{"id":"8f67e0c6-399e-4866-b309-bb580d66274f","arxiv_id":"2608.08082","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A prompt-specific commitment horizon, identified by comparing guided versus base-only continuations, marks an early point where classifier-free guidance can be removed with little loss in constraint success.","lead":"This paper asks when classifier-free guidance can be switched off during masked diffusion language model decoding without hurting constraint satisfaction. It defines a per-prompt commitment horizon using paired continuations and finds that for many prompts the base model can safely finish alone while most tokens are still masked.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The per-prompt horizon a* is estimated from one recorded trajectory but treated as a prompt-level property; fresh-trajectory intention-to-treat results show up to 0.07 absolute drops, so noninferiority does not yet transfer from the cross-fitted census to deployment.","rationale":"The reader's weakest assumption identifies the same load-bearing concern: a* is estimated from a single recorded guided trajectory but reported as a stable prompt-level horizon. The paper is unusually careful methodologically: it uses cross-fitting, prespecified noninferiority margins, audits of corrected artifacts, bit-exact replay checks, and it reports the fresh-deployment ITT table that exposes the risk. I therefore do not see a foundational mathematical error or fabrication. The issue is scope. Definition 1 defines a* as an ensemble quantity over guided prefixes, while the estimator conditions on one realized prefix; cross-fitting removes selection optimism within that prefix's rollouts but not variability across prefixes. The noninferiority receipt in Table 1 is legitimate for the cross-fitted, trajectory-conditioned estimand. However, the abstract's phrasing 'freezing each prompt at its own cross-fitted horizon is noninferior' can be read as a deployment claim, and Table 18 shows a pooled fresh-deployment gap of -0.04 with several subtasks at -0.06 or -0.07. That does not invalidate the paper's explicit conditional claim, but it requires the abstract and Section 5.2 to state that the result is conditioned on the recorded trajectory. The secondary mismatch between the estimator's 0.9/0.1 thresholds and the formal epsilon=0.03 would matter if a* were claimed to implement Definition 1 exactly, but the empirical noninferiority test is the actual evidence; still, reconciling the estimator with the formal definition would strengthen the paper. Keeping the verdict conditional, with the request to scope the claim and report trajectory-variability, is the right outcome.","tokens_in":53335,"tokens_out":7422,"duration_ms":77179,"concrete_test":"Take three representative subtasks (CommonGen kw k=10, CCG multi, CCG length) and n=50 prompts each. For each prompt, record R=20 independent guided trajectories; estimate a*_r on each trajectory using the same grid and N=24 cross-fitted rollouts. Compute the per-prompt range and IQR of a*_r across trajectories. Then deploy on fresh rollouts: freeze at the per-trajectory a*_r and at the per-prompt median a*_r, and compare terminal success with full CFG using paired lower 95% confidence bounds against -0.03 per subtask. If the per-prompt IQR spans at least two grid points for more than 30% of handoff prompts, or if any subtask's fresh-deployment lower confidence bound falls below -0.03, the prompt-level noninferiority claim fails; if the IQR is concentrated and fresh confidence bounds clear -0.03, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The abstract and Section 5.2 claim that freezing at each prompt's cross-fitted horizon is noninferior to full CFG at the prespecified margin epsilon=0.03 on all 13 subtasks. But Definition 1 defines a* as a prompt-level ensemble quantity via S_sw(s)=E^g[q0_s(X_s^g)], an average over guided prefixes. The estimator in Appendix C.1/C.3 instead records one guided trajectory per prompt, reconstructs states at seven grid fractions, and runs N=24 base and N=24 guided continuations from each; the paper explicitly calls the resulting a* a trajectory-level estimate of the ensemble-averaged horizon. Cross-fitting in C.2 corrects selection optimism within those shared rollouts, but it cannot correct trajectory-to-trajectory variability in the horizon. The paper's own fresh-deployment intention-to-treat table (Table 18) shows the symptom: at K=1, pooled handoff success is 0.69 versus 0.72 for full CFG, with per-subtask gaps of -0.06 (CommonGen k=10), -0.07 (CCG multi), -0.03 (CCG length), and -0.04 (CommonGen k=3). The survivors-only fresh deployment in Table 2 is closer (-0.01), so the gap concentrates in prompts and states where the recorded trajectory's horizon does not carry over. The central claim therefore holds as a within-trajectory, cross-fitted statement, but the paper's language ('prompt-specific', 'each prompt at its own horizon') invites a deployment reading that the evidence does not yet support. A secondary definitional gap: the estimator's safety thresholds q0>=0.9 and qg-q0<=0.1 are not tied to the formal epsilon=0.03, further weakening the identification of the measured a* with the formal a*_epsilon.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies when classifier-free guidance (CFG) is needed during masked diffusion language model decoding. It defines a prompt-level commitment horizon a* as the earliest persistent switch point at which switching all remaining decoding to the base model costs at most a tolerance epsilon in terminal constraint success. The theoretical section shows that the base committor is a martingale (Theorem 1), that the derivative of the guided one-step committor with respect to the guidance weight is a covariance with the successor committor (Eq. 11), and that the remaining guidance value is exactly the expected accumulation of local lifts (Proposition 1). Empirically, the paper estimates a* per prompt from a single recorded guided trajectory with N=24 continuations at each of seven grid fractions, uses cross-fitting to separate selection from evaluation, and reports that freezing at the estimated horizon is noninferior to full CFG on all 13 tested subtasks at a prespecified margin of 0.03. It also studies parallel decoding after handoff and repair by reopening committed positions, and shows that cheap rollout-free gate sources can approximately match the rollout-measured oracle.","tokens_in":53666,"tokens_out":6218,"duration_ms":62180,"significance":"If the central claim is read as a within-trajectory, cross-fitted statement, the paper is a careful and useful contribution. It separates the notion of constraint commitment from token-level realization, provides a clean martingale/covariance explanation of when guided steps can change terminal success, and backs its empirical claims with an unusually disciplined methodology: prespecified noninferiority margins, cross-fitting, hypergeometric half-splits, bootstrap inference at the prompt level, and explicit reporting of null controls and censoring. The theoretical identities in Section 4 and Appendices A are correctly derived; the paper is also commendably transparent about the limits of its own transport-law explanation and about artifacts found and corrected during the evaluation. The significance is tempered, however, by the gap between the formal prompt-level definition of a* and the trajectory-level estimator actually used, which matters for the deployment reading of the headline noninferiority claim.","major_comments":[{"comment":"The formal commitment horizon is a prompt-level ensemble quantity: S_sw(t)=E^g[q0_t(X^g_t)] averages over guided prefixes. The estimator in Appendix C.1 records one guided trajectory per prompt, reconstructs states at seven grid fractions, and runs N=24 base and N=24 guided continuations from each; Appendix C.3 explicitly calls the resulting a* a trajectory-level estimate of the ensemble-averaged horizon. Cross-fitting in C.2 corrects post-selection optimism within those shared rollouts, but it does not correct trajectory-to-trajectory variability in the horizon. The paper's own fresh-deployment intention-to-treat results (Table 18, and Table 10 per subtask) show the symptom: at K=1, pooled handoff success is 0.69 versus 0.72 for full CFG, with per-subtask gaps of -0.06 (CommonGen k=10), -0.07 (CCG multi), -0.04 (CommonGen k=3), and -0.03 (CCG length); the survivors-only fresh deployment in Table 2 is closer (-0.01). The abstract and Section 5.2 say that freezing at 'each prompt's own cross-fitted horizon' is noninferior to full CFG at epsilon=0.03 on all 13 subtasks, which invites a prompt-level or deployment reading that the evidence does not yet support. The paper should either restrict the noninferiority claim to the cross-fitted within-trajectory census or provide a deployment-level analysis that re-estimates a* on fresh rollouts and reports intention-to-treat confidence intervals against the margin.","section":"§3.3 (Definition 1) vs Appendix C.1–C.3"},{"comment":"The estimator uses safety thresholds q0>=0.9 and qg-q0<=0.1 at every later grid point, while the formal definition (Definition 1) and the reported noninferiority margin use epsilon=0.03. These are not equivalent: a switch point with q0=0.90 and qg=0.95 has an empirical gap of 0.05>0.03 and yet passes the estimator's criterion, while a point with q0=0.85 and qg=0.87 (gap 0.02<=0.03) fails the q0 threshold. Since the noninferiority test in Table 1 is evaluated separately, this does not by itself invalidate that table, but the paper should state explicitly that the estimated a* is not the epsilon=0.03 object of Definition 1 and justify the estimator thresholds, or replace them with the formal tolerance on S_sw. The current wording makes the connection between the theory and the headline margin appear tighter than it is.","section":"§5.2 and Appendix C.1"}],"minor_comments":[{"comment":"The sentence 'Section?? withdraws the two claims those results overturn' contains an unresolved cross-reference; the earlier claims are in Section 5.3 or Appendix F and should be cited explicitly.","section":"Appendix G"},{"comment":"There are two apparent typos: 'therefore report a* as a relational order parameter' is missing a subject, and 'rather th signal' should read 'rather than a signal'. These impede readability in an otherwise carefully written appendix.","section":"Appendix I"},{"comment":"The text says that prompts that already succeed without guidance are excluded from the handoff analysis, while Table 1 reports results on the full n=200 census including born-in prompts. Please clarify how born-in prompts are assigned in the full-census freeze evaluation (for example, whether they are assigned a*=0) and how their presence affects the reported confidence intervals.","section":"Section 5.2 and Table 1"},{"comment":"The caption correctly says the figure is illustrative, but the two failure-mode labels 'collapse' and 'hopeless' are used as technical terms before they are formally defined; a one-sentence definition in the caption or in Section 3.3 would help readers.","section":"Figure 1"}],"recommendation":"major_revision","confidential_remarks":"The paper is unusually transparent and already contains much of the evidence for the deployment caveat in its own appendices. The main revision should not be a pure wording fix: I would like to see either a fresh-rollout noninferiority analysis (with a* re-estimated per trajectory) or a clear demotion of the deployment reading, together with a short explicit reconciliation of the estimator thresholds with the formal epsilon. The theoretical core is sound and the empirical methodology is far above the usual standard for this area."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThis one is worth a serious look despite a real gap between the headline and the evidence. The core idea—define per-prompt guidance necessity by paired continuations from the same partial state, and identify the earliest safe switch point—is genuinely new relative to the scheduled-CFG literature. The empirical program is unusually careful: cross-fitted selection, prespecified noninferiority margin, post-hoc audits, and an explicit admission that the local covariance theory does not explain where the horizon lands. That honesty is substantial credit.\n\nWhat holds up: the martingale root and the covariance identity are standard but correctly derived; the cross-fitted freeze-at-a* census does meet the ε=0.03 margin on all 13 subtasks, as a within-trajectory statement. The finding that most variation in a* is within families (η²≈0.1), not between them, is interesting and credible. The reopening results are descriptive but well-controlled.\n\nWhere it gets soft: the abstract and Section 5.2 read like deployment claims, but a* is estimated from one recorded guided trajectory per prompt and then treated as a prompt-level property. Cross-fitting corrects selection optimism within shared rollouts; it cannot correct trajectory-to-trajectory variability in the horizon. The paper's own fresh-deployment ITT table (Table 18) shows a pooled K=1 gap of −0.04 and per-subtask drops up to −0.06/−0.07; survivors-only fresh deployment is closer (−0.01), so the shortfall concentrates where the recorded horizon doesn't transfer. The stress-test note on this is right, and it is the paper's main flaw: the noninferiority claim is a cross-fitted, within-trajectory result, not yet a deployment result. The abstract should say so.\n\nTwo smaller issues. First, the estimator's safety thresholds (q0≥0.9, qg−q0≤0.1) are not derived from the formal ε=0.03 definition; the measured a* is not literally the a*_ε in Definition 1. That is a definitional gap, not fatal, but it needs reconciling. Second, Appendix G still has unresolved 'Section ??' references, which looks unfinished.\n\nBottom line: the paper deserves refereeing. The math is sound, the evaluation is above the field's usual bar, and the central conditional claim holds. But the scope of the claim must be narrowed in revision, and the estimator-definition link needs tightening. The intended reader is someone working on diffusion LMs, adaptive guidance, or constrained decoding; they should read it for the methodology. I'd send it out.","headline":"Careful study of when CFG stops mattering in masked diffusion LMs; the headline noninferiority holds as a cross-fitted within-trajectory result but the abstract overreaches on deployment.","tokens_in":54286,"tokens_out":3806,"would_cite":false,"duration_ms":34392,"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":"The paper claims that classifier-free guidance in masked diffusion language models can be switched off at an early per-prompt commitment horizon with no more than a 0.03 loss in constraint success on all 13 subtasks, even while most…","keywords":["classifier-free guidance","masked diffusion language models","commitment horizon","committor","martingale","constrained text generation","guidance handoff","parallel decoding"],"falsifier":"Re-estimate each prompt's handoff time from several independently recorded guided trajectories rather than one, and evaluate the handoff on fully fresh rollouts that share no prefix with the estimating runs. If the paired handoff-minus-full-CFG difference has a lower confidence bound below −0.03 on more than one or two of the 13 subtasks, the prompt-level stability assumption — and the noninferiority that rests on it — is falsified. A cheaper companion check: the martingale theorem predicts zero drift in base-continuation success, so a large-sample estimate of $\\mathbb{E}[q^0_{t+1}-q^0_t \\mid X_t]$ that is systematically nonzero early in decoding would point to a broken estimator or Markov assumption.","tokens_in":53055,"feed_emoji":"🎯","tokens_out":15835,"duration_ms":130884,"temperature":0.7,"pith_summary":"Classifier-free guidance (CFG) is normally kept on for every decoding step of a masked diffusion language model, even though its value varies across prompts and over time. The paper's claim is that this necessity is front-loaded: for most prompts that benefit from guidance at all, there is an early per-prompt commitment horizon $a^*$ after which finishing with the unguided base model costs at most a small tolerance in constraint success, and freezing each prompt at its own cross-fitted $a^*$ is noninferior to full CFG on all 13 subtasks at the margin $\\varepsilon = 0.03$, even while most tokens are still masked. The reason to care is structural: the claim separates commitment from realization, meaning a constraint can be effectively locked in before the token identities that exhibit it are decided, which reframes CFG as an early trajectory corrector and opens later decoding stages to cheaper policies.","feed_headline":"Switch off guidance early: 13 of 13 tasks hold the 0.03 margin","feed_subtitle":"Classifier-free guidance can stop at each prompt's own horizon while most tokens are still masked.","key_machinery":"The load-bearing object is the commitment horizon $a^*_\\varepsilon$, defined as the earliest switch point $t$ such that switching all remaining decoding from the guided kernel to the base kernel keeps terminal success within $\\varepsilon$ of full CFG at every later switch point. Its definition rests on two committors — the base committor $q^0_t(x)$, the probability of eventual constraint satisfaction if all remaining steps use the unguided base model, and the guided committor $q^g_t(x)$ for continued CFG — whose difference is the remaining value of guidance. The theoretical engine is Theorem 1 (the base committor is a martingale under the base kernel) and the linear-response identity $$\\frac{\\partial}{\\partial w}\\mathbb{E}^{$K^{{g}}$_{t,w}}[$q^{0}$_{t+1}(X_{t+1})\\mid X_t=x]\\Big|_{w=0} = \\operatorname{Cov}_{$K^{0}$_t(\\cdot\\mid x)}(\\delta_t, $q^{0}$_{t+1}),$$ which says CFG's first-order effect at a step is exactly the covariance between its logit tilt $\\delta_t$ and the next state's base-continuation success. Proposition 1 closes the loop by writing the switch-off cost as the expected accumulation of these local lifts. The horizon is estimated by freeze interventions — one recorded guided trajectory per prompt with $N=24$ continuation rollouts at each of seven grid fractions — and selection is separated from evaluation by cross-fitting.","core_discovery":"The paper's central claim is that the need for classifier-free guidance in masked diffusion language models is prompt-specific and front-loaded, and that it can be located precisely by a per-prompt commitment horizon $a^*$ — the earliest decoding step from which switching every remaining step to the base, guidance-free model reduces the probability of final constraint satisfaction by no more than a chosen tolerance $\\varepsilon$. Defined through paired continuations from the same partial output, $a^*$ separates commitment from realization: at the median horizon the output is 55–78% masked, yet constraint success is already effectively decided. The paper supplies both a mechanism and a measurement: under the base decoder the success probability, or committor, is a martingale, and to first order each CFG step helps only through the covariance between its logit tilt and the successor committor; empirically, freezing each prompt at its own cross-fitted $a^*$ is noninferior to full CFG on all 13 subtasks at the prespecified margin $\\varepsilon = 0.03$. The same boundary marks the start of a later regime in which wider parallel decoding costs little in constraint success — though fluency still degrades with width — and for failed trajectories, reopening already-written positions improves recovery whether the failure was a collapse out of a satisfying configuration or never reaching one.","pith_inferences":["The paper's own attribution gap (Appendix D) is the sharpest open thread: guidance applied only at detected local-transport events recovers only about 18% of the full-versus-base success gap on keywords, which suggests the terminal benefit of CFG flows through higher-order, multi-step, or support-shifting channels that the first-order covariance law does not see; isolating those channels is a dire","Because the constraint family explains only about 10% of the variation in $a^*$ ($\\eta^2 = 0.097$), most of the horizon's variability lives inside prompts; identifying which prompt features push the horizon earlier or later would be a testable extension that could make deployed gates prompt-adaptive rather than family-calibrated.","The commitment-as-flatness view — outcome-level success stable while the token-realization distribution keeps changing — transfers naturally to continuous diffusion models and to other conditioning signals, where the same paired-continuation diagnostic could locate when guidance can be dropped; the paper does not make that transfer.","The repair-by-reopening result is descriptive and explicitly lacks a counterpart in the transport theory; its implication that reopening any committed position helps about as much as reopening the 'right' one suggests revision capacity itself, not position selection, is the scarce resource in late-stage recovery, which is a testable hypothesis for constrained decoding."],"forward_implications":["Guidance can be handed off to the base model: on all 13 subtasks, switching at each prompt's own horizon keeps constraint success within $\\varepsilon = 0.03$ of full CFG, so running CFG to the final step is not buying measurable constraint success for most prompts.","Commitment precedes realization: at the median horizon 55–78% of positions are still masked, so later decoding decisions can be organized around the commitment state rather than the revealed text.","A low-cost parallel regime opens after the horizon: widening the parallel block from $K=1$ to $K=16$ costs about 0.07 in success after handoff versus 0.25 under full CFG, though GPT-2 perplexity still degrades with width.","Failed trajectories are repairable by reopening: remasking already-written positions and resuming guided decoding improves recovery for both collapse and hopeless failures, and the gain does not depend on which position-selection signal is used.","The horizon can be localized without rollouts: a random forest on cheap per-prompt features matches the rollout-measured oracle in success (+0.015 [−0.001, +0.031] at $K=8$) at comparable forward-evaluation count, so long as it stays family-calibrated."],"supporting_citations":[{"why":"Defines classifier-free guidance, the mechanism whose removal the paper schedules.","marker":"Ho and Salimans 2022"},{"why":"Supplies LLaDA-Instruct, the masked diffusion language model on which the full 13-subtask census runs.","marker":"Nie et al. 2025"},{"why":"Provides the transition-path-theory committor concept that measures terminal success from a partial state.","marker":"E and Vanden-Eijnden 2010"},{"why":"Formalizes masked diffusion language model decoding, the reverse process the paper treats as partially decoded states.","marker":"Sahoo et al. 2024"},{"why":"Exemplifies the apply-CFG-at-every-step practice in discrete diffusion that the handoff policy is compared against.","marker":"Schiff et al. 2025"},{"why":"Supplies the CommonGen keyword-constraint subtasks of the census.","marker":"Lin et al. 2020"},{"why":"Supplies the IFEval instruction-following subtasks of the census.","marker":"Zhou et al. 2023"},{"why":"Supplies the COLLIE constrained-generation subtasks, including the failure-anchored c06a family.","marker":"Yao et al. 2024"},{"why":"Motivates the coarse-structure-before-detail picture behind commitment before realization.","marker":"Wang and Vastola 2023"}],"fun_headline_variants":["Early guidance cutoff keeps all 13 tasks at 0.03 margin","Per-prompt horizon: switch off guidance early, keep all 13 tasks","Guidance-free finish: 13/13 tasks within 0.03 margin","Commitment horizon: when CFG becomes unnecessary","Switch off CFG early, hold all 13 margins"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that each prompt's handoff time, estimated from a single recorded guided trajectory plus a small number of continuation rollouts at each grid point, is a stable property of the prompt; if re-running the same prompt yields a noticeably different handoff time, the headline noninferiority may not survive fresh deployment.","fun_headline_variants_meta":{"raw":{"variants":["Early guidance cutoff keeps all 13 tasks at 0.03 margin","Per-prompt horizon: switch off guidance early, keep all 13 tasks","Guidance-free finish: 13/13 tasks within 0.03 margin","Commitment horizon: when CFG becomes unnecessary","Switch off CFG early, hold all 13 margins"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000765,"raw_usage":{"total_tokens":3481,"prompt_tokens":1120,"completion_tokens":2361,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":736,"completion_tokens_details":{"reasoning_tokens":2270}},"tokens_in":736,"tokens_out":2361,"duration_ms":15461,"temperature":1.0,"reasoning_tokens":2270,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T00:26:55.205005+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-estimate each prompt's handoff time from several independently recorded guided trajectories rather than one, and evaluate the handoff on fully fresh rollouts that share no prefix with the estimating runs. If the paired handoff-minus-full-CFG difference has a lower confidence bound below −0.03 on more than one or two of the 13 subtasks, the prompt-level stability assumption — and the noninferiority that rests on it — is falsified. A cheaper companion check: the martingale theorem predicts zero drift in base-continuation success, so a large-sample estimate of $\\mathbb{E}[q^0_{t+1}-q^0_t \\mid X_t]$ that is systematically nonzero early in decoding would point to a broken estimator or Markov assumption.","supporting_citations":[],"review_version":1}