Research map / Large language models / Reasoning and chain of thought
Can LLMs Correct Themselves? A Benchmark of Self-Correction in LLMs
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 CorrectBench, a benchmark evaluating intrinsic, external, and fine-tuned self-correction methods, their mixtures, and reasoning LLMs across commonsense, math, and code tasks.
Problem
Whether LLMs can truly correct themselves is unclear: self-correction gains are inconsistent across tasks, and more intricate correction schemes may not outperform simpler methods or justify their cost.
Approach
A systematic benchmark that categorizes 11 self-correction methods into intrinsic, external, and fine-tuned types, evaluates them across 7 subdatasets and 9+ LLMs, and studies mixtures of methods and reasoning LLMs with accuracy and resource-cost analysis.
Key results
- Self-correction methods improve over Base on average; CoVe gains +18.85 points on GPQA and Self-Refine +22.13 on GPQA.
- Reflexion-v1 without tools declines substantially (e.g., -18.82 on GSM8K), while Reflexion-v2 with tools improves on HotpotQA (+7.22) and AQUA (+7.00).
- Simple CoT is competitive, with 91.96 on GSM8K (+5.50) and far lower response time than correction methods; DeepSeek-V3 baseline is close to its corrected variants.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use CoVe instead of Base | LLMs averaged across models · GPQA | accuracy: 18.56 -> 37.41 (+18.85) | Helped |
| Use Self-Refine instead of Base | LLMs averaged across models · GPQA | accuracy: 18.56 -> 40.69 (+22.13) | Helped |
| Use Reflexion-v1 without external tools instead of Base | LLMs averaged across models · GSM8K | solve rate: 86.46 -> 67.64 (-18.82) | Hurt |
| Add external tools to Reflexion (Reflexion-v1 to Reflexion-v2) | Reflexion · HotpotQA | accuracy: 69.52 -> 87.98 | Helped |
| Use CoT instead of Base | LLMs averaged across models · GSM8K | solve rate: 86.46 -> 91.96 (+5.50) | Helped |
| Use CoT instead of Base | LLMs averaged across models · HumanEval | pass@k: 72.71 -> 60.10 (-12.61) | Hurt |
| Use RARR instead of Base | LLMs averaged across models · MATH | solve rate: 75.12 -> 82.78 (+7.66) | Helped |
| Use CoVe instead of Base | LLMs averaged across models · HotpotQA | accuracy: 80.76 -> 83.04 (+2.28) | Helped |
| Use DCoT fine-tuning | LLaMA2-7B-hf · GSM8K | solve rate: 56.75 -> 41.20(-15.55) | Hurt |
| Fine-tune with SCORE | Gemma-7B-it · GSM8K | solve rate: 56.75 -> 75.30(+18.55) | Helped |
| Apply domain-specific SuperCorrect fine-tuning | Qwen2.5-Math-7B-Instruct · MATH | solve rate: 41.71 -> 70.16(+28.45) | Helped |
| Use RATT instead of Base | LLMs evaluated on MATH resource-cost set · MATH (150 samples) | accuracy (%): 68.5 -> 78.7 | Helped |
Surprising findings
- Reasoning LLM DeepSeek-V3 shows little benefit from additional self-correction methods, attributed to built-in correction.
- Simple CoT baseline matches or beats many complex correction methods in accuracy-efficiency trade-off.
Limitations
- No error bars or statistical significance reported in main results; bootstrap CIs only in appendix on a small subsample.
- Only ~100 samples per dataset after filtering, limiting statistical power.
- Inconsistent reporting of method coverage and inconsistent numbers between text and tables.
Experimental setup
Datasets: HotpotQA CommonsenseQA GPQA GSM8K AQUA MATH HumanEval
Benchmarks: HotpotQA CommonsenseQA GPQA GSM8K AQUA MATH HumanEval
Metrics: accuracy solve rate pass@k correction rate misjudgment rate response time tokens API calls
Baselines: CoT RCI CoVe Self-Refine Reflexion-v1 Reflexion-v2 RARR RATT CRITIC
Builds on: Chain-of-Thought Self-Refine Reflexion CRITIC CoVe
Closest papers by meaning
- Large Language Models Cannot Self-Correct Reasoning Yet (ICLR 2024 · similarity 0.94)
- S^3cMath: Spontaneous Step-Level Self-Correction Makes Large Language Models Better Mathematical Reasoners (AAAI 2025 · similarity 0.93)
- Toward Self-Improvement of LLMs via Imagination, Searching, and Criticizing (NeurIPS 2024 · similarity 0.91)
- Small Language Model Can Self-Correct (AAAI 2024 · similarity 0.91)
- Incentivizing LLMs to Self-Verify Their Answers (NeurIPS 2025 · similarity 0.91)
- ReGenesis: LLMs can Grow into Reasoning Generalists via Self-Improvement (ICLR 2025 · similarity 0.90)
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.