{"id":"7f0bcff2-5257-49a9-ae92-efc7a1ccb038","arxiv_id":"2608.11544","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Adding a Conditional Value-at-Risk tail penalty to a Lipschitz-regularized divergence creates a bounded non-Lipschitz particle flow that improves heavy-tail accuracy of pre-trained generative models.","lead":"Generative models trained on finite data often miss rare extreme events because too few samples represent the tails. This paper adds a tail-risk penalty to the training flow, producing a method that fine-tunes any pre-trained model to better match heavy-tailed data without accessing its internals.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Global radial CVaR penalty fails on nearly Gaussian marginals in anisotropic targets; its scalar, norm-based correction is the weakest load-bearing assumption of the tail-agnostic claim.","rationale":"The reader's weakest_assumption identifies exactly the same load-bearing point: the global scalar CVaR discrepancy with radial risk g(x)=||x|| is applied identically to all coordinates, and the paper's own Section 5.2 shows that this assumption fails for the nearly Gaussian ν=30 marginal. My stress-test confirms this is the most consequential limitation of the central claim. The theoretical apparatus (Clarke subgradients, bounded non-Lipschitz velocity) is internally consistent and is not the primary source of risk; the empirical failure on ν=30 is a direct consequence of the scalar, norm-based penalty rather than an implementation bug. The paper acknowledges the failure and proposes a per-coordinate risk as future work, but the abstract and central claims are not restricted to the moderate-heterogeneity regime where the method works. This warrants the existing CONDITIONAL verdict: the contribution is real and the theory is sound, but the tail-agnostic claim needs qualification. I see no basis for REJECT because the method does improve all tested isotropic cases, Neal's funnel, and all Fama-French 25 marginals; the concern is about scope, not correctness. The concrete test I propose would settle whether the failure is fully attributable to the scalar radial risk by checking the sign of per-marginal CVaR discrepancies and by testing a per-coordinate variant. Since the reader already reached CONDITIONAL and my analysis does not move the verdict, I mark it UNCHANGED.","tokens_in":27822,"tokens_out":6867,"duration_ms":82829,"concrete_test":"In the Section 5.2 anisotropic setting, compute per-marginal CVaR discrepancies ΔC_j = CVaR_{P_tar, |x_j|}^{α} - CVaR_{Q, |x_j|}^{α} for j=1,...,5 at the initial pre-trained distribution and during training. If sign(ΔC_30) is opposite to sign(ΔC_global) = CVaR_{P_tar, ||x||}^{α} - CVaR_{Q, ||x||}^{α}, this confirms the radial update pushes the ν=30 marginal in the wrong direction. Then rerun CVaR-GPA with a per-coordinate risk g_j(x)=|x_j| and per-coordinate CVaR penalty, and check whether the ν=30 marginal errors decrease. If they do, the failure is caused by the scalar radial risk; if they do not, the issue is deeper than the choice of risk function. Either outcome settles whether the current central claim requires a meaningful restriction to moderate tail heterogeneity.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central algorithmic claim rests on the CVaR discrepancy term in Eq. (2.4) using a single global, scalar risk function g(x)=||x||, with the velocity correction in Eq. (3.39)/Algorithm 2 line 11 applied radially to every particle. The paper's own Section 5.2 and Figure 5 show that for a 5-d anisotropic Student-t target with ν=(1,1.5,3,10,30), the nearly Gaussian ν=30 marginal gets worse under CVaR-GPA in both E_L1 and E_tail. The authors attribute this to the global CVaR correction being dominated by heavier-tailed directions. This is not a peripheral artifact: it directly contradicts the abstract's 'tail-agnostic' and 'robust' claims for anisotropic targets, and it shows that the scalar radial CVaR statistic is not a sufficient summary of tail structure when marginal tail behaviors are highly heterogeneous. The theory in Theorems 3.8 and 3.11 holds for any Q, but it does not guarantee that a single scalar ΔC provides useful directional information; in the ν=30 case it provides information with the wrong sign or magnitude for that marginal. A future per-coordinate risk function is suggested in Section 6, but the present algorithm and the central claim are not qualified accordingly.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a CVaR-penalized loss functional F^CVaR(Q;P_tar) = D^L_KL(Q||P_tar) + λ(ΔC)^2, where ΔC is the difference of CVaR values of a risk function g under Q and the target, and derives the corresponding Wasserstein gradient flow. The main theoretical results are a subgradient calculus for CVaR based on the Rockafellar–Uryasev representation, valid on empirical measures where the classical density-based derivative fails, and an explicit bounded but non-Lipschitz velocity field (Theorem 3.8, Corollary 3.11). The paper then presents the particle algorithm CVaR-GPA, which fine-tunes samples from a pre-trained model, and reports experiments on isotropic and anisotropic Student-t targets, Neal's funnel, and Fama-French 25 portfolio data, comparing global L1 and tail log-CCDF errors against a Lip-KL-GPA baseline.","tokens_in":28118,"tokens_out":5560,"duration_ms":64540,"significance":"If the claims hold, the paper makes a solid theoretical contribution: the first-variation subgradients of CVaR are derived rigorously for empirical measures, the velocity field is explicit and provably bounded, and the proposed algorithm is architecture-agnostic and tail-agnostic in the sense of not requiring an estimated tail index. The authors are also transparent about a known failure mode for a nearly Gaussian marginal in an anisotropic Student-t example. However, the experimental evidence is weakened by the absence of error bars or repeated trials, and the abstract's unqualified 'dramatically improves' and 'tail-agnostic/robust' statements are contradicted by the paper's own ν=30 result in Section 5.2. The central theoretical derivation appears sound, but the empirical scope and the strength of the claims need to be revised.","major_comments":[{"comment":"The reported results for the 5-dimensional anisotropic Student-t target contradict the abstract's unqualified claim that CVaR-GPA 'dramatically improves global and tail accuracy' on anisotropic Student-t targets. For the ν=30 marginal, E_L1 increases from 0.014 to 0.021 and E_tail increases from 0.078 to 0.14. The text acknowledges this and attributes it to the global scalar CVaR correction being dominated by heavier-tailed directions, but the abstract and introduction are not qualified accordingly. Please revise the central claims to state explicitly that the method is effective for moderate tail heterogeneity and may degrade for near-Gaussian marginals in highly anisotropic targets, or modify the algorithm/risk function so that this failure no longer occurs.","section":"§5.2, Figure 5"},{"comment":"All numerical results are presented as single values with no error bars, no number of random seeds, and no measure of variability. Given that the comparisons of E_L1 and E_tail are the main empirical evidence, please report means and standard deviations over multiple runs, or at least provide box plots or confidence intervals. Without this, the improvements in Figures 3 and 4, and the deterioration in Figure 5, cannot be assessed for statistical significance.","section":"§5, Figures 3–5"},{"comment":"The population-level theory requires E_{P_tar}[g] < ∞, which for the chosen radial risk function g(x)=||x|| means the target must have a finite first moment. This excludes the motivating Cauchy (ν=1) target at the population level; the authors note that the empirical target measure automatically satisfies the condition, but the abstract's 'minimal assumptions on the target distribution' and 'tail-agnostic' statements are therefore too strong. Please state explicitly that the population-level well-posedness of F^CVaR requires finite E_{P_tar}[g], and that the algorithm's use on infinite-mean targets rests on the finite-sample formulation only.","section":"Assumption 2.1, §5.1"},{"comment":"Both error metrics E_L1 and E_tail are defined through the radial CCDF of ||x||, which is the same risk function g used in the CVaR penalty (Eq. (1.3) and Corollary 3.11). Consequently, part of the reported tail improvement is aligned with the penalized quantity by construction. The per-marginal evaluations in Figures 4 and 5 provide some independent evidence, but the paper should either report a metric based on a different risk function or explicitly acknowledge that the tail-fidelity claim is demonstrated for the radial statistic used in the penalty.","section":"§5, Eqs. (5.45)–(5.46)"}],"minor_comments":[{"comment":"The notation in Eq. (2.9), CVaR^{Q,g}_α := β VaR^{Q,g}_α + (1−β) CVaR^{Q,g}_α, uses the same symbol CVaR^{Q,g}_α on both sides, making the definition self-referential. Please denote the upper conditional expectation, E_Q[g(X)|g(X)>VaR^{Q,g}_α], by a different symbol such as CVaR^+ or CVaR^{Q,g,+}_α.","section":"Eq. (2.9)"},{"comment":"The text in Section 5 states that the outer VaR endpoint is used, but Algorithm 2 line 8 allows either 'VaR or VaR'. Please make the choice explicit in the algorithm or its caption so the experiments are exactly reproducible.","section":"Algorithm 2, line 8"},{"comment":"The numbers above the bars in Figure 2 are dense and hard to read at log scale. Consider moving the exact values to a table or using larger fonts and fewer decimal places.","section":"Figure 2"},{"comment":"The paper does not state the computational overhead of CVaR-GPA relative to Lip-KL-GPA. Since the added cost is mainly sorting and scalar tail statistics, a short remark on runtime or complexity would help practitioners.","section":"Section 4"}],"recommendation":"major_revision","confidential_remarks":"The paper is within scope and the theoretical derivation appears sound. The main concern is the mismatch between the abstract's strong claims and the empirically reported degradation on the ν=30 marginal, together with the absence of error bars. These issues are fixable by qualification and additional experiments, so I recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. First, the subgradient calculus for CVaR on empirical measures is real and worth having; Theorem 3.5's formula is correct and closes a gap left by the density-based derivations in prior work. Second, the paper's \"tail-agnostic\" claim needs a qualifier: their own anisotropic Student-t experiment shows the global radial penalty degrades the nearly Gaussian ν=30 marginal, and they admit it. A single scalar CVaR discrepancy cannot summarize heterogeneous marginal tail behavior; the per-coordinate risk function floated in Section 6 is the obvious fix and should have been part of the experiments.\n\nThe theory is the strong part. The velocity field (3.36) is explicit, bounded, and non-Lipschitz, and the derivation of first variational subgradients via the Rockafellar-Uryasev representation on empirical measures is careful and, as far as I can tell, correct. The algorithm CVaR-GPA is simple and genuinely tail-agnostic in the limited sense that it never estimates the tail index. On heavy-tailed marginals, the empirical improvements over Lip-KL-GPA on Student-t, Neal's funnel, and Fama-French are consistent and plausible.\n\nSoft spots, in proportion. The stress-test concern is on target: the ν=30 result is not peripheral, because the abstract's \"robust, tail-agnostic\" language is broader than what the algorithm delivers. No error bars appear anywhere, and no code is released; for a method with neural-network training and several hyperparameters, that makes the empirical claims harder to judge. The comparison is only against the authors' own Lip-KL-GPA baseline, not against other tail-oriented fine-tuning methods, so the practical advantage is not fully established. The finite-first-moment theory is technically bypassed for the Cauchy case by working with empirical measures, but the paper is honest about this. The circularity concern about using ||x|| in both the penalty and the evaluation metric is minor; the metric is a reasonable way to measure tail accuracy, and the penalty is designed to move particles outward.\n\nOverall, the central argument holds: the subgradient result is publishable on its own, and the flow formulation is a useful contribution. The authors need to add error bars, release code, and reword the tail-agnostic claim to acknowledge the ν=30 failure and the need for per-coordinate risk. I would send this to peer review and treat it as conditionally acceptable after those fixes.","headline":"A genuinely useful new subgradient calculus for CVaR on empirical measures, packaged as a fine-tuning algorithm that works well on heavy tails but overstates its tail-agnostic robustness: the global radial penalty fails on the nearly Gaussian marginal of their own anisotropic test.","tokens_in":28658,"tokens_out":1910,"would_cite":true,"duration_ms":20993,"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":"CVaR penalty restores the vanishing tail velocity in generative particle flows.","keywords":["heavy-tailed distributions","extreme events","Conditional Value-at-Risk","Wasserstein gradient flows","Lipschitz-regularized KL divergence","particle algorithms","generative model fine-tuning","tail-agnostic learning"],"falsifier":"The paper's own 5-dimensional Student-t experiment is the concrete check: with tail indices $\\nu=(1,1.5,3,10,30)$, the nearly Gaussian $\\nu=30$ marginal's global and tail errors increase after fine-tuning. Repeating that experiment at larger sample sizes and with several seeds, and testing whether a per-coordinate risk function removes the degradation, would settle whether the single-global-CVaR mechanism is the limiting component of the method.","tokens_in":27641,"feed_emoji":"📉","tokens_out":11870,"duration_ms":116311,"temperature":0.7,"pith_summary":"The paper proposes CVaR-GPA, a particle fine-tuning algorithm that improves a pre-trained generative model's accuracy on heavy-tailed targets without being told the target's tail index. Its loss is the Lipschitz-regularized KL divergence plus a squared Conditional Value-at-Risk (CVaR) discrepancy, and the Wasserstein gradient flow of this loss has a bounded but non-Lipschitz velocity field that keeps pushing particles outward in the tail after the KL-only velocity has faded. The authors derive the variational subgradients of CVaR from its infimum representation, a step needed because the classical density-based first variation does not exist for empirical measures. On isotropic and anisotropic Student-$t$ targets, a funnel benchmark, and a real 25-dimensional portfolio-return dataset, the fine-tuned particles reduce global $L^1$ and tail log-CCDF errors relative to the Lipschitz-regularized baseline. The paper also reports the method's limitation: on a 5-dimensional Student-$t$ with one nearly Gaussian marginal, that marginal's errors increase.","feed_headline":"CVaR penalty restores the vanishing tail velocity","feed_subtitle":"A bounded non-Lipschitz gradient flow cuts global and tail error on heavy-tailed targets without knowing the tail index.","key_machinery":"The machinery is the CVaR-penalized potential $\\Phi^y_Q = \\phi^* - \\frac{2\\lambda}{1-\\alpha}\\,\\Delta C(Q;P_{\\mathrm{tar}})\\,(g-y)_+$, where $\\phi^*$ is the optimal $L$-Lipschitz potential defining the Lipschitz-regularized KL divergence, $g(x)=\\|x\\|$ is the radial risk function, and $y$ ranges over the quantile interval $T(Q)$. Choosing $-\\nabla \\Phi^y_Q$ as the velocity gives a bounded, non-Lipschitz field: the KL part contributes $-\\nabla\\phi^*$ everywhere, and the CVaR part adds an outward radial push of size $\\frac{2\\lambda}{1-\\alpha}|\\Delta C|$ exactly where $\\|x\\|>y$. The subgradient theorem provided in the paper is the essential step: for empirical measures, where the CDF of $g$ is a step function, the classical formula $(g-\\mathrm{VaR}_\\alpha)_+/(1-\\alpha)$ fails because the quantile interval is non-singleton, and every $y\\in T(Q)$ yields a valid variational subgradient.","core_discovery":"The central discovery is that a squared CVaR discrepancy term appended to the Lipschitz-regularized KL divergence repairs the premature vanishing velocity of Lipschitz-regularized Wasserstein gradient flows. The CVaR term contributes a radial velocity component, active only beyond a threshold radius, whose magnitude is proportional to the gap between the target's CVaR and the generated distribution's CVaR; this component is nonzero in the tail even when sample scarcity has made the KL-derived velocity vanish, and it tapers off as the distribution approaches the target. The paper proves that the resulting velocity field is bounded yet not Lipschitz, which is the kind of transport map needed to move a light-tailed pre-trained distribution toward a heavier-tailed target. It also proves that the penalized loss is a divergence whose unique global minimizer is the target distribution, and that the subgradient family indexed by thresholds in the quantile interval $T(Q)$ supplies the correct first-variation objects for empirical measures.","pith_inferences":["If the limiting mechanism is the single radial CVaR statistic, replacing $g(x)=\\|x\\|$ with a per-coordinate risk function should remove the $\\nu=30$ degradation while keeping the heavy-marginal gains; this is the paper's own stated future direction.","The tail-agnostic claim should be read as holding for the algorithm's empirical objective: the population-level theory needs a finite first moment, but the practical algorithm inherits finite support from its samples, which is why the Cauchy target can be handled.","Monitoring the jump in the velocity field across the threshold shell could serve as a live diagnostic of unresolved tail discrepancy, since the jump magnitude is proportional to the current CVaR gap.","Comparing inner- versus outer-endpoint choices for the activation threshold $y\\in T(Q)$ would probe whether the activation region's size trades global accuracy against tail accuracy; the paper leaves this as a design choice."],"forward_implications":["One fine-tuning run with the same hyperparameters improves both global $L^1$ error and tail log-CCDF error across targets with different tail indices, without estimating the tail decay rate.","The kinetic-energy stopping rule makes the effective flow depth depend on the target, not on a preset number of layers or iterations.","The algorithm needs only samples from the pre-trained model, so it can fine-tune any black-box generator without architectural access.","The bounded velocity keeps particle updates stable even when the target's population distribution has no finite moments, because the empirical target used in the algorithm has finite support.","The non-Lipschitz transport map is what allows a light-tailed pre-trained distribution to move toward a heavier-tailed target; Lipschitz maps would preserve the source's light tail."],"supporting_citations":[{"why":"Supplies the theory and empirical baseline for Lipschitz-regularized divergences and the heavy-tail learning results that CVaR-GPA improves on.","marker":"[9]"},{"why":"Defines the Lipschitz-regularized KL divergence and its variational representation used as the base term of the loss.","marker":"[14]"},{"why":"Introduces the generative particle algorithm and the Lipschitz-regularized baseline whose premature saturation motivates the CVaR penalty.","marker":"[18]"},{"why":"Shows Lipschitz flows preserve light tails, the obstruction that the non-Lipschitz velocity field is designed to break.","marker":"[23]"},{"why":"Gives the infimum representation of CVaR from which the variational subgradients are derived.","marker":"[32]"},{"why":"Prior CVaR-based flow fine-tuning whose density-assumption derivative the paper's subgradient framework repairs.","marker":"[33]"},{"why":"Provides the 25-dimensional monthly portfolio-return dataset used as a real-world benchmark.","marker":"[16]"},{"why":"Provides the funnel distribution used to test a mixture of heavy- and light-tailed marginals.","marker":"[29]"}],"fun_headline_variants":["Tail-agnostic CVaR flow repairs heavy-tail transport","No tail index? CVaR penalty still catches extremes","Bounded non-Lipschitz flow beats tail blindness","CVaR boost revives vanishing tail velocity","Fine-tune to heavier tails without tail guesses"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that one global CVaR discrepancy computed from radial distances, with a single activation threshold, can steer all coordinates of an anisotropic target at once; the paper's own $\\nu=30$ experiment shows this premise can fail when one marginal is nearly Gaussian while others are heavily tailed.","fun_headline_variants_meta":{"raw":{"variants":["Tail-agnostic CVaR flow repairs heavy-tail transport","No tail index? CVaR penalty still catches extremes","Bounded non-Lipschitz flow beats tail blindness","CVaR boost revives vanishing tail velocity","Fine-tune to heavier tails without tail guesses"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000642,"raw_usage":{"total_tokens":3007,"prompt_tokens":1050,"completion_tokens":1957,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":666,"completion_tokens_details":{"reasoning_tokens":1879}},"tokens_in":666,"tokens_out":1957,"duration_ms":15957,"temperature":1.0,"reasoning_tokens":1879,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T00:35:59.135366+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"The paper's own 5-dimensional Student-t experiment is the concrete check: with tail indices $\\nu=(1,1.5,3,10,30)$, the nearly Gaussian $\\nu=30$ marginal's global and tail errors increase after fine-tuning. Repeating that experiment at larger sample sizes and with several seeds, and testing whether a per-coordinate risk function removes the degradation, would settle whether the single-global-CVaR mechanism is the limiting component of the method.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the 25-dimensional monthly portfolio-return dataset used as a real-world benchmark."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the theory and empirical baseline for Lipschitz-regularized divergences and the heavy-tail learning results that CVaR-GPA improves on."},{"cited_title":"Dupuis and Y","cited_arxiv_id":null,"evidence_quote":"Defines the Lipschitz-regularized KL divergence and its variational representation used as the base term of the loss."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces the generative particle algorithm and the Lipschitz-regularized baseline whose premature saturation motivates the CVaR penalty."},{"cited_title":"Jaini, I","cited_arxiv_id":null,"evidence_quote":"Shows Lipschitz flows preserve light tails, the obstruction that the non-Lipschitz velocity field is designed to break."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Gives the infimum representation of CVaR from which the variational subgradients are derived."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Prior CVaR-based flow fine-tuning whose density-assumption derivative the paper's subgradient framework repairs."}],"review_version":1}