{"id":"rewardbench-2","title":"RewardBench 2: Advancing Reward Model Evaluation","authors":[{"given":"Saumya","family":"Malik"},{"given":"Valentina","family":"Pyatkin"},{"given":"Sander","family":"Land"},{"given":"Jacob","family":"Morrison"},{"given":"Noah A.","family":"Smith"},{"given":"Hannaneh","family":"Hajishirzi"},{"given":"Nathan","family":"Lambert"}],"first_public":"2025-06-02","date_note":"First arXiv version.","date_source":"https://arxiv.org/abs/2506.01937","year":2026,"venue":"ICLR 2026","paper_url":"https://arxiv.org/abs/2506.01937","arxiv":"2506.01937","code_url":"https://github.com/allenai/reward-bench","category":"evaluation","relationship":"uses","dataset_version":"Previously unreleased WildChat prompts supply approximately 70% of the 1,865-prompt benchmark.","summary":"Builds a reward-model benchmark around previously unseen human prompts from WildChat, combining quality and topic filtering with manual verification. It evaluates reward models across factuality, precise instruction following, math, safety, focus, and ties, and studies how benchmark scores relate to best-of-N sampling and RLHF.","finding":"WildChat supplies the human-prompt domains: factuality (475), precise instruction following (160), math (183), and focus (495). The safety and ties domains use other prompt sources.","evidence_url":"https://arxiv.org/html/2506.01937v2","evidence_note":"Section 3 and Table 2: WildChat prompts and benchmark composition. Sections 4–5: reward models and downstream use.","bibtex_key":"malik2025rewardbench2","record_url":"https://yuntiandeng.com/wildchat/papers/rewardbench-2/","relationship_label":"Uses WildChat","citation_title":"RewardBench 2: Advancing Reward Model Evaluation","citation_type":"article-journal","bibtex_type":"misc","citation_url":"https://arxiv.org/abs/2506.01937","citation_year":2025,"citation_note":"2025 arXiv preprint; published at ICLR 2026.","doi":"10.48550/arXiv.2506.01937"}