On October 24, 2026, an AI research paper titled “What Pairwise Evaluation Misses: Reward Scores across Prompts” was published on the arXiv platform (arXiv:2610.31090v1), focusing on the central role of reward models in preference optimization, response filtering, and ranking. The study points out that the widely used pairwise evaluation method compares only two responses under the same prompt, which automatically cancels out the prompt‑related scoring components shared by all responses, rendering it entirely incapable of detecting those components.
To quantify this overlooked bias, the authors propose a “human-referenced audit,” which combines inter-prompt variance analysis, sibling-rating regression, and prompt‑replacement experiments, and applies it to 18 publicly released reward models and six corresponding training conditions. The results show that, on the HelpSteer2 benchmark, the inter-prompt variance of six reward models ranges from 0.34 to 0.64, whereas the inter-prompt variance for human ratings is close to zero, indicating that model scores exhibit significant “clustering by prompt” rather than differentiation based on response quality.
Further experiments revealed that when the original prompt is replaced with a fixed instruction, the model scores still exhibit clear clustering, but their agreement with human ratings at the prompt level drops significantly, indicating that clustering alone does not equate to alignment with human judgment. Meanwhile, subtracting the prompt mean reduces the utility of the globally preferred response, and even after shuffling the prompt means during training, the model scores remain sensitive to incorrect prompts—suggesting that prompt-related variation is neither pure noise nor entirely reliable signal. The paper ultimately recommends that reward-model evaluation should separately assess “inter-prompt variation,” “sensitivity to incorrect prompts,” and “practicality for response selection.”