Research map / Large language models / Language, safety and interpretability
Small Language Model Can Self-Correct
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.
Proposes Intrinsic Self-Correction (ISC), fine-tuning small (6-13B) open LMs on constructed self-verify-then-revise data with Partial Answer Masking so they self-correct in a single generation pass.
Problem
Small generative language models produce confident but incorrect answers and cannot self-correct via multi-step prompt pipelines designed for large models like ChatGPT.
Approach
Instead of prompting separate verify and modify steps, train the model via instruction fine-tuning to self-verify and self-revise in one spontaneous pass, using Partial Answer Masking that excludes incorrect initial answers from the loss so the model learns verification and correct revisions rather than errors.
Key results
- ChatGLM-6B accuracy on OpenBookQA rises from 37.0% to 42.6% after intrinsic self-correction, the largest gain reported (+5.6 points).
- Partial Answer Masking improves post-correction accuracy for most models; e.g., CuteGPT-13B on OpenBookQA reaches 42.0 with PAM versus 32.4 without.
- Gains are small or absent for some models: Vicuna-7B improves only +0.4 on OpenBookQA and Llama2-7B shows +0.0.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply ISC instead of using the initial answer | ChatGLM-6B · OpenBookQA | accuracy: 37.0 -> 42.6 (+5.6) | Helped |
| Apply ISC instead of using the initial answer | CuteGPT-7B · OpenBookQA | accuracy: 25.2 -> 29.0 (+3.8) | Helped |
| Apply ISC instead of using the initial answer | CuteGPT-13B · OpenBookQA | accuracy: 37.2 -> 42.0 (+4.8) | Helped |
| Apply ISC instead of using the initial answer | Llama2-7B · OpenBookQA | accuracy: 52.2 -> 52.2 (+0.0) | No clear effect |
| Apply ISC instead of using the initial answer | Vicuna-7B · OpenBookQA | accuracy: 28.6 -> 28.80 (+0.4) | Helped |
| Apply ISC instead of using the initial answer | Vicuna-13B · OpenBookQA | accuracy: 33.8 -> 34.0 (+0.2) | Helped |
| Apply ISC instead of using the initial answer | CuteGPT-7B · CommonsenseQA | accuracy: 23.2 -> 28.9 (+3.7) | Helped |
| Apply ISC instead of using the initial answer | CuteGPT-13B · CommonsenseQA | accuracy: 35.6 -> 37.9 (+2.3) | Helped |
| Apply ISC instead of using the initial answer | Llama2-7B · CommonsenseQA | accuracy: 52.2 -> 52.3 (+0.1) | Helped |
| Apply ISC instead of using the initial answer | ChatGLM-6B · CommonsenseQA | accuracy: 34.3 -> 38.7 (+4.4) | Helped |
| Apply ISC instead of using the initial answer | Vicuna-7B · CommonsenseQA | accuracy: 25.9 -> 26.2 (+0.3) | Helped |
| Apply ISC instead of using the initial answer | Vicuna-13B · CommonsenseQA | accuracy: 32.4 -> 32.6 (+0.3) | Helped |
Surprising findings
- Models with high confidence in initial answers (e.g., Vicuna family, ~98% confidence) rarely self-correct, limiting gains.
- Without PAM, CuteGPT models tend to judge their answers as incorrect most of the time, reducing self-verification accuracy.
Limitations
- Gains are modest and inconsistent across models; some models (Llama2-7B, Vicuna) barely improve.
- Evaluation limited to multiple-choice QA on two datasets plus a zero-shot StrategyQA check.
- Self-verification is weak; models often remain confident in wrong answers and multi-round correction is not fully explored.
Experimental setup
Datasets: OpenBookQA CommonsenseQA
Benchmarks: OpenBookQA CommonsenseQA StrategyQA
Metrics: ACC-First ACC Confidence EvalACC R2R R2W W2R W2W
Baselines: Llama2-7B ChatGLM-6B Vicuna-7B Vicuna-13B CuteGPT-7B CuteGPT-13B
Builds on: chain-of-thought prompting LoRA nucleus sampling Self-Refine
Closest papers by meaning
- Large Language Models Cannot Self-Correct Reasoning Yet (ICLR 2024 · similarity 0.92)
- Can LLMs Correct Themselves? A Benchmark of Self-Correction in LLMs (NeurIPS 2025 · similarity 0.91)
- A Theoretical Understanding of Self-Correction through In-context Alignment (NeurIPS 2024 · similarity 0.89)
- S^3cMath: Spontaneous Step-Level Self-Correction Makes Large Language Models Better Mathematical Reasoners (AAAI 2025 · similarity 0.88)
- Training Language Models to Self-Correct via Reinforcement Learning (ICLR 2025 · similarity 0.88)
- Incentivizing LLMs to Self-Verify Their Answers (NeurIPS 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.