{"id":"8c02e375-f01d-4de6-afe2-303042121cc8","arxiv_id":"2608.12037","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":7,"one_line_summary":"Clustered α-smoothing preserves multiple modes in randomized smoothing by smoothing each cluster separately and certifies the smoothed mixture with a probabilistic lower bound.","lead":"This paper proposes clustered α-smoothing, a modification of randomized smoothing that preserves multiple modes in stochastic predictions while keeping robustness guarantees. The approach clusters prediction samples, smooths each cluster separately, and combines the results as a mixture, with reported gains on trajectory prediction and quadrotor control.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Lemma A.1's proof that the α-trimmed mean stays in R_m via convexity is invalid: coordinate-wise order statistics are not sample points, so Theorem 3.1 is unproven for general convex sets.","rationale":"The reader's verdict rejects the paper primarily on post-selection bias. While that concern is valid, the more load-bearing flaw is in the proof of the theorem itself: the certificate in Theorem 3.1 depends on Lemma A.1, and that lemma's proof contains a false assertion. Coordinate-wise trimming creates order-statistic vectors that are not elements of the sample and need not belong to a general convex R_m. Thus the claimed lower bound on the probability of the smoothed prediction lying in the union of regions is unsupported for convex regions. Even if one repaired the statistical estimation by sample splitting, the theorem's proof would still be invalid unless the regions are restricted to shapes where the coordinate-wise trimmed vectors remain inside (e.g., axis-aligned hyperrectangles). The paper's experiments use hyperrectangles, which suggests the practical results might survive a restriction, but the claimed theoretical result as stated does not. This independently justifies rejection and cannot be fixed by the sample-splitting repair alone.","tokens_in":23584,"tokens_out":14919,"duration_ms":130464,"concrete_test":"Re-derive the bound in Lemma A.1 by explicitly analyzing the coordinate-wise α-trimmed mean. Construct a 2-D example with R_m the unit disk, s=5, α=0.2 (so ⌊αs⌋=1), with four samples strictly inside the disk and one outside; enumerate possible configurations of the order-statistic vectors and check whether the trimmed mean can fall outside the disk while z=4≥s−⌊αs⌋. If any such configuration exists, the key step in Lemma A.1 is false. If none exists for the disk, try another non-rectangular convex set (e.g., a rotated rectangle), since the theorem claims all convex R_l. A successful counterexample would establish that Theorem 3.1 as stated is not proven.","verdict_should_be":"UNCHANGED","load_bearing_attack":"In Lemma A.1 (Appendix A.1), the proof lower-bounds P(˜H_{N,α,V}(x+δ)∈R_m) by the probability that at least s−⌊αs⌋ of the s samples in cluster m fall in R_m, and then asserts that in this event “the average in (2) only contains points in R_m”, so by convexity the trimmed mean lies in R_m. This is false for general convex R_m. The trimming in Definition 3 is applied coordinate-wise to the vector-valued samples: H_{V_m}(x)_{(i)} is the vector of i-th order statistics of each coordinate, not a sample point. These order-statistic vectors can combine the x-coordinate of one sample with the y-coordinate of another. Hence even when z≥s−⌊αs⌋ samples lie in R_m, the surviving order-statistic vectors can lie outside R_m. Convexity of R_m therefore does not imply the trimmed mean is in R_m, and the inequality in (22) does not follow. For axis-aligned hyperrectangles the step can be repaired (an outside point is extreme in some coordinate and is trimmed), but Theorem 3.1 explicitly assumes only convex R_l. Thus the paper's central theoretical guarantee is unsupported for the claimed generality. This is independent of the post-selection issue in Proposition 4.1.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes clustered α-smoothing for stochastic predictors: given noisy forward samples, it clusters the outputs, applies coordinate-wise α-trimmed averaging within each cluster, and combines the per-cluster predictors as a mixture. It claims a certified lower bound (Theorem 3.1) on the probability that the smoothed prediction lies in a union of convex region subsets of a fixed partition of the output space, under adversarial L2 perturbation of the input. The bound is instantiated in practice via a Clopper-Pearson procedure (Proposition 4.1) and Algorithms 1–2, which construct the partition and coverage sets from samples. The empirical section reports improved Wasserstein distance in trajectory prediction and reduced collision rate in quadrotor control relative to α-smoothing and RS-Reg.","tokens_in":23945,"tokens_out":7640,"duration_ms":79329,"significance":"The problem is well motivated: multi-modal stochastic predictors are common, and standard randomized smoothing collapses their modes. The per-mode smoothing idea is natural, and the empirical results are suggestive, with clearly described benchmarks and released code. If the certificate in Theorem 3.1 and the confidence procedure in Proposition 4.1 were both valid, this would be a useful contribution to certified regression for multi-modal predictors. However, the certificate is not established for the claimed generality and the confidence procedure is invalid as used, so the central claims of the paper are not supported in their current form.","major_comments":[{"comment":"The proof of Lemma A.1 is invalid for general convex sets R_m. In Definition 3, the α-trimming in (2) is applied coordinate-wise, so the vectors H_{V_m}(x)_{(i)} are coordinate-wise order statistics and need not be elements of the original sample set {h_w(x+ε_i)}. Therefore, from Z_m ≥ s − ⌊αs⌋ it does not follow that 'the average in (2) only contains points in R_m': coordinate-wise trimming can keep a coordinate contributed by a point outside R_m while discarding coordinates of points inside R_m. Convexity of R_m does not imply that the coordinate-wise trimmed mean lies in R_m, so the inequality in (22) is unsupported. The argument can be repaired for axis-aligned hyperrectangles, but Theorem 3.1 explicitly assumes only convex R_l, and the proof of Theorem 3.1 relies on Lemma A.1. Thus the central robustness guarantee is unproven at the claimed level of generality. This issue is independent of the post-selection problem in Proposition 4.1.","section":"Appendix A.1, Lemma A.1 and Theorem 3.1"},{"comment":"The high-confidence bounds are not valid as used. Proposition 4.1 is a correct Clopper-Pearson statement for a fixed, data-independent partition V and fixed sets R_l, but Algorithm 1 constructs R_m to cover a fraction p of the observed samples (line 9), and Algorithm 2 then computes the counts Z_{V_m}, Z_{R_m} and the Clopper-Pearson bounds from the same samples. Because the sets R_m and the induced partition V are selected using the same data that are subsequently used for the binomial counts, Z_{R_m} is biased upward and the nominal 1−β confidence of the bounds does not hold. The statement after Proposition 4.1 that 'after observing the samples, one may place the sets V_m and R_m so that the number of outputs inside them (approximately) induces the desired probability level' is precisely the invalid selection step. A valid procedure would require sample splitting or a post-selection adjustment. This invalidates the end-to-end certificate returned by Algorithm 2.","section":"Section 4, Proposition 4.1 and Algorithms 1–2"}],"minor_comments":[{"comment":"The phrase 'where the i-th ordering statist in H_{V_m}(x)_{(i)} is taken element-wise' contains a typo ('statist') and the notation is not defined explicitly; please define the coordinate-wise order-statistic vector clearly.","section":"Section 3, Definition 3"},{"comment":"The sentence mentioning 'correction to the fact that V is a partition of X' should refer to the output space Y, not the input space X.","section":"Section 3, Theorem 3.1"},{"comment":"The notation for the lower bounds yp_{R_l}, yp_{V_l} and their Neyman-Pearson-adjusted counterparts is inconsistent: in particular, the residual term in (8) uses p_{V_l} − p_{R_l} while Lemma A.3 is stated with a generic yq{ytilde p}; please make the relationship explicit.","section":"Equations (5)–(8) and Lemma A.3"},{"comment":"The 'Require' line allows α∈[0,1], but Definition 3 requires α∈[0,1/2); state the domain consistently.","section":"Algorithm 2"},{"comment":"The requirement that the same partition be used for all x′ in the δ-neighborhood is not shown to be enforced by Algorithm 1, which constructs V locally at x; please clarify how the algorithm guarantees this.","section":"Footnote 2"},{"comment":"Appendix F is a detailed analysis of a quantization artifact of α-smoothing, but it is not referenced by the main results; consider moving it to supplementary material or connecting it explicitly to the design choices in Section 5.","section":"Appendix F"}],"recommendation":"reject","confidential_remarks":"The two major issues are independent and both affect the paper's central claims. The convexity problem could in principle be fixed by restricting the certificate to axis-aligned hyperrectangles, and the selection problem by sample splitting or post-selection inference, but either change would require restating the main theorem and reworking the certification pipeline. I do not see a path to acceptance of the current version without a substantial rewrite."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a genuinely new framework—cluster the noisy outputs, trim per cluster, smooth locally, combine as a mixture—and the empirical story on trajectory prediction and quadrotor control is engaging. But the main theorem as stated is not proven, for two independent reasons, and the high-confidence bounds in Proposition 4.1 are also invalid as used. I'd still send it out; the core idea is good and the repairs are plausible.\n\nWhat's new and worth credit: clustering in randomized smoothing has been used before for LLMs, but preserving all clusters as certified modes for regression and control is a real extension. The per-cluster alpha-trimming, the mixture formulation, and the LP relaxation via anchor points are clever. Appendix F, on the quantization artifact of alpha-smoothing for bimodal outputs, is a thoughtful piece of analysis that shows the authors understand the failure mode they are targeting.\n\nWhere it breaks. First, Lemma A.1. The proof assumes that if enough samples lie in a convex set R_m, then the trimmed mean lies in R_m. But trimming is coordinate-wise: the surviving \"samples\" are vectors of order statistics, not actual sample points. Those vectors can combine coordinates from different samples and leave an arbitrary convex set. The inequality only follows for axis-aligned hyperrectangles. Since Theorem 3.1 claims arbitrary convex R_l, the central certificate is unsupported at the claimed generality. This is not a nitpick; it is a load-bearing step.\n\nSecond, the confidence guarantee. Algorithm 1 constructs the partition and the coverage regions R_m using the same samples that Algorithm 2 then feeds into Clopper-Pearson bounds. That is post-selection inference; the binomial count is biased upward and the claimed 1-beta confidence doesn't hold. A simple sample split would fix it, but as written the paper's signature guarantee is not valid.\n\nAlso worth noting, though minor by comparison: the quadrotor collision-rate reduction of 81% comes from a single run, with no seeds or variance reported. It is anecdotal.\n\nWhat holds up: the problem is real, the method is intuitive, the code is available, and the limitations section is honest about the convex-set restriction and the fact that certification is for the smoothed predictor, not the base. The trajectory Wasserstein results look plausible and reproducible.\n\nWho this is for: anyone working on randomized smoothing for regression or control in safety-critical settings. It deserves a serious referee, not a desk reject, because the flaws are identifiable and repairable. My recommendation: send it out, but with an expectation that the theory must be corrected—either by restricting the theorem to axis-aligned hyperrectangles and adding the missing proof, or by finding a different argument—and that the confidence bounds must use held-out samples. With those fixes, it could become a solid paper.","headline":"Nice idea with two independent theory bugs; both look fixable, so it deserves a rigorous review rather than a desk reject.","tokens_in":24465,"tokens_out":2400,"would_cite":false,"duration_ms":24245,"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":"Clustered $\\alpha$-smoothing certifies robust multi-modal predictions for stochastic predictors: rather than averaging all noisy outputs, it smooths each mode separately and mixes the results, with a formal lower bound on the probability…","keywords":["clustered α-smoothing","randomized smoothing","stochastic predictors","multi-modal regression","robustness certification","mode collapse","trajectory prediction","quadrotor control"],"falsifier":"Run Algorithm 2 on a predictor with two well-separated modes using a small $\\beta$ (e.g., $0.05$) and count, over many independent noise and weight draws, how often $\\tilde H_{N,\\alpha,V}(x+\\delta)$ actually lands in the certified region $\\tilde R$ for a fixed $\\delta$ with $\\|\\delta\\|_2\\le r$. If the empirical frequency falls below the certified lower bound even at the nominal $1-\\beta$ confidence level, the guarantee as stated fails, because the region was chosen from the same data used to estimate the binomial proportions; repeating the check with counts from an independent sample would isolate the cause.","tokens_in":23381,"feed_emoji":"🎯","tokens_out":8689,"duration_ms":81787,"temperature":0.7,"pith_summary":"The paper sets out to fix a failure of randomized smoothing for stochastic predictors whose outputs are multi-modal: averaging over noisy samples collapses distinct behaviors into a single unrepresentative mean. Its proposal, clustered $\\alpha$-smoothing, clusters the noisy predictions in output space, applies $\\alpha$-trimmed averaging inside each cluster, and recombines the per-cluster smoothers as a mixture. The central claim is Theorem 3.1, a lower bound on the probability that the clustered smoother's output at any perturbed input lies in a union of per-mode coverage regions, valid for every perturbation of bounded Euclidean norm. If the claim holds, safety-critical predictors such as trajectory forecasters and robot controllers can retain multi-modal uncertainty and still carry formal robustness certificates; the experiments report a $27\\%$ lower Wasserstein distance in trajectory prediction and an $81\\%$ lower collision rate in quadrotor control relative to state-of-the-art smoothing.","feed_headline":"Smoothing per cluster keeps multi-modal predictions certifiable","feed_subtitle":"Averaging within each mode, not across all outputs, preserves distinct behaviors and certified robustness.","key_machinery":"The central object is the clustered $\\alpha$-smoothed predictor, a mixture of per-cluster trimmed means whose weights are the empirical cluster sizes; the $\\alpha$-trimming removes the $\\lfloor\\alpha |I_m|\\rfloor$ extreme order statistics in each coordinate before averaging. The argument is carried by two standard tools: the likelihood-ratio lemma of randomized smoothing, which shifts binomial success probabilities when the input is perturbed by bounded-norm Gaussian noise (equations (4)), and a binomial confidence bound that turns observed counts into high-confidence intervals for those success probabilities (Proposition 4.1). To make the general multi-mode certificate tractable, the paper replaces the non-convex infimum in Theorem 3.1 with a linear program over anchor points and a controlled residual (Proposition 3.2). The generalized Voronoi construction in Algorithm 1 is what keeps each coverage region $R_m$ inside its partition cell $V_m$, which the theorem requires.","core_discovery":"The paper introduces a new smoothed predictor, $\\tilde H_{N,\\alpha,V}(x)=\\sum_{m=1}^{M}\\mathbf{1}_{z=m}\\tilde H_{N,\\alpha,V_m}(x)$, in which the output space is partitioned into cells $V_m$, each cell receives its own $\\alpha$-trimmed mean over the noisy samples that fall into it, and a categorical variable $z$ with weights $|I_m|/N$ mixes the per-cell components. Against this object, the paper proves that for any subset of modes $L$ and convex coverage regions $R_l\\subset V_l$, the probability that $\\tilde H_{N,\\alpha,V}(x+\\delta)\\in \\tilde R=\\cup_{l\\in L}R_l$ is at least the infimum in (5), provided the per-cell probabilities $p_{V_m}$ and $p_{R_l}$ are known or bounded. The argument combines a per-cell shift bound for Gaussian noise, a union bound over the selected modes, and a linear-program relaxation with anchor points to make the infimum computable. Sample versions of the bounds are obtained from binomial counts, so the whole procedure is implementable for black-box stochastic predictors.","pith_inferences":["A natural extension the paper does not pursue is input-dependent partitions $V(x)$: the current certificate fixes one partition around $x$ and reuses it for the whole perturbation ball, so adapting the modes to each input would require a new concentration argument.","The post-selection gap in the confidence statement (regions and counts from the same samples) could be closed by sample splitting—one batch to construct $V$ and $R$, an independent batch to count the binomial statistics—at the price of roughly doubling inference cost; the paper's Algorithm 2 does not do this.","In closed-loop settings the certificate applies per time step, so composing it over a receding horizon would need a union bound over steps; the quadrotor experiments evaluate the smoothed policy in closed loop but the formal guarantee is claimed per step.","The per-cluster $\\alpha$-trimming suggests an adaptive variant that selects $\\alpha_m$ per mode from local dispersion; nothing in the proof requires a shared $\\alpha$, so this is a testable modification."],"forward_implications":["For any stochastic predictor with multi-modal output, clustered $\\alpha$-smoothing certifies that the prediction under a bounded input perturbation falls in a union of per-mode coverage regions with high probability, so robustness guarantees no longer force mode collapse.","The certificate is agnostic to the clustering algorithm, so DBSCAN, $k$-means, or hand-designed behavioral partitions (as used in the trajectory experiments) all feed into the same bound.","In trajectory prediction, the method lowers the 2-Wasserstein distance between smoothed and original distributions by 27% on average compared to $\\alpha$-smoothing, and lowers the risk rate of a go/no-go decision from 16.5% to 2.5%.","In quadrotor control with a bi-modal policy, clustered smoothing preserves both navigation paths and cuts the collision rate by 81% relative to randomized smoothing, while additionally removing the discrete quantization artifact that $\\alpha$-smoothing exhibits on bimodal outputs.","Because the multi-mode bound uses a union bound over clusters and a confidence interval for each cluster probability, adding more modes makes the certificate more conservative, a trade-off the paper documents empirically."],"supporting_citations":[{"why":"supplies the likelihood-ratio certified-robustness argument that Theorem 3.1 adapts from classification to per-cluster regression bounds.","marker":"[8]"},{"why":"defines $\\alpha$-smoothing, the single-cluster special case (Remark 1) and the main baseline for Wasserstein and collision comparisons.","marker":"[30]"},{"why":"RS-Reg serves as a baseline in trajectory prediction and supplies the observation that one outlier can displace an averaged regression output.","marker":"[31]"},{"why":"TrajFlow is the stochastic trajectory-prediction model used in the first benchmark.","marker":"[26]"},{"why":"provides the quadrotor navigation benchmark with stochastic dynamics and obstacles used in the control experiment.","marker":"[3]"},{"why":"defines generalized Voronoi partitions, used in Algorithm 1 to keep coverage regions inside their partition cells.","marker":"[25]"},{"why":"L-GAP driving-simulator dataset supplies the human left-turn gap-acceptance trajectories for the trajectory-prediction experiments.","marker":"[47]"}],"fun_headline_variants":["Cluster-wise smoothing preserves mode diversity","Per-mode averaging certifies robust mixtures","Clustered alpha-smoothing keeps modes and bounds","Smooth each mode, certify the union","Mode-preserving smoothing with certifiable robustness"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The confidence statement presumes the partition and coverage regions are fixed before the binomial counts are collected; in Algorithms 1 and 2 the same samples choose those regions and produce the counts, so the stated $1-\\beta$ confidence is not justified without sample splitting or a post-selection correction.","fun_headline_variants_meta":{"raw":{"variants":["Cluster-wise smoothing preserves mode diversity","Per-mode averaging certifies robust mixtures","Clustered alpha-smoothing keeps modes and bounds","Smooth each mode, certify the union","Mode-preserving smoothing with certifiable robustness"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000147,"raw_usage":{"total_tokens":1228,"prompt_tokens":1028,"completion_tokens":200,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":644,"completion_tokens_details":{"reasoning_tokens":134}},"tokens_in":644,"tokens_out":200,"duration_ms":2490,"temperature":1.0,"reasoning_tokens":134,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T00:18:45.386777+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run Algorithm 2 on a predictor with two well-separated modes using a small $\\beta$ (e.g., $0.05$) and count, over many independent noise and weight draws, how often $\\tilde H_{N,\\alpha,V}(x+\\delta)$ actually lands in the certified region $\\tilde R$ for a fixed $\\delta$ with $\\|\\delta\\|_2\\le r$. If the empirical frequency falls below the certified lower bound even at the nominal $1-\\beta$ confidence level, the guarantee as stated fails, because the region was chosen from the same data used to estimate the binomial proportions; repeating the check with counts from an independent sample would isolate the cause.","supporting_citations":[{"cited_title":"Certified adversarial robustness via randomizedα-smoothing for regression models.Advances in Neural Information Processing Systems, 37:134127– 134150, 2024","cited_arxiv_id":null,"evidence_quote":"defines $\\alpha$-smoothing, the single-cluster special case (Remark 1) and the main baseline for Wasserstein and collision comparisons."},{"cited_title":"Rs-reg: Probabilistic and robust certified regression through randomized smoothing.Transactions on Machine Learning Research, 2025","cited_arxiv_id":null,"evidence_quote":"RS-Reg serves as a baseline in trajectory prediction and supplies the observation that one outlier can displace an averaged regression output."},{"cited_title":"Schumann, Javier Alonso-Mora, Arkady Zgonnikov, and Jens Kober","cited_arxiv_id":null,"evidence_quote":"TrajFlow is the stochastic trajectory-prediction model used in the first benchmark."},{"cited_title":"Robust control for dynamical systems with non-gaussian noise via formal abstractions.Journal of Artificial Intelligence Research, 76:341–391, 2023","cited_arxiv_id":null,"evidence_quote":"provides the quadrotor navigation benchmark with stochastic dynamics and obstacles used in the control experiment."},{"cited_title":"Robust construction of the voronoi diagram of a polyhedron","cited_arxiv_id":null,"evidence_quote":"defines generalized Voronoi partitions, used in Algorithm 1 to keep coverage regions inside their partition cells."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"L-GAP driving-simulator dataset supplies the human left-turn gap-acceptance trajectories for the trajectory-prediction experiments."}],"review_version":1}