{"id":"3ab75379-7092-4489-9ea3-039847012ae2","arxiv_id":"2607.10853","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"Variance-corrective time shifting cancels the variance inflation of naive temperature sampling so pretrained diffusion models can raise diversity without retraining.","lead":"A training-free fix lets pretrained diffusion models sample rare modes by raising temperature without the usual blur and noise. The method is a practical diversity knob for image and motion generators already in use.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.5","headline":"Well-separated-modes derivation (Sec. 3.1, App. A) is invoked for early high-noise stages where modes fully overlap, so the claimed recovery of the tempered marginal is unjustified precisely where compositional diversity is obtained.","rationale":"The reader correctly isolates the well-separated-modes assumption as the weakest link supporting the strongest claim. The additional observation that the paper deliberately applies the operator in the high-noise regime where that assumption is false sharpens the same concern but does not introduce a new internal contradiction or a more severe flaw; the empirical tables and ablations still stand, and the method remains a cheap, training-free diversity knob whose practical utility is already demonstrated. Hence the CONDITIONAL verdict (pending broader stress-tests of the approximation and public code) is unchanged. No stronger load-bearing defect—e.g., algebraic error in the VP/flow derivations of Apps. B–C, or systematic failure of the reported metrics—was found.","tokens_in":16474,"tokens_out":710,"duration_ms":32289,"concrete_test":"Re-run the Sec. 3.2 2-D Gaussian-mixture experiment (three well-separated modes, π=(0.8,0.1,0.1)) but restrict the variance-corrective time-shift + γ-scaling (γ=0.1) exclusively to the earliest third of the DDPM trajectory (high σ_t). Measure (i) empirical mode frequencies against the target π^{(γ)} and (ii) per-mode sample covariance against the original σ_data^{2}. If either the KL(p||π^{(γ)}) exceeds the low-noise values of Table 1 by >0.05 or the empirical variances inflate by >20 %, the early-stage justification fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central construction (Eqs. 8–13) equates the tempered noised marginal p_t^{(γ)} to a variance-rescaled mixture only under the explicit premise that σ_t is small enough for modes to remain separated, making cross terms in p_t(x)^γ negligible (App. A). The paper then applies the identical time-shift + γ-scaling primarily in the early (high-noise) third of the trajectory (Sec. 3.3, Sec. 4.2, Fig. 2/4/5) to obtain compositional diversity. At those noise levels the modes of any realistic p_0 are completely mixed, the binomial expansion of App. A fails, and there is no longer a guarantee that querying at σ_˜t = √γ σ_t recovers \nabla log p_t^{(γ)}. The empirical quality/diversity gains therefore rest on an unanalyzed regime outside the stated approximation; if the early-stage score is simply a distorted version of the original score, the method reduces to an ad-hoc noise-level shift whose mode-reweighting effect is accidental rather than theoretically controlled.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The paper proposes variance-corrective time shifting as a training-free way to sample from the high-temperature target p0^γ (0<γ<1) with pretrained diffusion and flow-matching models. Naive score scaling reweights modes but inflates per-mode variance and degrades quality; the fix queries the network at a shifted noise level (σ̃_t² = γ σ_t², with VP/flow analogues) and multiplies the score by γ. Under a well-separated Gaussian-mixture approximation this cancels the variance inflation while preserving mode reweighting. The authors further show that applying γ(t)<1 at early vs. late denoising stages yields compositional vs. local diversity. Experiments on DiT, Stable Diffusion (1.5/XL/3.5) and a motion diffusion model report gains in Vendi diversity at modest cost to VQA/accuracy and CMMD, outperforming CADS and Feynman–Kac correctors in the reported settings.","tokens_in":16772,"tokens_out":1398,"duration_ms":30434,"significance":"If the construction is reliable, it supplies a simple, zero-retraining diversity knob that works across both classical diffusion and modern flow-matching backbones, including practical stage-wise control of composition vs. appearance. The identification of the variance-expansion failure mode of naive temperature sampling is useful, and the appendices give concrete, implementable score/velocity wrappers for VP and rectified-flow models. The toy-mixture reweighting check (Table 1) and the direct-tempering/TSR ablation (Fig. 6) are clear strengths. The main limitation is that the rigorous recovery argument is stated only under well-separated modes, while the primary diversity gains are obtained in the early, high-noise regime where that assumption fails; closing or clearly scoping that gap would make the contribution substantially more solid for the community.","major_comments":[{"comment":"Sec. 3.1 (Eqs. 8–13) and Appendix A justify the time-shift formula only when σ_t is small enough that modes remain separated and cross terms in p_t(x)^γ are negligible. The paper’s main diversity results, however, apply the same rule in the early (high-noise) third of the trajectory (Sec. 3.3, Sec. 4.2, Figs. 2/4/5), where modes of any realistic p0 are fully mixed and the binomial expansion of App. A does not apply. There is therefore no guarantee that querying at σ̃_t = √γ σ_t recovers ∇ log p_t^(γ) in the regime used for compositional diversity. Either (i) provide analysis or diagnostics for the overlapping-mode / high-noise regime, or (ii) reframe early-stage tempering as an empirically effective heuristic whose mode-reweighting is not theoretically controlled by the current derivation, and restrict the ‘recovers the tempered marginal’ claim to the low-noise setting where the approxim","section":"Sec. 3.1, Eqs. 8–13; App. A; Sec. 4.2"},{"comment":"The quantitative claim of ‘consistent gains at minimal cost to sample quality and condition fidelity’ rests on a small set of hand-chosen γ values per model (γ=0.20/0.88/0.94/0.78) selected after inspecting the Vendi–fidelity curves in Fig. 4. Table 2 reports a single operating point per model without a full Pareto comparison against CADS/FK across matched fidelity levels, and CMMD is acknowledged to penalize the intended deviation from p0. A clearer multi-γ trade-off table (or fixed-fidelity Vendi comparison) would make the superiority claim load-bearing rather than pointwise.","section":"Sec. 4.2, Table 2, Fig. 4"},{"comment":"Motion-generation evaluation is only qualitative (Fig. 8 and supplementary video). Given that the abstract and contributions list Motion Diffusion as a third modality demonstrating ‘consistent gains,’ the absence of any diversity/fidelity metric (even a simple feature-space Vendi or classifier-based condition score) leaves that part of the claim unsupported in the main text.","section":"Sec. 4.2 (Motion generation); Fig. 8"}],"minor_comments":[{"comment":"Sec. 4.2: typo ‘In this seciton’ → ‘section’; also ‘controled’ → ‘controlled’ in the same paragraph.","section":"Sec. 4.2"},{"comment":"Fig. 1 (left) is a useful 1D illustration of variance inflation, but the caption does not state whether the curves are analytic or sampled; a one-line clarification would help.","section":"Fig. 1"},{"comment":"Related work: TSR [37] is correctly identified as suffering from variance expansion; a short explicit statement that TSR’s γ(t) schedule is a smooth direct-tempering schedule (as later used in Fig. 6) would make the distinction sharper earlier.","section":"Sec. 2"},{"comment":"Eq. (13) and the VP/flow wrappers (App. B–C) are clear; adding a one-line pseudocode box for the ϵ_temp / v_temp wrapper in the main text would improve reproducibility for practitioners who skip the appendices.","section":"Sec. 3.1; App. B–C"},{"comment":"Fig. 7 is dense; the main-text claim that early-stage tempering usually gives the largest diversity gain would be easier to verify if the per-tier peak-γ values were tabulated.","section":"Fig. 7; Sec. 4.3"}],"recommendation":"major_revision","confidential_remarks":"The skeptic’s concern about applying a low-noise derivation at high noise is real and should be fixed in revision; it is not a reason to reject if the authors either analyze the overlapping regime or clearly demote the early-stage claim to empirical. Novelty relative to TSR and CADS is adequate once the variance-correction is cleanly scoped. Fit for a solid CV/ML venue is good after the theory–practice gap is addressed."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The one thing worth knowing is that this paper makes high-temperature sampling actually usable for diversity on off-the-shelf diffusion and flow models. Naive score scaling by γ < 1 reweights modes but inflates per-mode variance and wrecks the reverse process; their fix is to query the network at a quieter timestep (σ̃^{2} = γ σ^{2} or the VP/flow analogues) and then scale by γ. That is the concrete contribution, and it is not in TSR, CADS, FK correctors or particle guidance.\n\nThey derive it carefully for VE under the well-separated mixture, give the cross-term argument in the appendix, and extend the same construction to VP and rectified flow. The toy mixture tracks the tempered mode weights. On DiT and the Stable Diffusion family the method improves Vendi while holding VQA/accuracy and CMMD better than the baselines they compare, with error bars, and the stage ablation (early for composition, late for local detail) is a useful practical observation. Comparisons look fair.\n\nThe soft spot is real but proportionate: the math assumes modes stay separated at the current noise level, yet the biggest diversity gains come from applying the identical shift in the early high-noise third where modes are fully mixed. So the claimed recovery of the tempered marginal is not justified exactly where they use it most. Empirically it still works, which suggests the procedure is more robust than the analysis, but they do not close the gap. γ and the stage schedule are also hand-chosen per model after looking at the curves; motion evidence is only qualitative.\n\nThis is for people who use or ship diffusion samplers and want a cheap diversity dial without retraining. The math is solid where it applies, the experiments are thorough enough for a methods paper, and there is no internal contradiction. I would send it to peer review; a referee can push on the approximation regime and ask for code. Worth engaging if you care about inference-time control.","headline":"Practical zero-cost diversity knob for pretrained diffusion via time-shifted temperature; clean under mode separation but main gains come from the high-noise regime the math does not cover.","tokens_in":17400,"tokens_out":503,"would_cite":true,"duration_ms":19514,"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":"Querying a diffusion network at a shifted timestep cancels the variance inflation of high-temperature sampling, turning temperature into a training-free diversity knob.","keywords":["diffusion models","temperature sampling","score scaling","time shifting","diversity","flow matching","inference-time control","mode reweighting"],"falsifier":"On a multimodal density whose modes deliberately overlap at the noise scales used for tempering, measure whether the empirical mode frequencies still match the tempered weights π_k^γ and whether sample quality remains comparable to the untempered baseline; a large mismatch would refute the claim that the time shift recovers the tempered marginal.","tokens_in":17385,"feed_emoji":"🎛️","tokens_out":909,"duration_ms":14441,"temperature":0.7,"pith_summary":"Diffusion models copy their training data so faithfully that rare modes are hard to sample. Raising the temperature of the target distribution (raising p0 to a power γ less than one) would flatten common modes and lift rare ones, but simply scaling the score by γ also inflates each mode’s variance and wrecks sample quality. This paper shows that querying the pretrained score network at a carefully shifted, less-noisy timestep and then scaling by γ cancels that variance inflation while keeping the desired mode reweighting. The fix needs no retraining, works for both diffusion and flow-matching backbones, and gives users a practical dial for diversity. Applying the temperature change only in early (high-noise) steps mainly changes global composition; applying it late mainly changes local appearance.","feed_headline":"Time-shifted score query fixes temperature sampling","feed_subtitle":"A zero-retrain trick turns high temperature into a practical diversity dial for diffusion and flow models.","key_machinery":"Variance-corrective time shifting: query the score (or velocity) network at a shifted timestep t̃ chosen so σ̃_t² = γ σ_t² (or the VP / flow-matching analogue), then scale the output by γ. This cancels the variance expansion of direct score scaling while keeping the mode reweighting of the tempered target p0^γ.","core_discovery":"Naive temperature sampling reweights modes correctly but expands per-mode variance, so the reverse process no longer matches the tempered marginal and samples degrade. Under a well-separated-modes approximation, that expansion is exactly a rescaling of noise level. Querying the pretrained network at the shifted time whose noise level is γ times smaller, then multiplying the resulting score by γ, restores the original variance schedule while preserving the reweighting. The same construction carries over to variance-preserving diffusion and rectified-flow models, turning temperature into a zero-cost diversity control.","pith_inferences":["The same time-shift identity may extend to products of guidance signals (e.g., classifier-free guidance times temperature), giving a joint knob for fidelity and diversity.","How much diversity headroom a prompt has may itself become a diagnostic of how peaked the learned conditional is, revealing structure in the training distribution without extra training.","If the well-separated-modes premise fails for dense manifolds (e.g., highly entangled video or 3D), a learned or adaptive shift schedule might still recover a usable approximation."],"forward_implications":["Any pretrained diffusion or flow-matching checkpoint can be used as a diversity-controllable generator by wrapping its score or velocity call with a one-line time-shift and γ-scale.","Users can trade composition-level diversity against local detail diversity by choosing which segment of the denoising trajectory receives γ < 1.","Condition fidelity need not collapse when diversity is increased, provided γ is not driven too low and the correction is applied.","Competing inference-time diversity methods that lack the variance correction will continue to produce blurrier or off-manifold samples at high temperature."],"fun_headline_variants":["Time-shifted score query restores variance for tempered diffusion","Variance-corrective shift turns temperature into diversity control","Shifted timestep plus gamma score fixes high-temperature samples","Query network at gamma-rescaled noise to enable mode reweighting","Variance-corrective time shifting makes temperature sampling practical"],"cache_read_input_tokens":128,"weakest_assumption_plain":"The method rests on modes being well separated at the noise level where temperature is applied; if modes overlap strongly, the simple time-shift formula no longer recovers the tempered distribution.","fun_headline_variants_meta":{"raw":{"variants":["Time-shifted score query restores variance for tempered diffusion","Variance-corrective shift turns temperature into diversity control","Shifted timestep plus gamma score fixes high-temperature samples","Query network at gamma-rescaled noise to enable mode reweighting","Variance-corrective time shifting makes temperature sampling practical"]},"model":"grok-4.5","effort":"low","cost_usd":0.00425,"raw_usage":{"total_tokens":1307,"prompt_tokens":802,"num_sources_used":0,"completion_tokens":64,"cost_in_usd_ticks":42500000,"prompt_tokens_details":{"text_tokens":802,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":441,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":802,"tokens_out":64,"duration_ms":4879,"temperature":1.0,"reasoning_tokens":441,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-14T08:45:40.269108+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"On a multimodal density whose modes deliberately overlap at the noise scales used for tempering, measure whether the empirical mode frequencies still match the tempered weights π_k^γ and whether sample quality remains comparable to the untempered baseline; a large mismatch would refute the claim that the time shift recovers the tempered marginal.","supporting_citations":[],"review_version":1}