Predicting Inference-Time Scaling Gains from Labeled Validation-Set Output Statistics

• URL: http://arxiv.org/abs/2606.02981v1
• Date: Tue, 02 Jun 2026 00:38:51 GMT
• Title: Predicting Inference-Time Scaling Gains from Labeled Validation-Set Output Statistics
• Abstract summary: Best-of-$N$ inference scaling improves accuracy by an amount that varies across models, but currently requires running the procedure end-to-end.<n>We fit ridge predictors on features computed from a single validation-set sampling pass, use bootstrap-Lasso as a stability analysis of the candidate feature set, and give a concentration analysis with an explicit linear-approximation residual.<n>A compact ridge predictor built from this core plus an entropy add-on reaches Spearman $= 0.90$ with actual best-of-$N$ gain under a reward-model verifier.
• Score: 0.42970700836450487
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Best-of-$N$ inference scaling (drawing $N$ candidate answers from a language model and returning the one a reward model ranks highest) improves accuracy by an amount that varies across models, but predicting that amount in advance currently requires running the procedure end-to-end. Prior work links cheap statistics of a model's sampled outputs and validation-set correctness (how often samples agree, how diverse they are, how confident the model is, and where correct samples appear) to model behavior, but does not isolate which of these form a stable, compact predictor of best-of-$N$ gain. We fit ridge predictors on features computed from a single labeled validation-set sampling pass, use bootstrap-Lasso as a stability analysis of the candidate feature set, and give a concentration analysis with an explicit linear-approximation residual. Across three base-model families, six post-training methods, and math and reasoning task domains, the stability analysis identifies a strict three-feature core spanning prompt-level agreement spread, label-assisted first-correct-sample position, and completion-length variance; a compact ridge predictor built from this core plus an entropy add-on reaches Spearman $ρ= 0.90$ with actual best-of-$N$ gain under a reward-model verifier. The intended use is labeled validation-set screening of candidate configurations before paying the full reward-model scoring cost.

Related papers

• Selective Test-Time Compute Scaling for Click-Through Rate Prediction via Uncertainty-Triggered Feature Path Exploration [9.542597285477683]
  Scaling test-time compute has proven highly effective for language models, yet this opportunity remains largely unexplored for industrial Click-Through Rate (CTR) prediction.<n>We propose UTTSI (Uncertainty-Triggered Test-Time Selective Inference), a training-free model-agnostic framework that scales inference depth proportionally to per-instance uncertainty.
  arXiv  Detail & Related papers  (2026-05-24T10:29:10Z)
• Test-Time Personalization: A Diagnostic Framework and Probabilistic Fix for Scaling Failures [26.927205237105156]
  We study scaling inference-time computation by sampling N candidates from a personalized policy model and selecting the best with a personalized reward model.<n>We prove that oracle selection yields expected utility growing logarithmically with the number of sampled candidates.<n>We propose a probabilistic personalized reward model whose learned variance effectively mitigates both failure modes.
  arXiv  Detail & Related papers  (2026-05-09T12:09:11Z)
• $V_{0.5}$: Generalist Value Model as a Prior for Sparse RL Rollouts [81.48669089692189]
  Generalist Value Models (such as $V_0.5$) achieve pre-trained value estimation by explicitly encoding model capabilities in-context.<n>We propose $V_0.5$, which adaptively fuses the baseline predicted by such value model with the empirical mean derived from sparse rollouts.<n>$V_0.5$ significantly outperforms GRPO and DAPO, achieving faster convergence and over some 10% performance improvement.
  arXiv  Detail & Related papers  (2026-03-11T14:57:41Z)
• DOTA: Distributional Test-Time Adaptation of Vision-Language Models [69.41389326333771]
  Vision-language foundation models can be unreliable when significant distribution gaps exist between training and test data.<n>We propose DOTA (DistributiOnal Test-time Adaptation), a simple yet effective method addressing this limitation.<n>This distribution-centric approach enables the model to continually learn and adapt to the deployment environment.
  arXiv  Detail & Related papers  (2024-09-28T15:03:28Z)
• Rejection via Learning Density Ratios [50.91522897152437]
  Classification with rejection emerges as a learning paradigm which allows models to abstain from making predictions.<n>We propose a different distributional perspective, where we seek to find an idealized data distribution which maximizes a pretrained model's performance.<n>Our framework is tested empirically over clean and noisy datasets.
  arXiv  Detail & Related papers  (2024-05-29T01:32:17Z)
• Conformalization of Sparse Generalized Linear Models [2.1485350418225244]
  Conformal prediction method estimates a confidence set for $y_n+1$ that is valid for any finite sample size. Although attractive, computing such a set is computationally infeasible in most regression problems. We show how our path-following algorithm accurately approximates conformal prediction sets.
  arXiv  Detail & Related papers  (2023-07-11T08:36:12Z)
• Conformal Nucleus Sampling [67.5232384936661]
  We assess whether a top-$p$ set is indeed aligned with its probabilistic meaning in various linguistic contexts. We find that OPT models are overconfident, and that calibration shows a moderate inverse scaling with model size.
  arXiv  Detail & Related papers  (2023-05-04T08:11:57Z)
• Robust Outlier Rejection for 3D Registration with Variational Bayes [70.98659381852787]
  We develop a novel variational non-local network-based outlier rejection framework for robust alignment. We propose a voting-based inlier searching strategy to cluster the high-quality hypothetical inliers for transformation estimation.
  arXiv  Detail & Related papers  (2023-04-04T03:48:56Z)
• Boost Test-Time Performance with Closed-Loop Inference [85.43516360332646]
  We propose to predict hard-classified test samples in a looped manner to boost the model performance. We first devise a filtering criterion to identify those hard-classified test samples that need additional inference loops. For each hard sample, we construct an additional auxiliary learning task based on its original top-$K$ predictions to calibrate the model.
  arXiv  Detail & Related papers  (2022-03-21T10:20:21Z)
• Conformal Prediction Under Feedback Covariate Shift for Biomolecular Design [56.86533144730384]
  We introduce a method to quantify predictive uncertainty in settings where the training and test data are statistically dependent.<n>As a motivating use case, we demonstrate with several real data sets how our method quantifies uncertainty for the predicted fitness of designed proteins.
  arXiv  Detail & Related papers  (2022-02-08T02:59:12Z)
• $\Delta$-UQ: Accurate Uncertainty Quantification via Anchor Marginalization [40.581619201120716]
  We present $Delta$UQ -- a novel, general-purpose uncertainty estimator using the concept of anchoring in predictive models. We find this uncertainty is deeply connected to improper sampling of the input data, and inherent noise, enabling us to estimate the total uncertainty in any system.
  arXiv  Detail & Related papers  (2021-10-05T17:44:31Z)
• Robust Validation: Confident Predictions Even When Distributions Shift [19.327409270934474]
  We describe procedures for robust predictive inference, where a model provides uncertainty estimates on its predictions rather than point predictions. We present a method that produces prediction sets (almost exactly) giving the right coverage level for any test distribution in an $f$-divergence ball around the training population. An essential component of our methodology is to estimate the amount of expected future data shift and build robustness to it.
  arXiv  Detail & Related papers  (2020-08-10T17:09:16Z)
• Meta-Learned Confidence for Few-shot Learning [60.6086305523402]
  A popular transductive inference technique for few-shot metric-based approaches, is to update the prototype of each class with the mean of the most confident query examples. We propose to meta-learn the confidence for each query sample, to assign optimal weights to unlabeled queries. We validate our few-shot learning model with meta-learned confidence on four benchmark datasets.
  arXiv  Detail & Related papers  (2020-02-27T10:22:17Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.