Research map / Large language models / Language, safety and interpretability
FairJudge : An Adaptive, Debiased, and Consistent LLM-as-a-Judge
Bo Yang, Lanfei Feng, Yunkui Chen, Xiao Xu, Yu Zhang, Shijian Li
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add SFT stage (rubric-aware supervised fine-tuning) (ablation) | FairJudge-8B · PandaLM Test Set | macro-F1: 68.74 -> 72.18 | Improved |
| add SFT stage (ablation) | FairJudge-8B · FairJudge-Benchmark-1K | macro-F1: 57.86 -> 67.63 | Improved |
| add SFT stage (ablation) | FairJudge-8B · JudgeLM Test Set | macro-F1: 63.86 -> 66.78 | Improved |
| add DPO debiasing stage (ablation) | FairJudge-8B · PandaLM Test Set | macro-F1: 70.05 -> 72.18 | Improved |
| add DPO debiasing stage (ablation) | FairJudge-8B · JudgeLM Test Set | Recall: 67.17 -> 66.93 | Hurt |
| add GRPO consistency stage (ablation) | FairJudge-8B · FairJudge-Benchmark-1K | Precision: 72.32 -> 71.15 | Hurt |
| add GRPO consistency stage (ablation) | FairJudge-8B · JudgeLM Test Set | macro-F1: 63.95 -> 66.78 | Improved |
| full SFT-DPO-GRPO training of Qwen3-VL-8B (ablation) | Qwen3-VL-8B · PandaLM Test Set | macro-F1: 54.05 -> 72.18 | Improved |
| add cross-mode consistency training (GRPO) and policy framing | FairJudge-8B · Pointwise-pairwise consistency, FairJudge-Benchmark-1K | consistency (%): 60.59 -> 65.52 | Improved |
| scale general-purpose model to 72B or 671B instead of judge training | general-purpose LLM judges · PandaLM, JudgeLM, FairJudge-Benchmark-1K | agreement/F1 | Mixed |
| increase FairJudge size from 2B to 4B to 8B | FairJudge · PandaLM Test Set | macro-F1: 61.68 -> 62.18 -> 72.18 | Mixed |
| use Fast mode (final decision only) instead of Full reasoning | FairJudge-8B · 1,000 samples on RTX 4090 | wall-clock time: 16 min 52 s -> 1 min 19 s | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- JudgeLM-7B — JudgeLM: Fine-tuned Large Language Models are Scalable Judges (ICLR 2025) · better on 5 of 5 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- JudgeBench: A Benchmark for Evaluating LLM-Based Judges (ICLR 2025 · similarity 0.91)
- Evaluating the Evaluator: Measuring LLMs’ Adherence to Task Evaluation Instructions (AAAI 2025 · similarity 0.91)
- Justice or Prejudice? Quantifying Biases in LLM-as-a-Judge (ICLR 2025 · similarity 0.90)
- Beyond the Surface: Enhancing LLM-as-a-Judge Alignment with Human via Internal Representations (NeurIPS 2025 · similarity 0.89)
- Validating LLM-as-a-Judge Systems under Rating Indeterminacy (NeurIPS 2025 · similarity 0.89)
- J1: Incentivizing Thinking in LLM-as-a-Judge via Reinforcement Learning (ICLR 2026 · similarity 0.89)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
