a state-dependent merge depth chosen by checking that a joint token commitment preserves a frozen teacher's rollout outcome, and a set-bottleneck certainty loss. The controlled ablations are well designed, and the paper is candid about what its guarantee does and does not cover. Second, the headline number—313.18 average AUP, 27.5% above TAD-S—cannot be checked from the manuscript. The AUP formula, reference accuracies, and endpoint-selection rule are all deferred to a technical appendix that is not in the arXiv posting, main results are single-seed, and no code is released. That is load-bearing, because the metric anchors at TPF=1 and integrates over a threshold sweep whose endpoints are described as \"may use\" a quality-constrained rule. On a curve comparison, endpoint choice can move the integrated value substantially. The margin may well be real; we just cannot verify it.\n\nWhat is genuinely good: the state-source ablation shows on-policy accuracy improving monotonically over training (48.51 vs 46.62 at 1,024 updates) while TPF stays matched; the verification-unit ablation (10 of 416 individually valid tokens fail the joint check) cleanly motivates joint testing; and the transfer-from-vanilla-LLaDA result gives a frank picture of how much of the gain comes from the TAD-S initializer. The paper also scopes Lemma 1 correctly: it guarantees teacher-outcome-preserving training targets, not inference-time consistency, and says so explicitly.\n\nSoft spots, in proportion. The missing appendix is the main problem. The endpoint policy is genuinely vague. Single-seed results leave the 27.5% claim without variance estimates. Also, the headline AUP is a complete-stack comparison; the decoder-matched attribution shows a much more modest +0.67 accuracy and +4.4% TPF, so the large AUP gap is not purely an effect of the distilled policy—though the paper does say this. A minor point: the consistency criterion is self-referential to the teacher's own rollout, so a poor teacher could preserve incorrect completions. The paper acknowledges inference quality remains empirical, which is fair.\n\nBottom line: this deserves a serious referee, but the evaluation reproducibility has to be fixed before the headline claim can be taken at face value. If you send it out, require the appendix, the code, and at least a few seeds.","headline":"Solid method and unusually honest scoping, but the headline AUP claim is unverifiable as written because the metric formula, endpoint rule, and code are all missing.","tokens_in":14031,"tokens_out":2649,"would_cite":true,"duration_ms":25338,"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":"On-policy transition distillation with consistency-guided adaptive compression improves few-step diffusion language model decoding by training on student-visited states and selecting the longest outcome-preserving joint action, achieving…","keywords":["diffusion language models","few-step distillation","on-policy distillation","adaptive compression","parallel decoding","consistency verification","quality-constrained AUP","transition distillation"],"falsifier":"Compare completions trained from verified merges against gold-answer quality on states where the frozen teacher's own rollout outcome is incorrect; if verified merges are no more accurate than unverified ones on those states, the teacher-outcome proxy is too coarse. A direct audit: on a held-out set, measure the fraction of merges that pass the outcome check but whose jointly committed tokens are later changed or undone by the decoder, and compare that fraction with the 10-of-416 failure rate reported for independent verification.","tokens_in":13019,"feed_emoji":"⚡","tokens_out":4289,"duration_ms":36688,"temperature":0.7,"pith_summary":"This paper claims that few-step distillation for diffusion language models has been held back by a mismatch: supervision is built on teacher trajectories, but at inference the student visits different partial states, especially when compression is aggressive. The proposed method, OPTD, trains on partial states sampled from the student's own rollouts and adaptively decides how many future tokens to commit in one transition by keeping the longest prefix of teacher-proposed tokens whose joint commitment preserves the frozen teacher's rollout outcome. This matters because it directly addresses the state-action mismatch that limits parallel decoding in dLLMs, and the paper reports consistent gains in the quality-efficiency trade-off across mathematical reasoning and code-generation benchmarks without using any gold response.","feed_headline":"OPTD tops few-step diffusion LLM speed-quality scores by 27.5%","feed_subtitle":"Training on student-visited states and verifying joint merges wins on four reasoning and code benchmarks.","key_machinery":"The central mechanism is consistency-guided adaptive compression, which decides at each student-visited state how many teacher transitions can be merged into one student transition. Candidate future positions must already predict their value in the teacher outcome, are ordered by teacher confidence, and are accepted only as a set: the chosen action is the longest prefix whose joint commitment leaves the frozen teacher's rollout outcome unchanged. The training objective is a set-bottleneck certainty loss, a max over positive gaps to the release threshold for the verified set, plus a frozen-teacher KL anchor over all other active positions. Lemma 1 guarantees that the fallback action, the teacher's ordinary first release, always preserves the outcome, so the selected target is always outcome-preserving under the frozen teacher.","core_discovery":"The central claim is that off-policy supervision in few-step dLLM distillation mismatches the states the student actually visits, and that this mismatch can be fixed by collecting partial states from the student's own rollouts and compressing only as far as a frozen teacher's outcome stays unchanged. From each student-visited state, the frozen teacher rolls out a completion, candidate future tokens are ordered by current-state confidence, and the longest prefix whose joint commitment reproduces the teacher's outcome becomes the training action. The verified set is trained with a set-bottleneck certainty loss that pushes every member past the decoder release threshold, while all other active positions are anchored to the frozen teacher by KL divergence. The paper reports that OPTD reaches 47.21% average accuracy and 7.87 tokens per forward across GSM8K, MATH-500, HumanEval, and MBPP, with an average quality-constrained AUP of 313.18, 27.5% above the best prior method, without ever observing a gold response.","pith_inferences":["Since the outcome-preservation guarantee covers only training-time targets, a natural extension is an inference-time consistency check that rejects student actions whose joint commitment changes a frozen teacher rollout, which could improve calibration of the learned release policy.","The verification cost scales with the number of candidate prefixes, but batched counterfactual checks could make adaptive compression practical for larger block sizes or higher parallelism regimes.","The observation that only 10.1% of states support an 8x compression span suggests an upper bound on achievable parallelism from this recipe; combining OPTD with stronger initializers or different decoder families may lift that cap.","The transfer experiment from base LLaDA retains much less parallelism than the TAD-S-initialized model, implying that most cross-block parallelism comes from the initializer; a matched-initializer study would be needed to separate the backbone effect from OPTD's own contribution."],"forward_implications":["Distillation target construction for dLLMs should be state-dependent rather than fixed-horizon, since the useful compression depth varies across student-visited states.","Training on student-visited states recovers accuracy without sacrificing efficiency relative to teacher-state supervision, as shown by matched learning curves.","Joint verification of token sets is necessary: independently valid tokens can still fail when committed together, so per-token checks are an insufficient safety filter.","The method constructs targets without gold responses, making it applicable where reference answers are unavailable or expensive.","Adaptive horizon selection reaches operating points that fixed horizons cannot, so the compression extent itself becomes a learned quality-efficiency control."],"supporting_citations":[{"why":"Supplies the TAD-S checkpoint used to initialize both student and frozen teacher, and serves as the best prior baseline for comparison.","marker":"(Zhou et al. 2026)"},{"why":"Provides the LLaDA-8B-Instruct base architecture and the vanilla dLLM baseline whose accuracy anchors the AUP computation.","marker":"(Nie et al. 2025)"},{"why":"Defines the d3LLM pseudo-trajectory distillation baseline and the AUP evaluation protocol that the paper adopts.","marker":"(Qian et al. 2026)"},{"why":"Supplies the Fast-dLLM training-free acceleration baseline, demonstrating the system-level speedups that OPTD builds on.","marker":"(Wu et al. 2025)"},{"why":"Introduces the on-policy distillation formulation that OPTD extends from autoregressive models to masked diffusion language models.","marker":"(Agarwal et al. 2023)"},{"why":"Contributes the progressive-distillation view of compressing multiple teacher steps into a single student step, which OPTD adapts to a state-dependent setting.","marker":"(Salimans and Ho 2022)"}],"fun_headline_variants":["On-policy steps close diffusion LLM distillation gap","Student rollouts improve few-step diffusion LLM training","Frozen teacher verifies merges: OPTD beats prior few-step methods","Adaptive on-policy compression lifts diffusion LLM speed-quality"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that a frozen teacher's rollout outcome from a student-visited partial state is a reliable and sufficiently fine-grained proxy for correctness, so any merged action that preserves that outcome is safe to train toward.","fun_headline_variants_meta":{"raw":{"variants":["On-policy steps close diffusion LLM distillation gap","Student rollouts improve few-step diffusion LLM training","Frozen teacher verifies merges: OPTD beats prior few-step methods","Adaptive on-policy compression lifts diffusion LLM speed-quality"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00066,"raw_usage":{"total_tokens":3059,"prompt_tokens":1027,"completion_tokens":2032,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":643,"completion_tokens_details":{"reasoning_tokens":1962}},"tokens_in":643,"tokens_out":2032,"duration_ms":14940,"temperature":1.0,"reasoning_tokens":1962,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T14:54:13.859964+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compare completions trained from verified merges against gold-answer quality on states where the frozen teacher's own rollout outcome is incorrect; if verified merges are no more accurate than unverified ones on those states, the teacher-outcome proxy is too coarse. A direct audit: on a held-out set, measure the fraction of merges that pass the outcome check but whose jointly committed tokens are later changed or undone by the decoder, and compare that fraction with the 10-of-416 failure rate reported for independent verification.","supporting_citations":[],"review_version":2}