From Curiosity to Caution: Mitigating Reward Hacking for Best-of-N with Pessimism

• URL: http://arxiv.org/abs/2604.04648v1
• Date: Mon, 06 Apr 2026 12:58:11 GMT
• Title: From Curiosity to Caution: Mitigating Reward Hacking for Best-of-N with Pessimism
• Abstract summary: We show that caution is a simple, computationally efficient approach that substantially mitigates reward hacking in BoN sampling.<n>We also provide a theoretical analysis in a simplified linear setting, which shows that caution provably improves over the standard BoN approach.
• Score: 30.96634743446629
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Inference-time compute scaling has emerged as a powerful paradigm for improving language model performance on a wide range of tasks, but the question of how best to use the additional compute remains open. A popular approach is BoN sampling, where N candidate responses are generated, scored according to a reward model, and the highest-scoring response is selected. While this approach can improve performance, it is vulnerable to reward hacking, where performance degrades as N increases due to the selection of responses that exploit imperfections in the reward model instead of genuinely improving generation quality. Prior attempts to mitigate reward hacking, via stronger reward models or heavy-handed distributional regularization, either fail to fully address over-optimization or are too conservative to exploit additional compute. In this work, we explore the principle of pessimism in RL, which uses lower confidence bounds on value estimates to avoid OOD actions with uncertain reward estimates. Our approach, termed as caution, can be seen as the reverse of curiosity: where curiosity rewards prediction error as a signal of novelty, caution penalizes prediction error as a signal of distributional uncertainty. Practically, caution trains an error model on typical responses and uses its prediction error to lower reward estimates for atypical ones. Our extensive empirical evaluation demonstrates that caution is a simple, computationally efficient approach that substantially mitigates reward hacking in BoN sampling. We also provide a theoretical analysis in a simplified linear setting, which shows that caution provably improves over the standard BoN approach. Together, our results not only establish caution as a practical solution to reward hacking, but also provide evidence that curiosity-based approaches can be a general OOD detection technique in LLM settings.

Related papers

• Beyond Penalization: Diffusion-based Out-of-Distribution Detection and Selective Regularization in Offline Reinforcement Learning [29.019573695533563]
  We propose DOSER (Diffusion-based OOD Detection and Selective Regularization), a novel framework that goes beyond uniform penalization.<n>We show that DOSER consistently attains superior performance to prior methods, especially on suboptimal datasets.
  arXiv  Detail & Related papers  (2026-05-06T01:21:53Z)
• Revisiting the (Sub)Optimality of Best-of-N for Inference-Time Alignment [10.749836992585363]
  Best-of-N (BoN) sampling is a widely used inference-time alignment method for language models.<n>Recent theoretical work has suggested that it is statistically suboptimal and vulnerable to reward hacking.<n>We show that properly tuned BoN is both computationally and statistically optimal in achieving high win-rate.
  arXiv  Detail & Related papers  (2026-03-05T22:54:19Z)
• Information-Theoretic Reward Modeling for Stable RLHF: Detecting and Mitigating Reward Hacking [78.69179041551014]
  We propose an information-theoretic reward modeling framework based on the Information Bottleneck principle.<n>We show that InfoRM filters out preference-irrelevant information to alleviate reward misgeneralization.<n>We also introduce IBL, a distribution-level regularization that penalizes such deviations, effectively expanding the optimization landscape.
  arXiv  Detail & Related papers  (2025-10-15T15:51:59Z)
• Intra-Trajectory Consistency for Reward Modeling [67.84522106537274]
  We develop an intra-trajectory consistency regularization to enforce that adjacent processes with higher next-token generation probability maintain more consistent rewards.<n>We show that the reward model trained with the proposed regularization induces better DPO-aligned policies and achieves better best-of-N (BON) inference-time verification results.
  arXiv  Detail & Related papers  (2025-06-10T12:59:14Z)
• Probabilistic Uncertain Reward Model [27.40414952747553]
  This paper proposes the Probabilistic Uncertain Reward Model (PURM) to learn the reward distributions that emerged from the preference data.<n>We show that PURM outperforms existing methods with more accurate reward and sound uncertainty estimations.
  arXiv  Detail & Related papers  (2025-03-28T14:39:52Z)
• Is Best-of-N the Best of Them? Coverage, Scaling, and Optimality in Inference-Time Alignment [54.787826863212146]
  Inference-time computation offers a powerful axis for scaling the performance of language models.<n>We analyze the performance of inference-time alignment algorithms in terms of (i) response quality, and (ii) compute.<n>We introduce $textttInferenceTimePessimism$, a new algorithm which mitigates reward hacking through deliberate use of inference-time compute.
  arXiv  Detail & Related papers  (2025-03-27T18:00:08Z)
• Mitigating Reward Over-Optimization in RLHF via Behavior-Supported Regularization [23.817251267022847]
  We propose the Behavior-Supported Policy Optimization (BSPO) method to mitigate the reward over-optimization issue.<n>BSPO reduces the generation of OOD responses during the reinforcement learning process.<n> Empirical results show that BSPO outperforms baselines in preventing reward over-optimization.
  arXiv  Detail & Related papers  (2025-03-23T16:20:59Z)
• Self-Evolutionary Large Language Models through Uncertainty-Enhanced Preference Optimization [9.618391485742968]
  Iterative preference optimization has recently become one of the de-facto training paradigms for large language models (LLMs) We present an uncertainty-enhanced textbfPreference textbfOptimization framework to make the LLM self-evolve with reliable feedback. Our framework substantially alleviates the noisy problem and improves the performance of iterative preference optimization.
  arXiv  Detail & Related papers  (2024-09-17T14:05:58Z)
• Jailbreaking as a Reward Misspecification Problem [80.52431374743998]
  We propose a novel perspective that attributes this vulnerability to reward misspecification during the alignment process.<n>We introduce a metric ReGap to quantify the extent of reward misspecification and demonstrate its effectiveness.<n>We present ReMiss, a system for automated red teaming that generates adversarial prompts in a reward-misspecified space.
  arXiv  Detail & Related papers  (2024-06-20T15:12:27Z)
• Model Reprogramming Outperforms Fine-tuning on Out-of-distribution Data in Text-Image Encoders [56.47577824219207]
  In this paper, we unveil the hidden costs associated with intrusive fine-tuning techniques. We introduce a new model reprogramming approach for fine-tuning, which we name Reprogrammer. Our empirical evidence reveals that Reprogrammer is less intrusive and yields superior downstream models.
  arXiv  Detail & Related papers  (2024-03-16T04:19:48Z)
• Improving Out-of-Distribution Detection via Epistemic Uncertainty Adversarial Training [29.4569172720654]
  We develop a simple adversarial training scheme that incorporates an attack of the uncertainty predicted by the dropout ensemble. We demonstrate this method improves OOD detection performance on standard data (i.e., not adversarially crafted), and improves the standardized partial AUC from near-random guessing performance to $geq 0.75$.
  arXiv  Detail & Related papers  (2022-09-05T14:32:19Z)
• Policy Gradient Bayesian Robust Optimization for Imitation Learning [49.881386773269746]
  We derive a novel policy gradient-style robust optimization approach, PG-BROIL, to balance expected performance and risk. Results suggest PG-BROIL can produce a family of behaviors ranging from risk-neutral to risk-averse.
  arXiv  Detail & Related papers  (2021-06-11T16:49:15Z)

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.