{"id":"7a3c8a13-317a-4ace-a45d-5fbc31519df8","arxiv_id":"2608.01740","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A disagreement signal between two cached forecasters lets diffusion sampling decide when to trust a forecast, improving quality at matched denoiser evaluations.","lead":"Diffusion image and video samplers can skip denoising steps by forecasting features, but the forecasts are not always trustworthy. RACER watches two cheap forecasts disagree and uses that disagreement to shrink or refresh predictions, reporting better quality at the same compute across four models.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Calibration trace may overlap evaluation set; if controller scalars are fit on benchmark prompts, the matched-NFE gains in Tables 1-3 are in-sample and the central improvement claim is unsubstantiated.","rationale":"I read the paper in good faith. The disagreement signal idea is plausible, and the AUROC 0.94 in Figure 1 is genuine evidence that the signal separates high-error steps. The theoretical propositions are simple but valid. However, the main tables' validity hinges on the calibration trace being independent of the evaluation set. The paper contains no statement of disjointness; in fact, it says video thresholds are calibrated per model and regime on a short trace, which is exactly the setting where leakage can occur. This is a concrete, checkable omission. The reader's weakest_assumption about observer validity is partially addressed by the empirical AUROC and by the reported worst-step correlations; it is a gap in the theory but not the primary threat to the empirical claim. The calibration overlap is a direct threat to the headline result. Therefore, the verdict should remain CONDITIONAL pending the calibration split disclosure.","tokens_in":12389,"tokens_out":5956,"duration_ms":68092,"concrete_test":"Ask the authors for the exact calibration protocol: number of prompts, source dataset, and whether any of those prompts appear in the DrawBench/VBench/COCO evaluation sets. If overlap exists, re-run the full main tables with the scalars fixed on a disjoint held-out set (or on a separate dataset) and compare the PSNR gain over the Chebyshev base. If the gain on SD3.5 at α=0.25 drops from +1.54 dB to below, say, +0.5 dB, the central improvement claim is not supported. Independently, reproduce Table 1 with the released code and a user-specified calibration split to verify the threshold values are insensitive to the split.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central empirical claim—RACER improves the strongest open-loop baseline at matched NFE on DrawBench, VBench, and COCO—depends on the controller scalars (λ, β, θκ, θref) being fixed without using the evaluation distribution. The paper states 'A short offline trace fixes the trust scalars' (Method) and 'video thresholds are calibrated per model and regime on a short trace' (Experimental Setup), but never specifies that this trace is disjoint from the evaluation prompts. If the trace is drawn from the same benchmarks, the refresh threshold and sigmoid scalars are effectively tuned to the test set, so the PSNR/SSIM gains in Tables 1–3 and the speed-quality curve in Figure 3 can reflect in-sample overfitting rather than a property of the disagreement signal. This missing support is more load-bearing than the observer-validity assumption: even a perfect AUROC for the signal does not protect against controller hyperparameters that are chosen on the evaluation prompts. The paper should specify the trace composition and overlap status.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"RACER is a training-free closed-loop controller for diffusion feature caching. The paper observes that two cached forecasts with different biases (a Chebyshev base and a Taylor observer) agree on smooth stretches and diverge at hard steps; this disagreement r_t is proposed as a reliability signal requiring no extra denoiser evaluations. RACER maps r_t to two responses: continuous shrinkage of the trusted feature toward the anchor via κ_t = exp(-λ max(k_t-1,0)) σ(β(θ_κ - r_t)), and an exact refresh-and-repay mechanism that recomputes a feature at high-disagreement steps and skips a later scheduled step to keep per-prompt NFE constant. The paper proves three propositions: a triangle-inequality error bound, an MSE-optimal shrinkage formula, and exact budget conservation. Experiments compare RACER against open-loop baselines on SD3.5-Large, FLUX.1-dev, Wan2.1-14B, and HunyuanVideo on DrawBench, VBench, and COCO, reporting matched-NFE gains, ablations, observer choice, and transfer to a Taylor base.","tokens_in":12686,"tokens_out":4589,"duration_ms":49542,"significance":"If the empirical results hold, the contribution is a simple, orthogonal control layer that can be stacked on any feature forecaster. The disagreement signal is cheap, the refresh-and-repay budget argument is clean, and the paper includes multiple ablations and a public code link. The theoretical results are elementary (Prop. 1 is a convexity/triangle inequality, Prop. 2 is standard forecast combination, Prop. 3 is structural), so the value rests on the empirical validation. The benchmark sweep across four models and three datasets is welcome, and the transfer experiments to a Taylor base help establish generality.","major_comments":[{"comment":"The matched-NFE gains in Tables 1–3 and Figure 3 are in-sample unless the offline trace used to fix λ, β, θ_κ, θ_ref is disjoint from the evaluation prompts. The paper says 'A short offline trace fixes the trust scalars' and 'video thresholds are calibrated per model and regime on a short trace' but never states the trace composition or overlap status. This is load-bearing because the central claim is a comparison at matched NFE over specific benchmarks; quantile calibration and threshold selection on the evaluation distribution would directly inflate the reported improvements. Please specify the trace (number of prompts, source, disjointness) for all models, and for image-side transfer from SD3.5 to FLUX.","section":"Experimental Setup, 'Implementation details'; Method, 'Closed-Loop Control'"},{"comment":"The claim that this is a 'deterministic error bound' is stronger than what is established. B_f and B_h are not defined in the main text; on real traces they are necessarily measured per trace. If B_f and B_h are empirical maxima from a calibration trace, the inequality holds for those traces by construction and gives no distribution-free guarantee for new prompts. The paper should either define B_f and B_h explicitly and provide a priori bounds (e.g., Lipschitz or Chebyshev remainder bounds) or reframe the result as an empirical envelope. This affects the abstract's 'deterministic error bound' claim.","section":"Theoretical Analysis, Proposition 1"},{"comment":"The theoretical validity of the disagreement signal requires an observer error bound B_o and a weak error correlation ρ_fo. The paper never reports B_o for the deployed Taylor observers, and reports error correlations only on 'worst steps' (0.17–0.44). Without B_o, Eq. (4) cannot certify that large disagreement implies large base error. The empirical AUROC of 0.94 is supportive but does not substitute for the stated theoretical conditions. Please report B_o estimates on the calibration trace or explicitly present (4)–(5) as motivating assumptions rather than verified bounds.","section":"Theoretical Analysis, Eqs. (4)–(5)"},{"comment":"The FORA, TaylorSeer, TeaCache, and ToCa entries are identical to the α=3.0 rows despite different reported NFE (9 and 8 vs. ~10). The note 'The four cache baselines land on one schedule for α=3.0 and 5.0' does not explain how the same PSNR/SSIM can accompany different NFE. If these baseline numbers were not re-measured at the aggressive schedules, the matched-NFE comparison at α=5.0/7.0 is not supported and the claim 'advantage holds in every regime' is compromised. Please correct the table or clarify the scheduling.","section":"Table 1, α=5.0 and α=7.0 rows"}],"minor_comments":[{"comment":"There are small spacing typos such as '2.6to5.4×' and '0.76×the latency'; these should be fixed. Also, Eq. (2) should explicitly state that the denominator is the norm of the base forecast, not the true feature.","section":"Throughout"},{"comment":"The Taylor observer g is not defined in the main text. Please give the exact order and step-size formula used for the Taylor expansion, or cite where it is specified.","section":"Method, Preliminaries"},{"comment":"The phrase 'image parameters are transferred from SD3.5 to FLUX' is vague: it should state which parameters are transferred, and whether the transfer includes the refresh threshold or only the trust scalars.","section":"Experimental Setup"}],"recommendation":"major_revision","confidential_remarks":"The calibration-trace overlap issue is the primary risk: if the trace is not disjoint from the evaluation benchmarks, the central empirical claim is unsubstantiated and the paper would need re-running, not just rewriting. The table inconsistency at α=5.0/7.0 is also concerning. I believe the paper's core idea has merit, so major revision with the required clarifications/experiments is the appropriate outcome."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the core idea is real and worth engaging. Two forecasters with different biases disagree where the feature trajectory bends, and that disagreement is a cheap output-side reliability signal that needs no extra denoiser call. Putting a closed-loop controller on top of an open-loop cache — continuous shrinkage plus an exact refresh-and-repay — is a genuine step beyond the binary accept/reject or input-side probes in the cited literature. The empirical sweep is broad, and the matched-NFE improvements over the Chebyshev baseline are consistent across four models and three benchmarks. The cross-model transfer (image params from SD3.5 to FLUX) and the signal ablation against horizon/random rules help support the claim that the disagreement carries the signal.\n\nThe theory is honest but modest. Proposition 1 is a triangle inequality, and the 'deterministic bound' in practice relies on empirical B_f and B_h; it would be fair to say it's an empirical envelope, not an a priori guarantee. Proposition 2 is a standard MSE combination result; Proposition 3 is true by construction. That's fine, but it shouldn't be oversold.\n\nThe real soft spot is the calibration trace. The paper says 'a short offline trace fixes the trust scalars' and video thresholds are 'calibrated per model and regime on a short trace,' but never states that the trace is disjoint from the evaluation prompts. If the trace comes from the same benchmarks, the gains in Tables 1–3 are partly in-sample. That is a reporting gap more than a demonstrated flaw: the FLUX transfer suggests the scalars aren't simply memorized for images, and the ablations show the disagreement signal itself is informative. But the authors need to state trace composition explicitly, report the actual scalar values, and ideally do a sensitivity analysis over them. Also missing: error bars on the main tables, a few LPIPS regressions that are acknowledged only in passing, and no measured bound B_o for the observer, which is one of the two conditions that legitimize the signal.\n\nWho this is for: anyone working on training-free diffusion acceleration or feature caching. It doesn't reshape the field, but it's a clean, useful mechanism with broad empirical results and a clear next step for follow-up work. A serious referee should see it; I'd accept it for review conditional on the calibration-trace disclosure and minor reporting fixes.","headline":"Genuinely new reliability signal for feature caching, but the calibration-trace overlap question has to be answered before the matched-NFE gains are fully trusted.","tokens_in":13168,"tokens_out":2262,"would_cite":true,"duration_ms":22527,"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":"This paper argues that the missing lever in training-free diffusion feature caching is not a better forecaster but a way to decide, at runtime, when and how much to trust each forecast.","keywords":["diffusion sampling acceleration","feature caching","training-free forecasting","closed-loop control","forecast reliability","disagreement signal","denoiser evaluation budget","RACER"],"falsifier":"Compute the true base forecast error and observer error step-by-step on many prompts and measure their correlation rho_fo and the observer error bound B_o. If rho_fo is high on hard steps (say above 0.8) or B_o is comparable to the base error, then Equations (4)-(5) show disagreement cannot reveal forecast error, so an ablation replacing r_t with the true error should show RACER's gains shrinking or reversing.","tokens_in":12292,"feed_emoji":"⚡","tokens_out":5095,"duration_ms":51119,"temperature":0.7,"pith_summary":"The paper claims that forecast reliability can be read from the cache itself: two cheap forecasts with different biases agree on smooth stretches of the feature trajectory and diverge where prediction turns hard, and their normalized disagreement flags high-error steps without any extra denoiser evaluation. On this signal, RACER closes the loop with two responses: it continuously shrinks uncertain forecasts toward the last computed feature, and at the riskiest steps it spends one real evaluation and repays it by skipping a later scheduled one, so the per-prompt evaluation count stays exactly fixed. At matched NFE, RACER improves the strongest open-loop baseline on SD3.5-Large, FLUX.1-dev, Wan2.1-14B, and HunyuanVideo across DrawBench, VBench, and COCO, and on SD3.5 it reaches equal quality at lower latency. The central claim is that reliable diffusion acceleration depends on how forecasts are used, not only on how they are made.","feed_headline":"Two forecasts' disagreement speeds up diffusion sampling","feed_subtitle":"A closed-loop controller shrinks or refreshes uncertain forecasts, improving four image/video models without extra denoiser calls.","key_machinery":"The load-bearing objects are the disagreement signal r_t and the trust function kappa_t. r_t is the relative norm difference between two forecasts produced from the same cache—the base forecaster and an observer with a different bias—so reliability becomes visible with O(F) arithmetic and no extra denoiser evaluation. kappa_t = exp(-lambda * max(k_t - 1, 0)) * sigmoid(beta * (theta_kappa - r_t)) sets the interpolation weight between the anchor feature and the forecast; kappa = 0 recovers pure feature reuse and kappa = 1 recovers full open-loop forecasting. A refresh-and-repay debt counter turns flagged risky steps into an exact budget-conservation mechanism, keeping the number of denoiser ev","core_discovery":"The central discovery is that the reliability of a diffusion feature forecast is observable from the cache alone. The paper defines a runtime signal r_t = ||h_hat_t - g_hat_t|| / ||h_hat_t|| between a base forecaster and a second forecaster with a different bias, called the observer. Because the two forecasts carry different biases, their disagreement grows exactly where the feature trajectory bends and the true forecast error spikes. Across four image and video models, this signal identifies the top-20% forecast-error steps at a mean AUROC of 0.94, versus 0.77 for the input-side signals that prior caches use. RACER converts this signal into a closed-loop controller: an exponential-sigmoid t","pith_inferences":["Editorial inference: if disagreement reliably flags hard steps, the same signal could also choose where to place the computed steps U itself, rather than only deciding whether to refresh within a fixed schedule.","Editorial inference: the validity conditions suggest a practical diagnostic—measuring the observer's error bound and its error correlation with the base on a short trace would predict when the disagreement signal will fail.","Editorial inference: because the signal only needs two forecasts with different biases, one could dynamically select among multiple base forecasters or observers rather than fixing a single pair.","Editorial inference: the query-by-committee motivation hints at an ensemble interpretation, but the paper stops at using disagreement as a risk score and does not develop ensemble weighting of multiple forecasts."],"forward_implications":["If RACER's central claim is correct, training-free feature caching can be improved without modifying the denoiser and without training: adding a second forecaster and a controller is enough.","The gains concentrate where open-loop caches break—aggressive acceleration regimes—so the same fixed NFE budget buys more quality exactly when skipping is most risky.","The controller is decoupled from the base forecaster: it improves a Taylor base as well as a Chebyshev base, so the benefit transfers across forecasting designs.","The per-prompt NFE remains exactly the base schedule's NFE, meaning the improvement is not additional compute but a reallocation of the same compute to the most unreliable steps.","The signal and controller transfer across models and datasets, with image-side calibration transferring from SD3.5 to FLUX with only a small trust-threshold change."],"supporting_citations":[{"why":"Supplies the Chebyshev global-basis forecaster that RACER uses as its base and improves.","marker":"Han et al. 2026"},{"why":"Defines the training-free feature-caching setup and the reuse paradigm RACER builds on.","marker":"Ma, Fang, and Wang 2024"},{"why":"Provides the query-by-committee principle that disagreement between predictors proxies uncertainty.","marker":"Seung, Opper, and Sompolinsky 1992"},{"why":"Gives the forecast-combination formula that Proposition 2 extends to anchor-forecast shrinkage.","marker":"Wang et al. 2023"},{"why":"TaylorSeers local expansion is the observer forecaster, a baseline, and the Taylor base in transfer experiments.","marker":"Liu et al. 2025b"},{"why":"Represents the runtime-check line of caches that RACER extends from binary recompute to continuous trust.","marker":"Zheng et al. 2025"},{"why":"Input-side signals that prior caches rely on, used as the comparison point for signal quality.","marker":"Liu et al. 2025a"},{"why":"FORA baseline and fast-forward caching in diffusion transformer acceleration.","marker":"Selvaraju et al. 2024"}],"fun_headline_variants":["Diffusion speedup via forecast disagreement","Disagreement between forecasts flags risky diffusion steps","Forecast disagreement is a free signal for diffusion cache control","When forecasts disagree, diffusion sampling gets faster","RACER uses forecast disagreement to close the loop"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The disagreement signal is a faithful reliability proxy only if the observer is a valid check: its own error must be bounded and only weakly correlated with the base forecaster's error, otherwise the two forecasts can agree while both are badly wrong.","fun_headline_variants_meta":{"raw":{"variants":["Diffusion speedup via forecast disagreement","Disagreement between forecasts flags risky diffusion steps","Forecast disagreement is a free signal for diffusion cache control","When forecasts disagree, diffusion sampling gets faster","RACER uses forecast disagreement to close the loop"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001481,"raw_usage":{"total_tokens":5843,"prompt_tokens":858,"completion_tokens":4985,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":602,"completion_tokens_details":{"reasoning_tokens":4914}},"tokens_in":602,"tokens_out":4985,"duration_ms":30157,"temperature":1.0,"reasoning_tokens":4914,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T21:49:44.443909+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute the true base forecast error and observer error step-by-step on many prompts and measure their correlation rho_fo and the observer error bound B_o. If rho_fo is high on hard steps (say above 0.8) or B_o is comparable to the base error, then Equations (4)-(5) show disagreement cannot reveal forecast error, so an ablation replacing r_t with the true error should show RACER's gains shrinking or reversing.","supporting_citations":[],"review_version":1}