RewardBench
by Allen Institute for AI (Ai2)
What is RewardBench? RewardBench is a rubric libraries & frameworks eval made by Allen Institute for AI (Ai2). A benchmark and toolkit for evaluating reward models, including DPO-style models and LLM judges. RewardBench 2 has 1,865 prompts across six domains (factuality, precise instruction following, math, safety, focus and ties), mostly unseen human prompts, scored best-of-4 so random guessing gets 25%. Its paper reports that models score about 20 points lower than on the original RewardBench.
Editor note
A benchmark and toolkit for evaluating reward models, including DPO-style models and LLM judges. RewardBench 2 has 1,865 prompts across six domains (factuality, precise instruction following, math, safety, focus and ties), mostly unseen human prompts, scored best-of-4 so random guessing gets 25%. Its paper reports that models score about 20 points lower than on the original RewardBench.
- What it tests
- Reward models and LLM-as-judge models picking the correct response.
- Original source
- github.com/allenai/reward-bench
- Last updated
- Repo last push 2026-10-14
- License
- Apache-2.0
Related rubric libraries & frameworks evals
- Inspect Evals — Many benchmarks under one runner: coding, agents, safety, knowledge.
- OpenAI Evals — Building and running your own evals; community registry.
- LLMJury rubric library — LLM-as-judge rubrics for support, RAG, summaries, code and tone.
- Prompt Assay judge templates — Starter LLM-as-judge prompts.
Discussion
No one has discussed RewardBench yet. Have you run it? Share what it caught, what it missed or how you set it up.