{"id":"6910c626-4ea4-48ef-a4b6-895f3e1799fc","arxiv_id":"2608.05687","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"Answer-first token commitment and answer-only collapse explain why unconstrained diffusion LLM decoding fails on reasoning tasks, and a frontier-window gate recovers the gap.","lead":"This paper shows that diffusion language models, which can write tokens in any order, often write the answer before their reasoning when solving math problems, and sometimes skip reasoning entirely. A one-line decoding change that restricts the order of token writing, with no retraining, restores most of the lost accuracy.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"EOS pressure is context-dependent, so equal pressure across decoders does not establish equal termination beliefs; the 'not beliefs, but reachability' claim in §3.3 is under-supported.","rationale":"The paper's headline contribution is a mechanistic claim: pure decoding fails because of reachability, not termination beliefs. The evidence is the pressure-commitment dissociation in §3.3. But the pressure measure is conditional on each decoder's own context, which diverges by design; the model's EOS tendency at masked positions is not a decoder-invariant latent. The gating experiment proves sufficiency of reachability but not the irrelevance of beliefs. This is exactly the reader's weakest assumption. A clean test is to measure EOS pressure under the gated decoder: if gating leaves pressure unchanged while eliminating collapse, the dissociation survives; if gating reduces pressure, the 'not beliefs' claim is weakened. The verdict remains CONDITIONAL as the reader recommended; the fix itself is well supported, but the causal explanation needs support or revision.","tokens_in":15417,"tokens_out":6188,"duration_ms":58390,"concrete_test":"Run the gated decoder (w=32, L=512) on the same 250 paired GSM8K problems and record the per-step EOS pressure trajectory, comparing it with the pure and semi-AR trajectories already logged. If the gated pressure stays within ~0.03 of the pure pressure at every step while collapse drops from 90.8% to 2–8%, then the dissociation is robust and reachability is the active ingredient. If the gated pressure is substantially lower (e.g., more than 5 points below pure at the same normalized step), then the gate's benefit may be mediated by a change in termination beliefs, and the 'not beliefs' claim in §3.3 would need to be withdrawn or re-derived with a context-matched pressure measurement.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central mechanistic claim is that pure decoding collapses because of reachability, not because the model's termination beliefs differ (§3.3, Discussion, abstract). This rests on measuring EOS pressure — the fraction of masked positions whose top-1 prediction is EOS — and finding it nearly identical across pure and semi-AR decoders (0.782 vs 0.766 at L=512) while EOS commitment differs by 2×. But EOS pressure is a function of the model's context, and at any given step the two decoders have different committed contexts and different masked-position sets. Under pure decoding, the context already contains distant EOS tokens by the time pressure is measured; under semi-AR it does not. The model's top-1 prediction at a masked position is therefore computed under different conditioning in the two decoders, so equal pressure does not establish equal termination beliefs. The gating experiment (§6) shows that restricting reachability is sufficient to recover accuracy, but it does not show that beliefs are irrelevant; the gate also changes the context and thus could change the beliefs that drive EOS commitment. Without a clean belief measurement, the headline claim 'the cause is not the model's termination beliefs' overreaches. The practical fix (frontier gating) is supported, but the explanatory reframing that distinguishes this paper from prior EOS-bias work is not.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies masked diffusion language models and argues that their advertised any-order commitment freedom is the cause of poor reasoning performance. Using an instrumented decoding loop on LLaDA-8B-Instruct and Dream-7B, the authors document two failure modes under unconstrained (pure) decoding: answer-only collapse, where the model emits a bare answer followed by EOS padding on up to 90% of GSM8K problems at L=512, and answer-first commitment, where the final answer is committed early and the derivation is backfilled around it. A preregistered-style 2x2 prompt-by-decoder design on 250 paired problems shows that chain-of-thought helps only under ordered commitment (interaction +34.8 pp, 95% CI [26.8, 42.8]), with no decoder difference in the answer-only condition; the pattern replicates on Dream-7B and MATH-500. A one-knob intervention, frontier-gated commitment, restricts which masked positions are eligible and recovers most of the accuracy gap without training, while a compute sweep shows the optimal window flips from w=1 at full refinement to unconstrained at low NFE. The paper concludes that the cause is reachability, not the model's termination beliefs, and reframes windowing methods as fixes for a reasoning pathology.","tokens_in":15641,"tokens_out":6694,"duration_ms":63377,"significance":"If the central claim holds, this is an important and useful reframing of diffusion LLM decoding: it identifies commitment order as the causal axis for chain-of-thought effectiveness, provides a cheap decoding-time intervention, and connects several previously separate observations about EOS bias, windowed decoding, and backfilled rationalization. The paper has notable strengths: the sampler is verified token-identical to the reference implementation, the 2x2 design and the gating sweep are specified before measuring the target results, two failed predictions are reported, robustness checks include stochastic sampling and prompt paraphrase, and the headline interaction is replicated on a second model and a second dataset with a direct-column near-tie. The main weakness is the belief-versus-reachability dissociation in Section 3.3, which is used to support the abstract's causal claim but rests on a context-dependent comparison. The gating result itself is well supported, but the paper's stronger explanatory claim needs either a cleaner belief measurement or a more modest statement.","major_comments":[{"comment":"The load-bearing dissociation between termination beliefs and reachability is not established by the reported EOS-pressure comparison. EOS pressure is defined at step t as the fraction of still-masked positions whose top-1 prediction is EOS, but at the same step the pure and semi-AR decoders are conditioning on different committed contexts and different masked-position sets, so near-equal pressure values (0.782 vs 0.766 at L=512) do not imply near-equal termination beliefs. The text also does not state at which step or over which problem subset these values are computed, and pressure varies over the trajectory. The gating experiment in §6.1 shows that restricting reachability is sufficient to recover accuracy, but because the gate changes the context on which later predictions are conditioned, it does not by itself show that termination beliefs are irrelevant. As written, the abstract's claim that 'the cause is not the model's termination beliefs' overreaches the evidence. Please either provide a matched-context measurement of termination beliefs (for example, by measuring EOS predictions under identical contexts for both policies) or revise the claim to say that reachability is the manipulated variable and the belief question remains open.","section":"§3.3"},{"comment":"The two-channel decomposition is not cleanly identified. The collapse-channel estimate of +13.6 pp is obtained by shrinking the canvas from L=512 to L=256 and attributing the interaction difference to the reduction in collapse rate, with the assertion that this changes 'little else.' But halving L also halves the number of denoising steps, changes the EOS-pressure profile, and shifts the commitment-order statistics (t* changes from 0.240 to 0.147 in Section 4.1). The residual order channel is separately supported by the non-collapsed comparison (0.478 vs 0.794, n=316, p≈3e-20), so the central interaction stands, but the decomposition as stated is a length-dependent effect rather than a pure collapse-channel estimate. Please re-label this quantity or provide a manipulation that changes collapse without changing canvas length, steps, and pressure.","section":"§5.2"}],"minor_comments":[{"comment":"The paper should report the step index, the aggregation window, and the dispersion or confidence intervals for the EOS-pressure values; the single pair 0.782 vs 0.766 is currently presented without enough detail to assess the near-identity claim.","section":"§3.3"},{"comment":"The checkpoints are not pinned to revisions and were loaded from default branches; providing commit hashes, or an explanation of why they are unavailable, would strengthen reproducibility.","section":"Appendix B"},{"comment":"The claim that wall-clock time is linear in NFE with R²=1.00 is based on four measured NFE points; please state the number of repetitions and the standard errors around the timing measurements.","section":"§6.2"},{"comment":"The heading 'Dream-7B-7B' appears to contain a duplicated model name and should be corrected.","section":"Appendix C"},{"comment":"The statement that predictions were 'fixed before the runs' would be more verifiable with a dated preregistration or an immutable log; as written it is an assertion about procedure.","section":"§5.1"}],"recommendation":"major_revision","confidential_remarks":"The paper is unusually transparent about failed predictions and implementation details, and the gating intervention is likely a valuable practical contribution. My main concern is that the belief-versus-reachability framing in the abstract and Discussion goes beyond what the EOS-pressure measurement can support. If the authors add a matched-context belief measurement or appropriately soften the causal claim, I would view the paper as suitable for publication. The remaining issues are local and can be addressed in revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Main take: this is a genuinely useful paper. It instruments the commitment log of masked diffusion LLMs and shows that unconstrained decoding on GSM8K freezes the answer before the reasoning exists, then backfills a derivation around it, and that collapse to answer-only outputs grows with canvas length. The 2x2 prompt-by-decoder design is clean, the interaction is large and replicated on a second model and dataset, and the frontier-gating intervention is a single-knob manipulation that recovers the gap without training. The paper also reports its failed predictions, verifies its sampler is token-identical to the reference implementation, and ships a frozen artifact manifest. That is real evidence and deserves credit.\n\nWhat is new: prior work knew block-wise decoding beat unconstrained decoding on GSM8K and that window samplers existed for efficiency, but the commitment-order instrumentation, the answer-first backfill diagnosis, and the framing of the window as a manipulated variable rather than an efficiency trick are the paper's own contributions. The dissociation between collapse and order channels is nicely separated by the canvas-length manipulation.\n\nThe weak spot is the mechanistic headline. The paper says the cause is \"not the model's termination beliefs\" because EOS pressure is nearly identical across decoders while EOS commitment differs. But pressure is measured at the same step under different decoders, which have different masked-position sets and different already-committed contexts. Equal pressure under different conditioning does not establish equal beliefs. The gating experiment shows reachability is sufficient to recover accuracy, but it does not show beliefs are irrelevant; the gate also changes the context and could change the beliefs that drive EOS commitment. That overreach should be fixed by either measuring pressure under matched contexts or softening the claim to \"reachability is sufficient and the practical fix works.\" The practical result survives; the explanatory reframing is under-supported.\n\nMinor points: the main paired comparisons use the first 250 GSM8K problems; replication on Dream-7B and MATH-500 helps, but the subset is still worth a robustness check on a random sample. The semi-AR answer-last row is correctly flagged by the authors as a by-construction baseline, so no issue there.\n\nWho this is for: anyone working on diffusion LLM decoding, reasoning evaluation, or chain-of-thought faithfulness. It deserves a serious referee. I would send it to review and ask for a revision of Section 3.3 rather than a desk reject.","headline":"A transparent, well-designed empirical study showing commitment order controls whether chain-of-thought helps in diffusion LLMs, with a strong practical fix and one overreaching mechanistic claim.","tokens_in":16235,"tokens_out":1664,"would_cite":true,"duration_ms":17875,"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":"Commitment order, not EOS beliefs, is the causal axis of reasoning failure in masked diffusion language models, and a decoding-time reachability constraint recovers chain-of-thought without retraining.","keywords":["diffusion language models","commitment order","chain-of-thought","EOS pressure","reachability","frontier gating","reasoning collapse","GSM8K"],"falsifier":"If a dLLM trained with a dedicated padding token (so EOS pressure at late positions is suppressed) still collapses to answer-only outputs under unconstrained pure decoding at the same rate as LLaDA-8B, the belief account is falsified and reachability is sufficient; if the collapse disappears, the paper's claim that termination beliefs are non-causal would be contradicted.","tokens_in":15216,"feed_emoji":"🧠","tokens_out":10928,"duration_ms":82802,"temperature":0.7,"pith_summary":"Masked diffusion language models advertise the freedom to commit tokens in any order, but on reasoning tasks this freedom is the axis of failure. By logging every commitment during decoding of LLaDA-8B on GSM8K, the paper shows unconstrained pure decoding freezes the final answer within the first quarter of the trajectory and collapses to answer-only outputs on up to 90% of problems as the canvas grows. The cause is not the model's termination beliefs—EOS pressure is nearly identical across decoders—but reachability: whether the sampler may act on those beliefs at distant positions. A 2×2 prompt–decoder experiment shows chain-of-thought helps only under ordered commitment, and a single-knob intervention, frontier-gated commitment, recovers the full gap without retraining while preserving parallel decoding. The paper reframes window-style samplers, previously seen as efficiency devices, as the minimal fix for a reasoning pathology.","feed_headline":"Commitment order, not EOS bias, is the real reasoning killer","feed_subtitle":"Restricting where tokens are committed restores chain-of-thought accuracy without retraining.","key_machinery":"The load-bearing mechanism is frontier-gated commitment, a decoding-time constraint that limits commitment eligibility to masked positions within a window of size w ahead of the left-most masked position. The sampler still scores every position and selects by confidence, but only among eligible candidates; w=∞ reproduces unconstrained pure decoding token-for-token and w=1 enforces left-to-right order. The paper uses this single knob as a manipulated variable to show that restricting reachability alone recovers the order channel—the part of the reasoning loss that persists even when reasoning text is produced—while also measuring the accuracy–parallelism frontier, where the optimal window flips from w=1 at full refinement to unconstrained at eight tokens per step. The companion measurement tool is the pressure–commitment dissociation: EOS pressure (the fraction of masked positions whose top-1 prediction is EOS) versus EOS commitment (the fraction of positions actually finalized as EOS), which separates what the model believes from what the sampler is permitted to do.","core_discovery":"The central claim is that in masked diffusion language models, the sampler's freedom to commit any token at any step—not the model's learned termination beliefs—determines whether chain-of-thought reasoning functions. Pure decoding of LLaDA-8B-Instruct on GSM8K commits the answer at a median normalized step of 0.147–0.240 while the median reasoning token is committed near 0.505, and it produces answer-only outputs on 36.6% of problems at L=256 and 90.2% at L=512. EOS pressure, the fraction of masked positions whose top-1 prediction is EOS, is nearly identical across pure and semi-autoregressive decoders (0.782 vs 0.766), while EOS commitment differs by more than 2×, showing the model wants to terminate equally but only the unconstrained sampler can act on that want at distant positions. The paper establishes causality by manipulating commitment order alone: frontier-gated commitment, which restricts eligibility to a window ahead of the left-most masked position, lifts pure decoding from 0.528 to 0.824–0.852 without training or logit modification. The general lesson is that any-order generation is not a neutral flexibility but a commitment policy that can make the conclusion precede and then bend the derivation.","pith_inferences":["We infer that the commitment-log instrumentation used here could serve as a general diagnostic for reasoning faithfulness in diffusion models: because the log records causal order directly, it makes post-hoc rationalization observable in a way that autoregressive models can only probe indirectly; applying it to code generation or open-ended dialogue, where the paper says the order channel is untes","We note the paper's belief-exoneration rests on the assumption that EOS pressure computed over masked positions is a fair cross-decoder measure; if masked-position sets themselves bias top-1 predictions, the dissociation would be weaker, though the gating experiment still demonstrates reachability is sufficient.","The failure of the paper's adaptive window schedules at low NFE is explained by commit density rather than window availability, which suggests that future fixes might attack the per-step commit budget itself (e.g., by scheduling fewer commitments per step) rather than spatial eligibility.","The measured crossover at 4–8 tokens per step may generalize to other any-order decoders such as non-autoregressive Transformers, where similar early-commitment pathologies could degrade multi-step reasoning; this is an extrapolation, not a paper claim."],"forward_implications":["Existing block and window samplers, previously motivated by efficiency, work because they restrict reachability; the paper shows a sliding window of 32 matches a block of 32 within 1.2 points, so the active ingredient is reachability, not block structure.","Retraining dLLMs to remove EOS bias addresses a symptom; a decoding-time gate recovers the full order channel on GSM8K and Dream-7B, so the documented EOS-overflow remedies are not necessary for this failure.","The optimal commitment window flips with the refinement budget: w=1 is best at full refinement and any-order freedom buys nothing, while at 8 tokens per step gating is purely harmful, implying diffusion decoding needs enough steps to exploit order.","Operationally, a gated decoder at 4 tokens per step retains near-autoregressive accuracy at a measured 4× wall-clock speedup, giving a concrete operating point for fast ordered decoding.","Answer-first commitment is universal under pure decoding, and 57–66% of erroneous equations are load-bearing (the erroneous value is precisely what the committed answer requires), so unfaithful reasoning is not an accident but a structural consequence of the commitment order."],"supporting_citations":[{"why":"Supplies the LLaDA-8B model and the observation that block-wise sampling beats unconstrained decoding on GSM8K, the effect this paper re-measures and re-explains.","marker":"Nie et al. (2025)"},{"why":"Names the <eos> overflow effect that the paper's pressure–commitment dissociation refutes.","marker":"Kim et al. (2025)"},{"why":"Documents the encroachment of EOS tokens into reasoning space, which the paper shows is gated by reachability.","marker":"Liu et al. (2026)"},{"why":"Provides the window-style sampler that the paper repurposes as a manipulated variable, anchoring the window at the left-most masked token.","marker":"Shu et al. (2026)"},{"why":"Supplies the Dream-7B model used to replicate the 2×2 interaction and the gating recovery.","marker":"Ye et al. (2025)"},{"why":"Provides the GSM8K benchmark on which the central experiments are run.","marker":"Cobbe et al. (2021)"},{"why":"Provides MATH-500, the second dataset where the order channel operates alone.","marker":"Hendrycks et al. (2021)"},{"why":"The official sampling guidelines that recommend block decoding, the configuration this paper's question begins from.","marker":"LLaDA Team (2025)"}],"fun_headline_variants":["Any-order token commitment breaks LLM reasoning","Answer-first decoding kills chain-of-thought, simple fix works","Commitment order, not EOS, decides reasoning success","Restricting token order rescues reasoning in diffusion LLMs"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The dissociation between EOS pressure and commitment assumes that the fraction of masked positions whose top-1 prediction is EOS is a decoder-invariant measure of termination belief, even though pure and semi-AR decoders have different masked-position sets and already-committed contexts at the same step.","fun_headline_variants_meta":{"raw":{"variants":["Any-order token commitment breaks LLM reasoning","Answer-first decoding kills chain-of-thought, simple fix works","Commitment order, not EOS, decides reasoning success","Restricting token order rescues reasoning in diffusion LLMs"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000381,"raw_usage":{"total_tokens":2098,"prompt_tokens":1101,"completion_tokens":997,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":717,"completion_tokens_details":{"reasoning_tokens":932}},"tokens_in":717,"tokens_out":997,"duration_ms":7738,"temperature":1.0,"reasoning_tokens":932,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-08T00:58:11.004754+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"If a dLLM trained with a dedicated padding token (so EOS pressure at late positions is suppressed) still collapses to answer-only outputs under unconstrained pure decoding at the same rate as LLaDA-8B, the belief account is falsified and reachability is sufficient; if the collapse disappears, the paper's claim that termination beliefs are non-causal would be contradicted.","supporting_citations":[{"cited_title":"well, keep thinking","cited_arxiv_id":null,"evidence_quote":"Names the <eos> overflow effect that the paper's pressure–commitment dissociation refutes."}],"review_version":1}