IF-RewardBench: Benchmarking Judge Models for Instruction-Following Evaluation
This benchmark addresses the problem of unreliable judge models for LLM developers by providing a more comprehensive and accurate evaluation method for instruction-following capabilities, which is an incremental improvement over existing benchmarks.
This paper introduces IF-RewardBench, a new benchmark for evaluating judge models' ability to assess instruction-following in large language models. It uses a preference graph for listwise evaluation, revealing significant deficiencies in current judge models and showing a stronger positive correlation with downstream task performance than existing benchmarks.
Instruction-following is a foundational capability of large language models (LLMs), with its improvement hinging on scalable and accurate feedback from judge models. However, the reliability of current judge models in instruction-following remains underexplored due to several deficiencies of existing meta-evaluation benchmarks, such as their insufficient data coverage and oversimplified pairwise evaluation paradigms that misalign with model optimization scenarios. To this end, we propose IF-RewardBench, a comprehensive meta-evaluation benchmark for instruction-following that covers diverse instruction and constraint types. For each instruction, we construct a preference graph containing all pairwise preferences among multiple responses based on instruction-following quality. This design enables a listwise evaluation paradigm that assesses the capabilities of judge models to rank multiple responses, which is essential in guiding model alignment. Extensive experiments on IF-RewardBench reveal significant deficiencies in current judge models and demonstrate that our benchmark achieves a stronger positive correlation with downstream task performance compared to existing benchmarks. Our codes and data are available at https://github.com/thu-coai/IF-RewardBench.