What it measures
Evaluates reward models and judges on how well optimizing against them predicts real human Chatbot Arena gains.
Reading the measurement
Separate satisfying instructions from factual accuracy, human preference and safe behavior. Inspect strict all-constraint success as well as average criterion credit. Judge choice and adversarial prompt selection can materially change the result.
Published evaluation
Reported mean correctness accuracy across MMLU-Pro, MATH, GPQA, MBPP-Plus and IFEval. This measures preference evaluation, not the judge model solving those benchmarks. Native fractions are converted to percentages.
Where this fits in an evaluation suite
Format compliance, preference, truthfulness, and safety require separate measurements. Keep desirable writing style from masking factual errors.
Start with a capability hypothesis, then choose a second test with a different failure mode. Keep the checkpoint, reasoning effort, agent tools, sampling policy, and scoring procedure fixed when measuring a training intervention. Record failures and missing outputs in the denominator.
A reproducible comparison
Reported mean correctness accuracy across MMLU-Pro, MATH, GPQA, MBPP-Plus and IFEval. This measures preference evaluation, not the judge model solving those benchmarks. Native fractions are converted to percentages.
Publish the exact run configuration, item coverage, score units, and available uncertainty. If a source supplies only an aggregate or a single run, preserve that limit. A new model name or a rounded score alone does not establish a capability improvement.
PRIMARY EVIDENCE & FURTHER READING
- How to Evaluate Reward Models for RLHF — published result table ↗
All 9 authors
Evan Frick, Tianle Li, Connor Chen, Wei-Lin Chiang, Anastasios N. Angelopoulos, Jiantao Jiao, Banghua Zhu, Joseph E. Gonzalez, Ion Stoica
Reviewed 2026-10-26. Explanation and interpretation by RLTasks. Benchmark ownership remains with its original authors.