{"id":"8eba2bdb-07ac-4146-9819-8ecc6a7b1770","arxiv_id":"1908.08558","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"low","formal_verification":"none","parameter_count":1,"one_line_summary":"Localized conformal prediction reweights training scores by proximity to the test point and adjusts the quantile level to preserve distribution-free finite-sample coverage.","lead":"This paper proposes localized conformal prediction, which builds prediction intervals using mostly training points near the test point while keeping a finite-sample coverage guarantee. It is relevant for anyone using conformal prediction who needs intervals that adapt to local data structure.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Finite-sample guarantee depends on the localizer H being a fixed, permutation-invariant function of the feature multiset; response- or order-dependent H is outside Theorem 3.2 and needs an explicit caveat.","rationale":"I read the paper in good faith and checked the permutation argument in Lemma B.2, the worst-case reduction in Lemma B.3, and the min-over-permutations device in the proof of Theorem A.7. The proofs are internally consistent for the stated class of localizers: H is a function of the feature multiset, H(x,x,X)=1, and the score function is fixed or symmetric in the data set. The conditional-on-the-unordered-set argument correctly handles the fact that the test point has a special role, and the use of min_{l} tilde_alpha_l restores permutation invariance for the data-dependent tilde_alpha produced by Algorithm 1. I could not identify a flaw in the coverage proof itself. The reader's weakest assumption identifies the genuine boundary: the finite-sample guarantee is tied to H being feature-only and permutation-invariant, and it would not cover a localizer built from the test response or from the data ordering. This is a scope limitation, not a mathematical error, so it does not change the conditional verdict; it does suggest that the paper should state the restriction more prominently. The reproducibility gaps noted by the reader (no code, no Monte Carlo error bars) remain the basis for the conditional recommendation.","tokens_in":14,"tokens_out":40329,"duration_ms":528919,"concrete_test":"Run Algorithm 1 on the simulation setting of Example 4.1 with a response-based localizer H_{i,j}=1{|Y_i-Y_j|≤h}, recomputing H and the grid-selected tilde_alpha for every candidate test y, with n=500, 1000 repetitions, and α=0.95; record empirical coverage. If coverage is materially below 0.95, the feature-only restriction in Theorem 3.2 is confirmed as load-bearing.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central theorem is sound for the class of localizers defined in Section 3: H(x1,x2,X) is in [0,1], H(x,x,X)=1, and H depends only on the feature multiset X. The load-bearing step is Lemma B.2, where conditioning on the unordered set T fixes the weight matrix p^H_{i,j} and yields the identity v_i^*(σ)=v_{σ_i}^*; this identity is what converts the permutation average into E[∑_i w_i 1{V_i≤v_i^*}|T]. If a user builds H from the test response Y_{n+1}, for example H_{i,j}=1{|Y_i-Y_j|≤h}, then for each candidate y the weight matrix, the distribution \\hat F, and the selected \\tilde α all change with y. Lemma B.3's monotonicity argument in V_{n+1} with fixed weights no longer applies, and Algorithm 1's single-\\tilde α construction has no stated guarantee. The paper defines H on the feature set only, so this is an explicit scope restriction rather than an internal inconsistency; however, the abstract's 'assumption-free' wording and the lack of a caveat about response-based localization make the limitation easy to miss. This is the boundary of the central claim and should be stated prominently.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes localized conformal prediction (LCP), a modification of conformal prediction in which training samples are reweighted by a localizer H(x_i, x_{n+1}) that emphasizes points near the test feature. For a fixed score function V and a localizer depending only on the feature set, the paper proves finite-sample marginal coverage by replacing the nominal level α with a data-dependent level \\tilde α chosen to satisfy condition (G2) (or with \\tilde α leading to \\bar v^* = ∞). The authors provide a practical Algorithm 1, discuss local and asymptotic conditional coverage, extend the idea to data-dependent scores and to covariate shift, and report simulations showing comparable coverage and more locally adaptive intervals than classical conformal prediction.","tokens_in":21268,"tokens_out":17513,"duration_ms":162565,"significance":"The localization idea is a natural and useful extension of conformal inference, and the paper is the first to provide a finite-sample coverage guarantee under local reweighting. The permutation argument in Lemma B.2 is clean for the symmetric case, and the formulation in terms of \\tilde α is interesting. If the proof of Theorem A.7 can be completed, the result would be a valuable contribution to distribution-free predictive inference. The simulations support the qualitative claim that LCP adapts to heteroscedasticity, and the explicit treatment of covariate shift in the appendix extends the reach of the method.","major_comments":[{"comment":"The step from Eq. (7) to Eq. (8) is not valid. Since min_l \\tilde α_l ≤ \\tilde α_{σ_{n+1}}, the quantiles Q(min_l \\tilde α_l; \\hat F_{σ_i}) are no larger than Q(\\tilde α_{σ_{n+1}}; \\hat F_{σ_i}); therefore the left-hand side of Eq. (8) is bounded above, not below, by the left-hand side of Eq. (7). The assertion that Eq. (8) holds for the identity permutation, which is then used to obtain E[∑_i w_i 1_{V_i ≤ u^*_i}|T] ≥ α, is unsupported. This gap is load-bearing because Theorem A.7 underlies Theorem 3.2 and Corollary 3.3, the main coverage guarantees. Replacing 'min' by 'max' does not fix the issue, since the subsequent step requires Q(\\tilde α; ·) ≤ Q(min_l \\tilde α_l; ·). The authors need to either repair the symmetrization argument or explicitly restrict the theorem to a permutation-invariant choice of \\tilde α.","section":"Supplement B.3, proof of Theorem A.7, Eq. (7)–(8)"}],"minor_comments":[{"comment":"The abstract calls the guarantee 'assumption-free,' but the proof requires the localizer H to be a fixed function of the feature set only, not of the responses or of the data ordering. This is an explicit scope restriction in Section 3, but it should be stated prominently in the abstract and introduction so that readers do not apply the method to response-dependent localizers, for which Lemma B.2 would fail.","section":"Abstract and Section 3"},{"comment":"The simulation tables report coverage for 1000 repetitions, but no Monte Carlo standard errors are given. With 1000 repetitions the standard error is approximately 0.006–0.01, so differences of 0.01 between methods are not interpretable without error bars.","section":"Tables 1 and 2"},{"comment":"In the second bullet of the proof of Lemma B.3, the text writes 'we always have Q(α; \\hat F) ≥ Q(α, ∑_{j=1}^{n+1} p^H_{n+1,j} δ_{V_j})'; the level in both quantiles should be \\tilde α, not α, and the inequality direction should be checked carefully.","section":"Supplement B.3, Lemma B.3 proof"},{"comment":"The paper states that the general recipe for data-dependent score functions is too computationally expensive to carry out in practice. This limitation is important and should be noted in the abstract or introduction, not only in Section 5.2, so that readers do not over-interpret the data-dependent extension as a practical algorithm.","section":"Section 5.2"}],"recommendation":"major_revision","confidential_remarks":"The proof gap in Theorem A.7 is serious and touches the central coverage claim. I recommend major revision rather than rejection because the symmetric-case proof and the overall idea appear sound; the authors need to either fix the symmetrization step or adjust the theorem statement to a permutation-invariant \\tilde α. I would also encourage them to add the response-dependent-localizer caveat to the abstract, since the current 'assumption-free' wording is broader than the theorem actually covers."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The central claim holds up. This is the first conformal construction I know of that gives the test point a special role through a localizer while still returning finite-sample marginal coverage. The trick is to adjust the quantile level using auxiliary localizers centered at each training point; the G2 condition handles the two hardest test-score cases and restores exchangeability. The permutation proof in Lemma B.2 is clean, and the H=1 case correctly reduces to standard conformal prediction. That is a real contribution, not a minor tweak.\n\nWhat the paper does well beyond the main theorem: the covariate-shift extension is thoughtful, the local-coverage theorem in Section 5.1 is a useful bridge to the conditional-coverage literature, and the simulations show the method capturing heteroscedastic structure that ordinary conformal bands miss. The authors are also honest that the data-dependent score version is computationally impractical, so the practical recipe is fixed score functions (e.g., via sample splitting).\n\nThe soft spots are real but not load-bearing. First, the stress-test concern is correct: the finite-sample guarantee requires H to be a fixed function of the feature multiset only. The definition in Section 3 does say this, so it is an explicit scope restriction rather than an internal inconsistency, but the abstract's \"assumption-free\" wording makes it easy to miss. A reader could reasonably build H from responses or from the data ordering and think the coverage still applies. The authors should state the restriction prominently in the abstract and introduction. Second, the simulation tables report no Monte Carlo error bars; with 1000 repetitions the standard errors are small, but reporting them would make the coverage comparisons more credible. Third, no code is shipped, which slows reproducibility. These are minor issues, not flaws in the math.\n\nOverall: the paper deserves a serious referee. I would send it out with a request for a scope caveat and minor reporting improvements, but I would expect it to survive review. This is a paper for people working on distribution-free inference and conformal prediction; it will be cited for the localization construction.","headline":"Genuinely new localized conformal method with a sound finite-sample proof; the main caveat is that the guarantee only covers feature-based localizers, and the abstract overstates that scope.","tokens_in":21771,"tokens_out":1632,"would_cite":true,"duration_ms":18607,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62G15"],"pacs":[],"model":"deepseek-v4-flash","headline":"Localized conformal prediction keeps its coverage guarantee","keywords":["conformal prediction","localized conformal inference","finite-sample coverage","distribution-free prediction intervals","local coverage","weighted exchangeability","covariate shift"],"falsifier":"Search is over all small configurations: pick $n=3$ or $4$, discrete features with all permutations, a binary score $V\\in\\{0,1\\}$, and a localizer matrix $H$ satisfying $H_{ii}=1$ and depending only on the feature set, then check whether every $\\tilde{\\alpha}$ satisfying condition (G2) also has $P(V_{n+1}\\le Q(\\tilde{\\alpha};\\hat{F}))\\ge\\alpha$ exactly; the first counterexample would refute Theorem 3.2. A cheaper check is to reproduce Example 3.7's asymptotic setting with a box localizer and $\\tilde{\\alpha}=\\alpha$, where the reported under-coverage $\\alpha(1-\\alpha)/(2-\\alpha)$ verifies that naively setting $\\tilde{\\alpha}=\\alpha$ is not sufficient.","tokens_in":20804,"feed_emoji":"🎯","tokens_out":8421,"duration_ms":70982,"temperature":0.7,"pith_summary":"The paper introduces localized conformal prediction, a method that builds a prediction interval for a new response $Y_{n+1}$ by weighting only the training samples closest to the test feature $X_{n+1}$, rather than treating all samples equally. Its central claim is that despite deliberately breaking the exchangeability of the scores, one can choose the quantile level $\\tilde{\\alpha}$ through a simple condition (G2) so that the resulting interval continues to satisfy $P(Y_{n+1}\\in \\hat{C}(X_{n+1}))\\ge \\alpha$ for every distribution $P$ and every finite sample size. This matters because score heterogeneity across the feature space is a common failure mode of ordinary conformal bands, and localization gives a way to adapt to local structure without giving up distribution-free coverage. The guarantee extends to covariate shift settings when the density ratio is known.","feed_headline":"Localized conformal prediction keeps its coverage guarantee","feed_subtitle":"Weight training points near the test sample and still get distribution-free finite-sample coverage.","key_machinery":"The load-bearing object is the localizer $H(x_1,x_2,\\mathcal{X})\\in[0,1]$ with $H(x,x,\\mathcal{X})=1$, which determines normalized weights $p^H_{i,j}=H_{ij}/\\sum_k H_{ik}$ for the weighted empirical distributions $\\hat{F}_i$. The proof's engine is Lemma B.2, a permutation argument that conditions on the unordered data set and expresses the conditional coverage probability as a weighted average of indicators $\\mathbb{1}\\{v_i\\le v_i^*\\}$; condition (G2) then certifies that this average stays above $\\alpha$ for the two hardest possible test scores, $V_{n+1}=0$ and $V_{n+1}=\\bar{v}^*$. Algorithm 1 searches over candidate $\\tilde{\\alpha}$ values to satisfy this condition, then inverts the score inequality to form the interval.","core_discovery":"The discovery is that the exchangeability broken by a localizer can be restored by symmetrizing the quantile choice. For any fixed score function $V$ and any localizer $H$ satisfying $H(x,x)=1$ and depending only on the unordered feature set, define a weighted empirical distribution $\\hat{F}$ with weights proportional to $H(X_{n+1}, X_i)$. If $\\tilde{\\alpha}$ is chosen so that $\\bar{v}^*=Q(\\tilde{\\alpha};\\hat{F})=\\infty$ or condition (G2) holds, then Theorem 3.2 gives $P\\{V_{n+1}\\le Q(\\tilde{\\alpha};\\hat{F})\\}\\ge\\alpha$, so Algorithm 1's interval $\\hat{C}(X_{n+1})=\\{y:V(X_{n+1},y)\\le Q(\\tilde{\\alpha};\\hat{F})\\}$ has marginal coverage at least $\\alpha$ for i.i.d. data, for every $P$. A randomized version of the level choice makes the coverage exactly equal to $\\alpha$. With the constant localizer $H_{i,j}=1$, the construction reduces to standard conformal prediction.","pith_inferences":["Inference: The condition (G2) works as a coverage certificate for any score function, so localized conformal prediction can be layered on top of modern scores such as conformalized quantile regression, not just fixed scores, likely narrowing intervals further where score heterogeneity remains.","Inference: Because the localizer may depend on the feature set, a user could choose $H$ adaptively (e.g., data-adaptive bandwidth from the paper's Supplement C) to balance locality and interval width; testing whether such adaptive localizers preserve coverage under the same exchangeability condition would be a natural follow-up.","Inference: In outlier-detection applications mentioned only as future work, a distance-based localizer could shrink prediction bands around dense regions and flag points whose response is unlikely given their local neighborhood; this could be tested by measuring detection rates on contaminated regression data.","Inference: The symmetrization idea (considering localizers centered at each training point) suggests a general recipe for any weighted conformal scheme: any weighting scheme that is symmetric in the unordered data set can be turned into a coverage-guaranteed method by tuning the quantile level with the analogue of (G2)."],"forward_implications":["For any fixed score function and any localizer satisfying the feature-set condition, Algorithm 1 yields a distribution-free interval with coverage at least $\\alpha$ for every $P$.","With the identity localizer $H_{i,j}=1$ and $\\tilde{\\alpha}=\\alpha$, the construction reproduces ordinary conformal prediction exactly (Proposition 3.4).","A randomized choice of $\\tilde{\\alpha}$ between two grid levels makes the marginal coverage exactly $\\alpha$, removing the slack from the non-randomized rule (Corollary 3.5, Theorem 5.5).","Under covariate shift with known density ratio $w$, the weighted version (Theorem A.7 and Corollary A.8) retains the same finite-sample guarantee, and localization can prevent a few extreme-weight training samples from dominating the interval.","For localizers with shrinking bandwidth and mild smoothness conditions, the interval achieves asymptotic conditional coverage at each feature value (Theorem 5.1)."],"supporting_citations":[{"why":"Establishes the exchangeability-based conformal guarantee that this paper generalizes and that Proposition 3.4 recovers when $H_{i,j}=1$.","marker":"[1]"},{"why":"Supplies the distribution-free prediction band framework and the regularity conditions used in Theorem 5.1 for asymptotic conditional coverage.","marker":"[4]"},{"why":"Shows finite-sample conditional coverage is impossible without infinite intervals, motivating the local-coverage notion the paper targets.","marker":"[7]"},{"why":"Provides the weighted-exchangeability result under covariate shift that the localized construction extends and uses in its Appendix A.","marker":"[8]"}],"fun_headline_variants":["Localized conformal prediction: coverage without exchangeability","Breaking exchangeability, keeping coverage: localized conformal","Local conformal intervals still guarantee coverage","Symmetrized quantiles restore conformal coverage","Localized conformal prediction: finite-sample coverage, no exchangeability"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"Everything rests on the requirement that the localizer $H$ depend only on the feature values $\\{X_1,\\dots,X_{n+1}\\}$ as an unordered set (and similarly that any data-dependent score depend only on the unordered data), never on the response values or the order of the samples; if a user builds $H$ from $Y_i$ or from the data ordering, the exchangeability argument in Lemma B.2 breaks and the finite-sample coverage guarantee no longer follows.","fun_headline_variants_meta":{"raw":{"variants":["Localized conformal prediction: coverage without exchangeability","Breaking exchangeability, keeping coverage: localized conformal","Local conformal intervals still guarantee coverage","Symmetrized quantiles restore conformal coverage","Localized conformal prediction: finite-sample coverage, no exchangeability"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000298,"raw_usage":{"total_tokens":1690,"prompt_tokens":873,"completion_tokens":817,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":489,"completion_tokens_details":{"reasoning_tokens":741}},"tokens_in":489,"tokens_out":817,"duration_ms":6619,"temperature":1.0,"reasoning_tokens":741,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T11:36:31.087609+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Search is over all small configurations: pick $n=3$ or $4$, discrete features with all permutations, a binary score $V\\in\\{0,1\\}$, and a localizer matrix $H$ satisfying $H_{ii}=1$ and depending only on the feature set, then check whether every $\\tilde{\\alpha}$ satisfying condition (G2) also has $P(V_{n+1}\\le Q(\\tilde{\\alpha};\\hat{F}))\\ge\\alpha$ exactly; the first counterexample would refute Theorem 3.2. A cheaper check is to reproduce Example 3.7's asymptotic setting with a box localizer and $\\tilde{\\alpha}=\\alpha$, where the reported under-coverage $\\alpha(1-\\alpha)/(2-\\alpha)$ verifies that naively setting $\\tilde{\\alpha}=\\alpha$ is not sufficient.","supporting_citations":[{"cited_title":"Springer Science & Business Media, 2005","cited_arxiv_id":null,"evidence_quote":"Establishes the exchangeability-based conformal guarantee that this paper generalizes and that Proposition 3.4 recovers when $H_{i,j}=1$."},{"cited_title":"Distribution-free prediction bands for non-parametric regression","cited_arxiv_id":null,"evidence_quote":"Supplies the distribution-free prediction band framework and the regularity conditions used in Theorem 5.1 for asymptotic conditional coverage."}],"review_version":1}