Research map / Large language models / Agents, code and math
CriticEval: Evaluating Large-scale Language Model as Critic
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces CriticEval, a human-annotated benchmark evaluating LLM critique ability across four dimensions (feedback, comparison, correction, meta-feedback) on nine tasks, using GPT-4 with human reference critiques as a judge.
Problem
Existing benchmarks for LLM critique ability are narrow in critique dimensions, response qualities and tasks, and GPT-4-based evaluation of textual critiques is unverified across dimensions and tasks.
Approach
Evaluate critiques in four dimensions over nine tasks with human-verified response quality levels, and make GPT-4 reliable for textual critique scoring by anchoring it with human-annotated reference critiques (GPT-4-turbo drafts refined by human experts).
Key results
- GPT-4-turbo achieves the best overall score (72.55 objective, 7.81 subjective), with Spearman meta-feedback correlation of 62.90 on test and 66.18 on dev, versus 79.03 for average humans.
- Removing human reference critiques lowers GPT-4-turbo meta-feedback correlation from 66.18 to 47.26 (-18.92), averaging -13.36 across LLMs.
- InternLM2-20B-Chat (subjective overall 6.20) surpasses GPT-3.5-turbo (5.89), and critique-tuned Llama-2-13B models (Auto-J-13B) outperform Llama-2-70B-Chat on feedback.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Provide human-annotated reference critiques to the evaluator (ablation) | GPT-4-turbo · CriticEval dev meta-feedback | correlation with human judgments: 47.26 -> 66.18 | Helped |
| Provide human-annotated reference critiques to the evaluator (ablation) | Qwen-1.5-72B · CriticEval dev meta-feedback | correlation with human judgments: 22.35 -> 38.97 | Helped |
| Provide human-annotated reference critiques to the evaluator (ablation) | Claude-instant-1 · CriticEval dev meta-feedback | correlation with human judgments: 19.88 -> 36.88 | Helped |
| Provide human-annotated reference critiques to the evaluator (ablation) | GPT-3.5-turbo · CriticEval dev meta-feedback | correlation with human judgments: 16.38 -> 17.28 | Helped |
| Use human-annotated instead of self-generated feedback (ablation) | LLM correction · CriticEval math and coding tasks | objective correction score: 13.01 -> 43.66 | Helped |
| Use self-generated feedback instead of no feedback (ablation) | LLM correction · CriticEval subjective tasks | subjective correction score: 5.16 -> 5.48 | Helped |
| Use human-annotated instead of self-generated feedback (ablation) | LLM correction · CriticEval subjective tasks | subjective correction score: 5.48 -> 7.12 | Helped |
| Increase response quality from low to high | LLM feedback critique · CriticEval test | subjective feedback score: 5.14 -> 4.66 | Hurt |
| Increase response quality from low to high | LLM meta-feedback · CriticEval test | meta-feedback objective correlation: 22.73 -> 28.84 | Helped |
| Use hard rather than easy response pairs | LLM comparison critique · CriticEval test | objective comparison score: 39.73 -> 29.80 | Hurt |
| Scale model parameters upward | LLMs evaluated on CriticEval · CriticEval | critique ability | Helped |
| Fine-tune Llama-2-13B on critique datasets | Llama-2-13B · CriticEval feedback | objective feedback correlation: 30.61 -> 36.05 | Helped |
Surprising findings
- Self-generated feedback can hurt correction on math and coding tasks (SF 13.01 < EF 14.44 < HF 43.66 objective).
- Higher-quality responses are harder to critique for feedback because they contain more subtle errors, with medium-quality responses scoring above low-quality ones on objective feedback.
Limitations
- Reference critiques are human-revised GPT-4 drafts, which may bias annotators and leave them sub-optimal for open-ended tasks.
- GPT-4 subjective judgments do not perfectly match human judgment; a gap to average human correlation remains.
- Only greedy decoding is used, and tasks like tool use, knowledge-intensive and hallucination are not covered.
Experimental setup
Datasets: WMT20 MLQE ChatArena AlpacaEval OBQA CommonsenseQA PIQA Anthropic HHH Summarization from human feedback AQuA-RAT MathQA GSM8K NumGLUE
Benchmarks: CriticEval
Metrics: Spearman correlation pairwise comparison accuracy pass rate subjective GPT-4 score (1-10)
Baselines: GPT-4-turbo GPT-3.5-turbo Claude-instant-1 Auto-J-13B UltraCM-13B TigerScore-13B UltraRM-13B Ziya-7B
Builds on: Auto-J UltraFeedback MT-Bench CriticBench
Closest papers by meaning
- CRITIC: Large Language Models Can Self-Correct with Tool-Interactive Critiquing (ICLR 2024 · similarity 0.91)
- JudgeLM: Fine-tuned Large Language Models are Scalable Judges (ICLR 2025 · similarity 0.89)
- RevisEval: Improving LLM-as-a-Judge via Response-Adapted References (ICLR 2025 · similarity 0.89)
- SciEval: A Multi-Level Large Language Model Evaluation Benchmark for Scientific Research (AAAI 2024 · similarity 0.89)
- Neither Valid nor Reliable? Investigating the Use of LLMs as Judges (NeurIPS 2025 · similarity 0.88)
- Benchmarking LLMs' Judgments with No Gold Standard (ICLR 2025 · similarity 0.88)
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.