Research map / Large language models / Alignment and preferences
From Curiosity to Caution: Mitigating Reward Hacking for Best-of-$N$ with Pessimism
Zhuohao Yu, Steven Wu, Adam Block
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add pessimism (caution) to the reward model score | Best-of-N · GSM8K | final accuracy at N=512: 71.5 -> 81.1 | Improved |
| add pessimism (caution) to the reward model score | Best-of-N · GSM8K | peak accuracy: 79.3 -> 82.6 | Improved |
| add pessimism to the reward model score | Best-of-N · GSM8K | degradation: 7.7 -> 1.5 | Improved |
| use pessimism only instead of combining with RM | Best-of-N · MATH-500 | peak accuracy: 12.3 -> 13.6 | Improved |
| add pessimism to the reward model score | Best-of-N · MATH-500 | final accuracy at N=512: 8.5 -> 10.1 | Improved |
| use uncertainty penalty only instead of RM + pessimism | Best-of-N · BBH-Hard | final accuracy at N=512: 11.0 -> 22.1 | Improved |
| add pessimism to the reward model score | Best-of-N · BBH-Hard | degradation: 15.6 -> 7.5 | Improved |
| add reward-model score to pessimism-only selection | Best-of-N · MATH-500 | degradation: 1.7 -> 2.3 | Hurt |
| increase pessimism strength lambda from 0.0 to 0.8 (ablation) | Caution · GSM8K | final accuracy at N=512: 71.2 -> 81.0 | Improved |
| use pessimism strength 1.0 (caution only) instead of 0.8 (ablation) | Caution · GSM8K | final accuracy at N=512: 81.0 -> 79.8 | Hurt |
| distill reward model features instead of random network features (RND) (ablation) | Caution · GSM8K | final accuracy at N=512 (lambda 0.8): 74.9 -> 81.0 | Improved |
| use a full-capacity predictor instead of lightweight predictor (ablation) | Caution · GSM8K | final accuracy at N=512: 82.2 -> 80.2 | Hurt |
Closest papers by meaning
- Inference-Time Reward Hacking in Large Language Models (NeurIPS 2025 · similarity 0.91)
- Exploring Data Scaling Trends and Effects in Reinforcement Learning from Human Feedback (NeurIPS 2025 · similarity 0.89)
- InfoRM: Mitigating Reward Hacking in RLHF via Information-Theoretic Reward Modeling (NeurIPS 2024 · similarity 0.89)
- Tail Competition Explains Scaling in Best-of-N Sampling for Verifiable Problems (NeurIPS 2026 · similarity 0.89)
- Reward Hacking Benchmark: Measuring Exploits in LLM Agents with Tool Use (ICML 2026 · similarity 0.88)
- On the Limits of Test-Time Compute: Sequential Reward Filtering for Better Inference (ICML 2026 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
