PyTorchKR Papers

최신 논문

총 171편
태그·기간 등 상세 필터

카테고리 탐색

arXiv:2611.13186v1cs.LG

Prediction-Powered LLM Ranking with Adaptively Collected Human Feedback

Ranking large language models (LLMs) through human evaluations is essential for downstream decision-making. Because human feedback is expensive, evaluation platforms often collect it adaptively, prioritizing informative model comparisons using previously observed feedback. To further improve precision with limited human feedback, evaluators can draw on abundant unrated responses and relatively inexpensive LLM judges; however, judge bias and adaptive dependence challenge standard i.i.d. inference frameworks. We propose Prediction-Powered Ranking (PP-Rank), a framework that combines LLM-judge predictions with human feedback to improve precision while providing valid inference under adaptive sampling. Data-driven prediction weights minimize the asymptotic covariance trace within the proposed class. In theory, we establish asymptotic normality for general evaluation targets and simultaneous ranking inference. A joint martingale analysis handles dependence between adaptive human feedback and auxiliary data. We further derive a rule for reducing exploration that prioritizes lower-cost comparisons while matching a conservative variance bound, and characterize allocation limits for regular $\sqrt{T}$-rate estimation of pairwise means. Semi-synthetic experiments on UltraFeedback-Extended show improvements over human-only baselines in score estimation, confidence interval width, and ranking certification. Synthetic experiments demonstrate robustness to judge quality and adaptive allocation, while showing that PP-Rank can substantially reduce evaluation costs without sacrificing precision.

arXiv:2611.09792v2cs.AI

Knowing When to Intervene: Causal Evaluation of AI Agents

Current evaluations of agentic AI systems primarily rely on benchmarks. However, the static nature of benchmarks makes them uninformative about the value of auditing interventions at deployment time. For reliable deployment, we often need to determine within an agent's trajectory whether or not to intervene, e.g., by verifying the agent's current output or steering it toward different behavior, using either human review or more costly models or time-consuming agentic supervision. The value of such interventions is fundamentally a causal inference question, asking whether the causal effect of such an intervention is beneficial for the final outcome. Estimating this causal effect is particularly challenging due to the dynamic nature of the problem: the impact of an intervention depends on the current context and environment, as well as the agent's future behavior. We develop the first methodologically grounded framework for evaluating the causal effects of randomized interventions within agent trajectories. By leveraging a relatively small dataset of agent trajectories collected under randomized interventions such as verification or screening, we can estimate the dynamic causal effect of a rich class of (contextual) intervention policies. The structural parameters that we estimate can also be used to construct safety leaderboards that capture the risk of various agentic actions, which can be used by practitioners for audit prioritization. Applying our methods to tau-bench Retail, ToolSandbox, and ImpossibleBench, we find that intervention effects vary substantially across agent actions and states. More broadly, our work suggests a rich research agenda at the intersection of causal inference and agentic evaluation, which we believe is essential for the reliable and safe deployment of AI agents in the real world.

arXiv:2608.08743v2stat.ML

A Distribution Mapping Approach to Counterfactually Fair Reinforcement Learning

Reinforcement learning (RL) seeks to optimize sequential decisions to maximize population-level benefits over time. However, when deployed in high-stakes settings such as healthcare, RL decisions might systematically restrict some subpopulation's access to valuable services in a manner contrary to the values and goals of stakeholders. Counterfactual fairness (CF) offers a promising framework to address this problem based on causal reasoning. This paper develops a data preprocessing algorithm that, when used in tandem with policy learning, enables CF in RL. Our algorithm relies on a novel quantile distribution mapping method for sequentially estimating the counterfactual states and rewards in the data preprocessing step, subsuming common additivity assumptions used for counterfactual prediction as a special case. We theoretically prove that the per-step level of counterfactual unfairness and infinite-horizon suboptimality gap can be bounded under mild regularity conditions. We also empirically test our algorithm in numerical experiments as well as in application to a real-world interventional digital health dataset.

arXiv:2607.19510v2cs.LG

Total Variation Distance Estimation in Autoregressive Models

Modern LLM deployments use a number of implementation choices and inference optimizations (e.g., batching, custom kernels, and quantization) on top of fixed weights, so two engines serving "the same model" can produce meaningfully different distributions. We study the problem of estimating the total variation (TV) distance between two length-$n$ autoregressive distributions to additive error $\varepsilon$, under three access models. (1) Under sample access, we use $\widetilde{O}(n^2 K/\varepsilon^2)$ queries, where $K$ is the maximum support of the next-token distribution. This improves upon the $\widetilde{O}(n^3 m/\varepsilon^5)$-query estimator of Meel et al. (2025), where $m \geq K$ is the total size of the token alphabet. (2) Under logit access, we use $O(n/\varepsilon^2)$ queries, and this is tight. (3) Under noisy logit access, we smoothly interpolate between the above two guarantees: if probability values are given to relative error $\sigma$, we use $\widetilde{O}((n+n^2\sigma^2)/\varepsilon^2)$ queries. We complement our theoretical results with an empirical evaluation of our algorithms, for example measuring the distance between SGLang and vLLM serving identical weights. Our experiments highlight the robustness and practicality of estimating the total variation distance, which remains estimable where the KL divergence is infinite. Our code is available at https://github.com/XunZhiyang/llm-tv-estimation.

arXiv:2605.08705v2math.ST

Minimax Optimal Estimation of Transport-Growth Pairs in Unbalanced Optimal Transport

Unbalanced optimal transport extends classical optimal transport to measures with different total masses, but statistical guarantees for estimation of the transport map and the growth map remain limited. We study unbalanced optimal transport with quadratic cost and Kullback-Leibler marginal penalties and argue that the natural target of estimation is a transport-growth pair. Consequently, we develop two estimators for the transport-growth pair under two setups: a nearest-neighbor estimator for transporting between general densities, and a kernel-based estimator for transporting between smooth densities. We prove that the kernel-based estimator achieves the minimax optimal rate up to a logarithmic factor by deriving a matching lower bound on the minimax risk. Our main technical contribution is a reduction of estimation to a stability bound on the unbalanced optimal transport value. This bound controls the transport map and growth map risks through the duality gap of the true dual potentials for the estimated measures. These results provide a statistical foundation for estimation of the transport-growth pair in unbalanced optimal transport.

arXiv:2605.06749v3stat.ME

A Statistical Framework for Algorithmic Collective Action with Multiple Collectives

As learning systems increasingly shape everyday decisions, Algorithmic Collective Action (ACA), i.e., users coordinating changes to shared data to steer model behavior, offers a complement to regulator-side policy and corporate model design. Real-world collective actions have traditionally been decentralized and fragmented into multiple collectives, despite sharing overarching objectives, with each collective differing in size, strategy, and actionable goals. However, most of the ACA literature focuses on single collective settings. To address this, we propose the first comprehensive statistical framework for ACA with multiple collectives acting on the same system. In particular, we focus on collective action in classification, studying how multiple collectives can influence a classifier's behavior. We provide quantitative statistical bounds on the success of the collectives, considering the role and the interplay of the collectives' sizes and the alignment of their goals. We make such bounds computable by each collective with only partial knowledge of other collectives' sizes and strategies. Finally, we numerically illustrate our framework on simulations inspired by interventions for climate adaptation in smart cities, demonstrating the usefulness of our bounds.

arXiv:2605.02327v3stat.ME

Denoising data using convex relaxations

We study the problem of denoising observations \(Y_i=X_i+Z_i\), where the latent variables \(X_i\) are sampled from a low-dimensional manifold in \(\mathbb{R}^n\) and the noise variables \(Z_i\) are isotropic Gaussian. We propose a convex-relaxation estimator that first reduces dimension by principal component analysis and then projects the observations onto the convex hull of the projected latent manifold. We construct a statistical oracle that estimates its supporting hyperplanes from empirical Gaussian tail probabilities of the noisy sample. Under a lower-mass condition on the latent distribution, we prove finite-sample guarantees for the oracle and derive error bounds for the resulting denoiser. The analysis combines risk bounds for least-squares projection under convex constraints (due to Chatterjee) with entropy bounds for the manifold. We also verify the assumptions of the framework for a Cryo-Electron Microscopy observation model by establishing suitable covering number and Lipschitz estimates for the associated group action and imaging operators.

arXiv:2403.03007v5stat.CO

Scalable and Calibrated Sampling for Bayesian Generalized Linear Mixed Model via Stochastic Gradient Markov Chain Monte Carlo

Generalized linear mixed models (GLMMs) are widely used for analyzing correlated data, particularly in large-scale biomedical and social science applications. Scalable Bayesian inference for GLMMs is challenging due to an intractable marginal likelihood and a high computational cost incurred by conventional Markov chain Monte Carlo (MCMC) methods. We develop a stochastic gradient MCMC (SGMCMC) algorithm tailored to GLMMs that enables accurate posterior inference in the large-sample regime. Our approach uses Fisher's identity to construct a (biased) Monte Carlo estimator of the gradient of the marginal log-likelihood, making SGMCMC feasible when direct gradient computation is impossible. We analyze the additional variability, introduced by both data subsampling and gradient approximation, to derive a post-hoc covariance correction that yields properly calibrated posterior uncertainty. We show through simulated studies that the proposed method provides accurate posterior means and variances in settings with a large number of groups, outperforming existing approaches, including control variate methods. We further demonstrate the method's practical utility in an analysis of electronic health records data, where accounting for variance inflation materially changes scientific conclusions.

arXiv:2611.12690v1stat.ME

Average versus expected causal effects under interference: Modeling assumptions and estimation bottlenecks

Causal inference under treatment interference requires assumptions to reduce the exponentially large number of potential outcomes. We develop hierarchical finite-population models for potential outcomes by separating these assumptions into three layers: interference neighborhoods, exposure mappings, and outcome models. Given an interference neighborhood and exposure mapping, we show that potential outcomes admit a nonparametric decomposition into direct, spillover, and treatment--spillover interaction components. This decomposition provides a common coordinate system for nonparametric, semiparametric, and parametric models of finite-population potential outcomes. For a given potential-outcome model, we define two classes of estimands. Average causal effects (ACEs), such as direct and total treatment effects, contrast potential outcomes under fixed treatment--exposure conditions. Expected causal effects (ECEs), such as expected direct and indirect effects, contrast expectations of potential outcomes under a specified allocation policy. For both classes, we bound the variance of the canonical Horvitz--Thompson estimator by three terms: a diagonal weight term, a term from overlapping interference neighborhoods, and a residual design-dependence term. Our upper and lower bounds show that under graph-unaware designs, nonparametric design-unbiased estimation of ACEs can require sample sizes growing exponentially in the interference degree, whereas polynomial growth can suffice for design-aligned ECEs. Under neighborhood interference, we construct graph-aware independent-set designs with variance bounds polynomial in the maximum degree for direct and total treatment effects on every graph, whereas graph cluster randomization can incur worst-case variance exponential in the degree.

arXiv:2611.12612v1stat.ME

Graph-Guided Bayesian Factor Analysis: A Fast Integrated EFA-CFA Approach for High-Dimensional Data

Factor analysis has been central to the analysis of high-dimensional biomedical data. However, existing exploratory and confirmatory factor analysis methods face several limitations: less interpretable factor findings, difficulty in identifying which variables can and cannot be represented by low-dimensional factors, challenges in estimating the covariance among factors, and maximum-likelihood fits that break down when variables outnumber subjects. We propose graph-guided Bayesian factor analysis (gg-BFA) to address these gaps. We first use a Bayesian model to explore and learn the latent graph structure, identifying which variables can and cannot be represented by low-dimensional factors and how many factors are needed. A dense-subgraph search proposes partitions of the variables into communities, each defining one factor, and a forward stepwise search compares their fractional variational posteriors and stops when the evidence lower bound first decreases. We show that this search recovers the true communities even when the number of variables grows much faster than the sample size. We then develop a confirmatory Bayesian model whose variational posterior of the covariance contracts at the parametric rate up to a logarithmic factor, yielding posterior inference for the factors and their covariance, as well as predictions of missing values. Its four members range from a community mask to several cross-loadings per variable. In five biomedical panels, the complete analysis took under ten minutes each on one GPU, and cross-loadings improved held-out covariance and the prediction of masked measurements. In TCGA Kidney, six of the first eight factors were enriched for known pathways, and 190 of the 210 factor correlations had 95% credible intervals excluding zero. Overall, gg-BFA is computationally fast and provides interpretable factor analysis findings and reliable inference results.

arXiv:2611.12004v1stat.ME

Conformal Prediction for Multi-Step Forecasting: Conditional Coverage and Sharp Score Calibration Rates

Uncertainty quantification in prediction is a central problem in statistics and machine learning. Conformal prediction provides a flexible approach to constructing prediction sets with valid marginal coverage under exchangeability. In this paper, we investigate conformal prediction for both one-step and multi-step forecasting from a single stationary time series, addressing temporal dependence and the additional challenges of joint prediction across multiple horizons. We propose a kernel-based framework, called Time series Residual Analysis for Calibration with Kernels (TRACK), for estimating conditional quantiles of residual scores and constructing prediction sets from rolling-window forecasters. The framework accommodates heterogeneous uncertainty across horizons and accounts for dependence induced by rolling estimation, while allowing predictor memory and forecast horizons to grow with the sample size. Our finite-sample bounds quantify how temporal dependence and overlapping score windows affect calibration and coverage. We further derive bounds for conditional quantile estimation and pointwise conditional coverage discrepancy. The guarantees extend to more general predictors that admit an accurate rolling-window approximation. We also derive matching minimax bounds over classes of known score maps, clarifying how overlapping score windows affect calibration accuracy. Experimental results from simulations and applications to wind energy generation and bike sharing demand show that our methods improve regional conditional coverage while maintaining competitive interval widths.

arXiv:2611.11923v1stat.ME

Characterizing Underrepresentation when Generalizing Causal Survival Estimates

Randomized trial findings are routinely generalized to broader target populations to inform decisions in medicine and public policy. When the trial sample is misaligned with the target population, generalized effect estimates can yield suboptimal decisions. We characterize the target subgroups that are underrepresented in the trial cohort in time-to-event settings. Here, underrepresentation arises in two distinct ways: baseline underrepresentation, when a subgroup's covariate profile is rare in the trial relative to the target, and time-varying underrepresentation, when differential censoring erodes a subgroup over follow-up. Existing diagnostics either conflate these mechanisms or apply only to the baseline case. We show that the asymptotic variance of the efficient generalizability estimator admits a decomposition in which only two terms can diverge as representation weakens: a sampling-weight term capturing baseline misalignment and a censoring-weight term capturing differential loss to follow-up. Underrepresented subgroups are therefore those with imprecise generalized estimates, and the larger of the two terms gives the responsible channel. We use this decomposition in TRACE (Temporal Rashomon Analysis of CEnsored data), which combines a self-normalized variance objective with a tree-based search to localize underrepresented subgroups, attribute each to baseline mismatch or censoring, and find the follow-up intervals where censoring drives the variance inflation. We solve the joint problem with a parametric dynamic program nested inside a Rashomon ensemble of near-optimal trees. In synthetic benchmarks, TRACE recovers the planted underrepresented regions. In a case study generalizing the National Job Training Partnership Act (JTPA) Study to U.S. JTPA-eligible adults, restricting to reliably generalizable subgroups narrows the 95% confidence interval for the refined estimand by 60%.

arXiv:2611.11726v1cs.LG

The Dark Matter of Clinical Data: An Identification Limit on Disentangling Physiology from Care Process in electronic health records

Representations learned from electronic health record (EHR) data encode patient physiology together with the clinical process that decides what is measured, when, and for whom. We formalize the data-generating process with latent physiology, ordering intent, laboratory values that exist only when measured, a learned representation, and clinical environments, and analyze two strategies for undoing this entanglement. First, no division of the information carried by laboratory "value" and "ordering" channels is identified from a single dataset: when ordering depends on the analyte, its distribution among untested patients is not recoverable, and a sentinel code for the undefined recorded value makes it carry all the entropy of the ordering indicator, fixing partial information decomposition (PID) for every redundancy measure satisfying the Williams-Beer axioms and invariant under injective re-encoding. Second, cross-site invariance of representation-conditional outcomes is a falsifiable implication of biological invariance taken jointly with exclusion and representation-sufficiency conditions, but not an identification strategy: we construct observationally equivalent models that disagree about biological invariance, whether the invariance test fails or passes. For binary physiology this happens exactly when case mix differs, the representation is lossy, and physiology affects the outcome. A worked example yields a 9.6-percentage-point cross-site gap under an unchanged biological mechanism. Tests performed for every patient remove the ordering gate, and measured tests can restore identification: biological invariance is identified when a test exactly summarizes outcome-relevant physiology, or, for binary physiology, when two tests with independent errors are available at sites with different case mix. Cross-site analyses should be reported as tests of a stated joint hypothesis.

arXiv:2611.11667v1stat.ME

Nonasymptotic causal structure learning in generalized linear DAGs with continuous optimization

We study the problem of learning directed acyclic graphs from observational data in nonlinear models and provide finite-sample, nonasymptotic statistical guarantees in these models. Although widely studied, statistical guarantees for gradient-based methods in nonlinear models are missing from the literature. Under a generalized linear model for each node, we analyze a global nonsmooth objective combining an exact acyclicity constraint with a quasi-minimax concave penalty, which can be optimized via standard gradient-based solvers. We establish finite-sample guarantees under this model, including the most widely used logistic and Poisson models. Experiments corroborate structure recovery accuracy and computational performance, including both standardized data and negative controls as baselines.

arXiv:2611.11483v1stat.ME

Causal Machine Learning for Multilevel Studies with Unmeasured Cluster-Level Covariates via Variational Autoencoders

Unmeasured confounders are a key obstacle to accurately estimating treatment effects in observational studies, and in multilevel data, they often arise at the cluster level. Flexible machine learning methods cannot directly resolve such unmeasured confounding. To address this, we propose the variational autoencoder-augmented R-learner (VAE-R), which leverages variational autoencoders (VAEs) to infer unmeasured cluster-level covariates from the within-cluster dependence of outcome and treatment residuals, and uses the resulting adjustments to estimate the conditional average treatment effect (CATE) and the average treatment effect (ATE). The VAE decoder is specified through an offset likelihood so that the latent variables capture only the cluster-level variation left unexplained by measured covariates. VAE-R accommodates unmeasured cluster-level confounders in observational studies and unmeasured cluster-level predictors of the outcome in multisite randomized trials. We also provide data-driven strategies for selecting the two hyperparameters of the VAE, the latent dimension size and the Kullback-Leibler weight. Our simulation study shows that VAE-R nearly eliminates the bias of the ordinary R-learner's CATE and ATE estimates in multilevel observational data and improves their precision in multisite randomized trials. Finally, we illustrate VAE-R with two datasets, one from a multilevel observational study and the other from a multisite randomized trial.

arXiv:2611.11460v1stat.ML

Identifiable alignment of unpaired representations

Aligning two models without paired data requires a distance that both share, yet existing methods use Euclidean or spherical geometry. This is either an arbitrary choice or imposed by their loss function. We show that, under a precise form of the manifold hypothesis, the pullback of the data metric gives such a distance for any two models of one data-generating process. This holds whether the models are encoders, decoders or both, and whatever their architecture, loss, prior or training data. The alignment map preserves this distance and the distribution of representations, and is identifiable if and only if the data distribution breaks every symmetry of the data manifold. We show that this symmetry breaking holds generically. Identifiability makes the alignment coupling the unique optimum of Gromov--Wasserstein and of its weighted combinations with a linear program. The linear program alone has the same optimum if almost every data point has a distinct distribution of distances. The coupling, and with it the alignment map, is recovered exactly from permuted samples, even without identifiability, and consistently from independent ones. Our method, Neural Gromov--Monge Alignment (NGMA), uses no pairs or labels and outperforms every unpaired and label-informed baseline. It relies on pullback distances to align image models with different priors and single-cell models trained on disjoint cells, where Euclidean distances fail.

arXiv:2611.11381v1stat.AP

Does Predictive Performance Translate to Causal Performance? Evaluating tabular foundation models as nuisance learners for causal effect estimation

Tabular foundation models based on prior-data fitted networks (PFNs) achieve strong predictive performance, but their value for causal inference is unclear. We evaluated whether PFNs, which predict through in-context learning, could match or outperform conventional nuisance learners in supporting accurate causal estimation and reliable uncertainty quantification (inference). Learners modeled propensity scores and outcome regressions within doubly robust estimators of average causal effects, with and without cross-fitting, across synthetic and semi-synthetic settings with varying confounding complexity and effect heterogeneity. Under simple parametric data-generating processes, the PFNs were competitive with correctly specified parametric regressions and ensembles containing them. In more complex settings, they outperformed all conventional learners evaluated, including richly specified Super Learner ensembles, for both accuracy and uncertainty quantification. As sample size increased, only the two PFN models (TabPFN and Mitra) consistently approached nominal confidence interval coverage, whereas under-coverage persisted for conventional learners. In some settings, PFNs achieved near-nominal coverage and well-calibrated standard errors even without cross-fitting. At the smallest sample size (n=200) and in the higher-dimensional semi-synthetic settings, PFNs and other flexible learners exhibited poor coverage without cross-fitting. For PFNs, this appeared to be driven by propensity score overfitting consistent with in-sample memorization of treatment labels. These findings highlight the promise of prior-data fitted networks as in-context nuisance learners for causal estimation, while underscoring the need for further evaluation in challenging settings.

arXiv:2611.10955v1stat.ME

Causal Outcome Learning with Noisy Proxies

Modern observational biomedical datasets contain increasingly extensive baseline and outcome measurements, yet detecting a clinically meaningful treatment effect remains difficult because investigators must identify an outcome that captures treatment response while adjusting for confounding distributed across numerous noisy baseline variables. We introduce SCOUT (Score-based Causal Outcome Learning Using Treatment signals), a framework that learns a clinically interpretable, treatment-responsive outcome while adjusting for confounding throughout the learning process. SCOUT constructs low-dimensional adjustment representations of high-dimensional baseline measurements to estimate the average treatment effect (ATE) and conditional average treatment effect (CATE) for the learned outcome. Importantly, the final adjustment space is at most two-dimensional, allowing treatment-effect heterogeneity to be estimated in this space rather than in the full high-dimensional covariate space. SCOUT also introduces a fourth-order statistic for testing treatment-effect heterogeneity that emphasizes unusually large deviations from the ATE. In synthetic and semi-synthetic experiments, the algorithm achieved competitive ATE estimation accuracy and substantially lower CATE estimation error. In an observational study of patients with bipolar disorder, SCOUT identified a clinically recognizable lithium-associated adverse-effect profile, detecting both a significant ATE (Holm-adjusted $p$ = 0.0026) and significant treatment-effect heterogeneity (Holm-adjusted $p$ = 0.033).These results demonstrate how jointly learning outcomes and controlling for confounding can facilitate the discovery of treatment effects in high-dimensional observational data.

arXiv:2611.10787v1stat.ML

Trustworthy Recursive Self-Improvement with Statistical Guarantees

Recursive self-improvement offers a promising path toward AI agents that improve through experience by refining their own harnesses without retraining their underlying large language models (LLMs). Making this self-improvement process reliable, however, requires more than generating new updates: agents must decide which changes are worth keeping. A greedy gate that accepts updates based solely on higher validation scores can mistake an observed gain for genuine progress and lead to performance degeneration, whereas overly conservative statistical gates may reject beneficial updates and hinder performance improvement. We introduce Trustworthy Recursive Updates with Statistical Testing (TRUST), a framework that adapts an online multiple hypothesis testing procedure to control the rate of false evolution across rounds, with an LLM analyzing historical context to help determine how much evidence is required to update the incumbent. Under interpretable assumptions, we prove that TRUST limits the proportion of accepted updates that result in unacceptable performance loss. We further derive a lower bound on the final performance gain. Across the benchmarks HotpotQA, DROP, and APPS with the models DeepSeek-V4-Flash and Qwen-3.8-Flash, TRUST reduces the average ratio of accepted inferior updates from the greedy gate's 45.59% to 2.78%, while achieving higher final scores than the state-of-the-art statistical gate baselines for recursive self-improvement. These findings demonstrate that TRUST strikes a balance between statistical reliability and performance gain in recursive self-improvement.

arXiv:2611.10738v1stat.ML

Matchability-Aware Conformal Prediction for Open-Ended Language Model Generation

Large language models (LLMs) often do not produce calibrated uncertainty estimates for their answers. We propose CAELUM, a framework that builds semantic prediction sets while separating aleatoric from epistemic uncertainty. CAELUM samples answers using an ensemble of clarification prompts, clusters them by semantic equivalence, and estimates the aleatoric and epistemic components via kernel-based von Neumann entropy before calibrating nonconformity scores with split conformal prediction. We use a weight $w_\lambda(x)=1+U_{\mathrm{ale}}(x)+\lambda\,U_{\mathrm{epi}}(x)$ to control for the contribution of the epistemic uncertainty of each instance and fit a coefficient $\lambda^\star$ to minimize prediction set size. When the objective is nonflat and selection is stable across splits, $\lambda^\star$ could be a diagnostic, where large values may indicate epistemic dominance (retrieve, abstain) and $\lambda^\star\to 0$ aleatoric dominance (clarify). For individual queries, a log ratio $R(x)$ compares the weighted epistemic component with the aleatoric component. Theoretically, we demonstrate that the coverage is capped by matchability, the fraction of items whose reference answer appears in any sampled cluster. Empirically, this matchability can be predicted from the original question features using a lightweight classifier. We use those predictions to define a stratum and derive a conformal coverage bound within it. Across 30 LLM and dataset combinations (4-32B), mean coverage is 0.74 in the learned stratum versus 0.65 overall, averaged over the 78 evaluations using the learned classifier. Project page: https://unco3892.github.io/caelum/