{"id":"fb84c62b-9d99-4fad-999d-640d5778409b","arxiv_id":"2608.07572","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":2,"one_line_summary":"BRACE replaces derivative-based polynomial forecasting with barycentric rational extrapolation for cached diffusion transformer features, reporting improved FID, ImageReward, and VBench at equal FLOPs.","lead":"This paper proposes a new way to speed up diffusion image and video generators by predicting the model's internal features with rational functions instead of polynomial formulas. It matters because faster generation without retraining is a practical goal for deploying generative AI at scale.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"In the flagship DiT-XL/2 setting (C=3, gamma=0.5), Eq. (7) reduces to quadratic Lagrange interpolation, so the claimed rational self-normalization is not actually exercised by the paper's headline results.","rationale":"The reader's weakest assumption includes the same core observation: the headline DiT-XL/2 configuration uses C=3 and gamma=0.5, which makes the barycentric weights identical to Lagrange weights and therefore turns the predictor into a quadratic polynomial. That observation is directly verifiable from Eqs. (7)-(8) and is fatal to the paper's central narrative. If the method is a polynomial in its most important benchmark, then the paper cannot claim that rational self-normalization or rational structural robustness explains its results. The remaining positive evidence for the rational mechanism is confined to FLUX/Hunyuan experiments with C=2, which are not the primary demonstration and are not supported by the missing stability proof. The stability theorem itself is also deferred to a missing supplement and its derivation mixes canonical mapped coordinates with original timestep coordinates, but I focus on the more decisive internal inconsistency: the flagship experiment does not instantiate the proposed mechanism. This concern reinforces the reader's REJECT verdict. The proposed concrete test would settle whether the issue is merely a poorly chosen hyperparameter or a genuine absence of any rational advantage: if the degenerate gamma=0.5 setting is best, the paper's own numbers argue against its mechanism; if a rational setting is better, the paper still needs to rerun its headline experiment and update all claims accordingly.","tokens_in":16727,"tokens_out":13446,"duration_ms":134212,"concrete_test":"Symbolically simplify Eq. (7) with C=3, gamma=0.5 and equidistant nodes to confirm it equals the quadratic Lagrange interpolant. Then re-run the Table 1 DiT-XL/2 protocol with genuinely rational settings (C=3, gamma=0.7 and gamma=1.0; or C=4 with Chebyshev-Gauss-Lobatto node positions) at N=4,5,6 and compare FID/sFID/IS. If the gamma=0.5 Lagrange setting remains best, the reported advantage is not attributable to rational structure; if a rational setting matches or beats it, the headline table should be replaced with that setting and the current rational-advantage claim remains unsubstantiated.","verdict_should_be":"REJECT","load_bearing_attack":"The central claim is that BRACE's rational structure is what tames sharp DiT feature irregularities. With C=3 and gamma=0.5, Eq. (8) gives weights w=[1/2, -1, 1/2]. For three equally spaced nodes these are exactly the barycentric weights of Lagrange interpolation, so Eq. (7) simplifies algebraically to the unique quadratic polynomial through the three cached features; the denominator is a removable artifact, not a self-normalizing rational denominator. Every BRACE row in the primary ImageNet/DiT-XL/2 benchmark (Table 1) uses this degenerate configuration. Thus any FID/sFID gains over TaylorSeer/HiCache in Table 1 are not evidence for the paper's rational-forecasting mechanism; the method in that benchmark is a feature-driven quadratic extrapolator. The genuinely rational case only appears in the FLUX/Hunyuan C=2 runs, which use different gamma values and are secondary to the headline ImageNet table. Since the paper's motivation and stability argument (Eq. 11) depend on a denominator D(x) staying bounded away from zero for a genuinely rational predictor, the headline experiments do not test the load-bearing mechanism. This is an internal inconsistency in the paper's own equations, not a disagreement with external consensus.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents BRACE, a training-free feature-caching method for accelerating Diffusion Transformer inference. Instead of derivative-based polynomial extrapolation (TaylorSeer, HiCache), BRACE proposes to forecast future hidden states by direct barycentric rational aggregation of a small FIFO cache of historical features, using an adapted Chebyshev weighting scheme. The method is evaluated on DiT-XL/2 for class-conditional ImageNet generation, FLUX.1-dev for text-to-image, and HunyuanVideo for text-to-video, reporting quality-efficiency trade-offs that are claimed to be state of the art. The paper also provides an empirical analysis of feature trajectories and a theoretical stability bound on the extrapolation error.","tokens_in":16970,"tokens_out":9022,"duration_ms":73654,"significance":"If the claims are substantiated, BRACE would be a useful plug-and-play acceleration module with a fresh numerical-analysis flavor, and the cross-architecture evaluation (image and video) is broader than in many caching papers. The idea of replacing finite-difference derivatives with feature-driven rational aggregation is attractive, and the reported latency gains are consistent with the caching paradigm. However, the paper's central mechanism is not actually exercised by its headline ImageNet experiments (the chosen configuration degenerates to a quadratic polynomial), the key stability proof is deferred to a missing supplement, and the empirical reporting lacks code, error bars, and an explanation of implausible metric values. These issues currently limit the paper's significance.","major_comments":[{"comment":"The flagship DiT-XL/2 benchmark (Table 1) uses C=3 and gamma=0.5. With these values, Eq. (8) gives weights [1/2, -1, 1/2], which are exactly the barycentric weights of Lagrange interpolation at three equally spaced nodes. Consequently, Eq. (7) simplifies algebraically to the unique quadratic polynomial through the three cached features; the denominator is a removable artifact and the predictor is not a genuine rational function. Thus every BRACE row in Table 1 tests a feature-driven quadratic extrapolator, not the barycentric rational mechanism that the paper's motivation (Section 3.2) and stability analysis (Section 3.4) rely on. The observed FID/sFID gains over TaylorSeer and HiCache in Table 1 therefore do not provide evidence for the paper's central claim that rational self-normalization absorbs sharp feature irregularities. This is an internal inconsistency between the tested configuration and the proposed mechanism, and it needs to be resolved: either report DiT results with a non-degenerate rational setting (e.g., C=2, or C=3 with gamma != 0.5), or explicitly reframe the DiT results as a polynomial extrapolation baseline.","section":"Sections 3.3.3-3.3.4, Table 1"},{"comment":"The proof of the key lower bound |D(s)| >= lambda/k for all sigma in (0,1] is deferred to a 'Supplemental Material' that is not included with the arXiv submission. Since this bound is the basis for the claimed O(k * L) stability guarantee, the central theoretical justification is unverifiable from the manuscript. Moreover, the asserted existence of a universal lambda > 0 for arbitrary gamma in Eq. (8) is not obvious and is in fact false for some parameter values: for C=3 and gamma=0.1, the denominator polynomial 0.5*sigma*(sigma+1) - sigma*(sigma+2) + 0.1*(sigma+1)*(sigma+2) changes sign on (0,1], so D(s) has a pole in the extrapolation interval. The paper needs to state the precise conditions on gamma and C under which Eq. (11) holds, and the proof must be included in the manuscript.","section":"Section 3.4, Eq. (11)"},{"comment":"No code, project-page link, or evaluation protocol details are provided, and no error bars, standard deviations, or multiple-seed results are reported for any of the FID, ImageReward, or VBench numbers. The differences in Table 1 are small (e.g., FID 2.46 vs. 2.51 at 3.56x), so without variance estimates the claimed consistent improvements are not fully convincing. In addition, Table 2 reports ImageReward 1.0021 for BRACE at 4.16x and 0.9919 for TaylorSeer, both higher than the unaccelerated 50-step FLUX reference (0.9898); it is implausible that a caching method that skips most of the network improves semantic alignment over the full model, and the discrepancy suggests a mismatch in evaluation (e.g., different prompts, seeds, or metric versions) that must be clarified.","section":"Tables 1-3 and Reproducibility"}],"minor_comments":[{"comment":"Line 7 of Algorithm 1 notes that the target domain mapping yields |x(t)| > 1, but Eq. (6) maps cached timesteps only to [-1,1]; the text should state explicitly that extrapolation targets lie outside the canonical interval and explain why this is consistent with the stability analysis.","section":"Algorithm 1, Section 3.3.2"},{"comment":"The reference row for FLUX.1-dev 50 steps contains the garbled entry '0.8800∞1.0000 0.0000'; this appears to be a formatting error (likely a misplaced symbol and missing separator) and should be corrected.","section":"Table 2"},{"comment":"FOCA and SpeCA are discussed in the related work as recent forecasting-based methods, but they are not included in any experimental comparison; the authors should either add them as baselines or justify their omission.","section":"Section 2.2 and Experiments"},{"comment":"Several text elements in Figures 2 and 3 render as broken glyph sequences such as '/uni00000013/uni00000014...'; the figures need to be regenerated with a correct font embedding.","section":"Figures 2 and 3"},{"comment":"The ablation of the boundary sensitivity gamma is performed only for FLUX (C=2). Since the DiT-XL/2 default uses C=3 and gamma=0.5, it would be informative to also ablate gamma for C=3 on DiT to show whether non-degenerate rational configurations are competitive in the headline benchmark.","section":"Table 5"},{"comment":"The claim that for C<=3 the Chebyshev-Lobatto nodes coincide with equidistant points is correct, but then the adapted weights with gamma are no longer the standard Chebyshev-Lobatto weights; the text should acknowledge this and provide a rationale beyond the empirical tuning in Table 5.","section":"Section 3.3.4"}],"recommendation":"major_revision","confidential_remarks":"The paper cites FOCA and SpeCA in the related work but never compares against them, and the project page link is absent from the submission. The missing supplement for the stability proof is a concern that the editor may want to verify before requesting a revision. The empirical core of the paper—the efficiency gains on DiT, FLUX, and HunyuanVideo—may be salvageable, but the narrative needs substantial reworking to align with what is actually tested."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the paper has a genuinely new idea—barycentric rational extrapolation for DiT feature forecasting—and the experimental comparisons look fair. But the stress-test is right: the headline ImageNet results (C=3, gamma=0.5) give weights [1/2, -1, 1/2], which are exactly the barycentric weights for Lagrange interpolation on three equally spaced nodes. So Eq. (7) simplifies to a quadratic polynomial in that configuration, and the 'self-normalizing denominator' is a removable artifact. That means Table 1, the paper's main evidence, does not test the rational mechanism it claims. The FLUX and Hunyuan runs use C=2 with gamma 0.7/0.4, which do produce a genuine rational function, but those are secondary and use different tuning.\n\nThat's a load-bearing mismatch, not a quibble. The abstract and Section 3.3.3 promise that the rational denominator absorbs sharp irregularities; the headline experiments show only that a quadratic predictor with these weights slightly beats Taylor/HiCache. The gains are small anyway—FID 2.46 vs 2.51 on ImageNet, VBench 80.14 vs 79.93 on Hunyuan—and there are no error bars and no released code, so I wouldn't put too much weight on the exact numbers.\n\nWhat the paper does well: the empirical analysis of feature trajectories (sharp irregularities, unstable finite differences) is reasonable and matches the motivation; the comparisons against TaylorSeer, HiCache, and FORA are external and appear honestly reported; and the FLUX results consistently improve on HiCache across multiple metrics, which suggests the rational form may have some value there. The stability bound in Eq. (11) is a nice idea but depends on a constant lambda whose proof is deferred to a missing supplement, so as written it's an assertion, not a proof.\n\nThe paper is for the feature-caching community. I'd send it to review but with a clear expectation of heavy revision: the authors need to either run ImageNet with a genuinely rational configuration (e.g., C=2, gamma not 0.5) or reframe the contribution as a polynomial forecasting method with adapted Chebyshev weights. As is, the central claim is unsupported.","headline":"BRACE's central claim is that rational extrapolation tames sharp DiT features, but the main ImageNet experiments use a configuration that reduces to quadratic Lagrange interpolation, so the paper's flagship results don't test its own mechanism.","tokens_in":17523,"tokens_out":5547,"would_cite":false,"duration_ms":47885,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"BRACE replaces derivative-based polynomial extrapolation with a self-normalizing barycentric rational function for diffusion transformer feature caching, and reports superior quality at 3.5x to 5.5x speedups.","keywords":["Diffusion Transformers","feature caching","barycentric rational interpolation","rational extrapolation","Chebyshev weights","training-free acceleration","diffusion model inference","image and video generation"],"falsifier":"Evaluate the denominator $D(\\sigma)=\\sum_{i=1}^{C} w_i/(\\sigma+C-i)$ at the exact weights used in the paper ($C=2$ with $\\gamma=0.7$ and $C=3$ with $\\gamma=0.5$) over a fine grid of $\\sigma\\in(0,1]$; if $|D(\\sigma)|$ approaches zero anywhere, the bound in Eq. (11) cannot hold.","tokens_in":16482,"feed_emoji":"⚡","tokens_out":13549,"duration_ms":102501,"temperature":0.7,"pith_summary":"Diffusion transformers produce high-fidelity images and video but require many sequential network evaluations, and existing cache-then-forecast accelerators predict skipped features with derivative-based polynomial extrapolations that diverge when feature trajectories turn sharply. BRACE is a training-free alternative: it keeps a small sliding window of cached features and predicts the next feature with a barycentric rational function that aggregates raw features directly, using adapted Chebyshev weights instead of finite-difference derivatives. The paper argues that the rational function's self-normalizing denominator absorbs local non-smoothness, so extrapolation error stays bounded by the skip interval and the feature's Lipschitz constant rather than by high-order derivatives. Across DiT-XL/2, FLUX.1-dev, and HunyuanVideo, BRACE reports the best quality–efficiency trade-offs among compared methods at 3.5x to 5.5x speedups, with negligible added computation.","feed_headline":"A rational function tames sharp jumps in diffusion transformer caching","feed_subtitle":"BRACE's self-normalizing rational predictor keeps images and video sharp at 3.5x–5.5x speedups.","key_machinery":"The central object is the barycentric rational extrapolator in its second form: for cached states $(F_{\\tau_j}, x(\\tau_j))$, the prediction at target $x(t_{\\mathrm{pred}})$ is $F_{\\mathrm{pred}} = \\left(\\sum_j w_j F_{\\tau_j}/(x(t_{\\mathrm{pred}})-x(\\tau_j))\\right)/\\left(\\sum_j w_j/(x(t_{\\mathrm{pred}})-x(\\tau_j))\\right)$, with adapted Chebyshev weights $w_j = (-1)^{j+1}\\delta_j$ and $\\delta_j \\in \\{\\frac12, 1, \\gamma\\}$ for the first, interior, and last nodes. An affine domain mapping sends cached timesteps to $[-1,1]$; a FIFO sliding window of capacity $C\\le3$ keeps the extrapolation local and makes the Chebyshev nodes coincide with equidistant points. The mechanism doing the work is the self-normalizing denominator $D(x)=\\sum_j w_j/(x-x_j)$: instead of estimating derivatives, the predictor is a rational blend of raw features, and the paper's Eq. (11) shows the error is controlled by the skip interval $k$ and the Lipschitz constant $L$ whenever the denominator stays bounded below by a positive structural constant $\\lambda$. That denominator is what absorbs sharp irregularities and prevents polynomial overshoot.","core_discovery":"DiT feature trajectories, though globally smooth, are punctuated by localized sharp irregularities, and the paper claims that rational extrapolation is structurally better suited than polynomial extrapolation for forecasting across those transitions. BRACE predicts the next feature map with the second barycentric form $$F_{\\mathrm{pred}}=\\frac{\\sum_{j=1}^{C} w_j\\, F_{\\tau_j}/(x(t_{\\mathrm{pred}})-x(\\tau_j))}{\\sum_{j=1}^{C} w_j /(x(t_{\\mathrm{pred}})-x(\\tau_j))},$$ with adapted Chebyshev weights $w_j=(-1)^{j+1}\\delta_j$, $\\delta_j\\in\\{\\tfrac12,1,\\gamma\\}$ for first, interior, and last nodes. No derivatives are estimated: the predictor is a weighted rational blend of raw cached features, and the denominator self-normalizes. The stability analysis (Eq. 11) bounds the pointwise error by $(\\sum_i |w_i|\\, L/\\lambda)\\, k$, linear in the skip interval $k$ and the Lipschitz constant $L$, in contrast to Taylor extrapolation's remainder $O(s^{m+1}\\|F^{(m+1)}\\|_\\infty)$, which explodes when high-order derivatives spike at sharp turns. The paper reports that this rational predictor outperforms TaylorSeer, HiCache, and FORA on DiT-XL/2, FLUX.1-dev, and HunyuanVideo at matched 3.5x–5.5x speedups.","pith_inferences":["In the headline DiT-XL/2 configuration ($C=3$, $\\gamma=0.5$), the adapted Chebyshev weights coincide exactly with Lagrange weights, so the barycentric formula reduces to quadratic polynomial interpolation; the rational structure is therefore exercised mainly in the $C=2$ FLUX and HunyuanVideo settings.","The unproved $\\lambda>0$ constant is the hinge of the theoretical stability argument, so the supplement's proof of that bound is the first thing to check before the claimed error guarantee can be accepted.","The same barycentric forecast recipe could be tested on other sequence-prediction problems with smooth-but-spiky trajectories, such as video frame extrapolation or ODE trajectory forecasting, where derivative-based extrapolators face the same instability."],"forward_implications":["At matched FLOPs, BRACE achieves lower FID and sFID than TaylorSeer and HiCache on ImageNet with DiT-XL/2 across 3.56x to 4.98x speedups.","On FLUX.1-dev text-to-image at up to 5.55x speedup, BRACE keeps ImageReward and CLIP scores near the full 50-step baseline while reuse- and polynomial-based baselines drop.","On HunyuanVideo at 5.56x speedup, BRACE reaches a VBench score of 80.14, the closest to the 80.68 of the unaccelerated 50-step model.","Because forecasting uses only a FIFO of cached features and an affine mapping, the added compute is negligible and no retraining is required.","The error bound in Eq. (11) predicts that BRACE's degradation grows linearly with the skip interval rather than with high-order derivatives, which is the formal reason it can sustain aggressive skip intervals."],"supporting_citations":[{"why":"Supplies the second barycentric form and the Chebyshev–Gauss–Lobatto weight structure that BRACE adapts.","marker":"[2]"},{"why":"Documents the numerical instability of barycentric extrapolation that motivates the sparse cache and adapted weights.","marker":"[42]"},{"why":"Defines the derivative-driven Taylor forecast paradigm that BRACE replaces and is the primary baseline it outperforms.","marker":"[21]"},{"why":"Provides the scaled-Hermite polynomial baseline used as a competing cache-then-forecast method.","marker":"[8]"},{"why":"Represents the cache-and-reuse baseline (FORA) that BRACE compares against on all three architectures.","marker":"[35]"},{"why":"Defines the Floater–Hormann rational weights used in the ablation that isolates the effect of the adapted Chebyshev weights.","marker":"[9]"},{"why":"Introduces the DiT architecture whose feature trajectories are analyzed and accelerated.","marker":"[31]"},{"why":"Supplies the phase-transition view of diffusion latent dynamics that motivates the sharp-irregularity observation.","marker":"[24]"},{"why":"Provides the FLUX.1-dev text-to-image model used for ImageReward, CLIP, and CycleReward evaluation.","marker":"[17]"},{"why":"Provides HunyuanVideo, the text-to-video model used for the VBench evaluation.","marker":"[15]"}],"fun_headline_variants":["Rational forecasting beats polynomial caching for diffusion transformers","BRACE tames sharp DiT features with rational blend of cached data","Rational caching predictor smooths diffusion transformer speedups","BRACE rational blend of cached features accelerates DiTs without quality loss","Barycentric rational forecasting tames sharp DiT caching jumps"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The stability guarantee rests on two premises: the feature trajectory is Lipschitz continuous, and a structural constant $\\lambda>0$ keeps the barycentric denominator bounded away from zero for every skip ratio in $(0,1]$; the proof of the latter is deferred to a supplement not included in the preprint.","fun_headline_variants_meta":{"raw":{"variants":["Rational forecasting beats polynomial caching for diffusion transformers","BRACE tames sharp DiT features with rational blend of cached data","Rational caching predictor smooths diffusion transformer speedups","BRACE rational blend of cached features accelerates DiTs without quality loss","Barycentric rational forecasting tames sharp DiT caching jumps"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001084,"raw_usage":{"total_tokens":4579,"prompt_tokens":1042,"completion_tokens":3537,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":658,"completion_tokens_details":{"reasoning_tokens":3455}},"tokens_in":658,"tokens_out":3537,"duration_ms":23374,"temperature":1.0,"reasoning_tokens":3455,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T00:33:06.667247+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Evaluate the denominator $D(\\sigma)=\\sum_{i=1}^{C} w_i/(\\sigma+C-i)$ at the exact weights used in the paper ($C=2$ with $\\gamma=0.7$ and $C=3$ with $\\gamma=0.5$) over a fine grid of $\\sigma\\in(0,1]$; if $|D(\\sigma)|$ approaches zero anywhere, the bound in Eq. (11) cannot hold.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the second barycentric form and the Chebyshev–Gauss–Lobatto weight structure that BRACE adapts."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the scaled-Hermite polynomial baseline used as a competing cache-then-forecast method."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the Floater–Hormann rational weights used in the ablation that isolates the effect of the adapted Chebyshev weights."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the phase-transition view of diffusion latent dynamics that motivates the sharp-irregularity observation."}],"review_version":1}