Research map / Large language models / Reasoning and chain of thought
SelfCheck: Using LLMs to Zero-Shot Check Their Own Step-by-Step Reasoning
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 SelfCheck, a zero-shot step-by-step self-verification method that uses target extraction, information collection, step regeneration and comparison to score LLM reasoning chains, then uses the scores for weighted voting on math problems.
Problem
LLMs make errors in multi-step reasoning, and existing step-checking methods need extra training data, exemplars, or external resources; direct self-checking by LLMs is ineffective.
Approach
Decompose step checking into four separate LLM calls (target extraction, information collection, step regeneration from only the collected context, and result comparison) so the regenerated step is decorrelated from the original; aggregate step results into a confidence score used for weighted voting.
Key results
- On GSM8K, MathQA and MATH (GPT-3.5, 10 solutions), SelfCheck improves over majority voting by 2.8%, 5.4% and 2.2% respectively in the 2-solution setting table (GPT-3.5 rows).
- With GPT-4 generator and GPT-3.5 checker, SelfCheck reaches 88.1% on GSM8K vs 87.1% majority voting and 81.2% vs 78.3% on MathQA.
- Global single-prompt checking achieves only 55.0% verification accuracy on MathQA, close to random, versus 66.7% for SelfCheck.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use SelfCheck confidence scores for weighted voting | GPT-3.5 generator/checker · GSM8K | answer accuracy: 71.7 -> 74.3; 2.8±0.9 | Helped |
| Use SelfCheck confidence scores for weighted voting | GPT-4 generator/checker · GSM8K | answer accuracy: 87.1 -> 86.9; -0.2±0.2 | No clear effect |
| Use GPT-3.5 to check GPT-4 solutions and weight votes | GPT-4 generator · GSM8K | answer accuracy: 87.1 -> 88.1; 1.0±0.3 | Helped |
| Use SelfCheck confidence scores for weighted voting | GPT-3.5 generator/checker · MathQA | answer accuracy: 59.2 -> 64.6; 5.4±1.1 | Helped |
| Use SelfCheck confidence scores for weighted voting | GPT-4 generator/checker · MathQA | answer accuracy: 78.3 -> 80.9; 2.6±0.4 | Helped |
| Use GPT-3.5 to check GPT-4 solutions and weight votes | GPT-4 generator · MathQA | answer accuracy: 78.3 -> 81.2; 3.0±0.4 | Helped |
| Use SelfCheck confidence scores for weighted voting | GPT-3.5 generator/checker · MATH* | answer accuracy: 35.8 -> 38.0; 2.2±0.7 | Helped |
| Use SelfCheck confidence scores for weighted voting | GPT-4 generator/checker · MATH* | answer accuracy: 47.9 -> 51.3; 3.4±0.6 | Helped |
| Use GPT-3.5 to check GPT-4 solutions and weight votes | GPT-4 generator · MATH* | answer accuracy: 47.9 -> 48.9; 1.0±0.8 | Helped |
| Use multi-stage step checking instead of global solution checking (ablation) | SelfCheck on MathQA with GPT-3.5 · MathQA | verification accuracy: 55.0% -> 66.7% | Helped |
| Use multi-stage step checking instead of direct single-stage step checking (ablation) | SelfCheck on MathQA with GPT-3.5 · MathQA | verification accuracy: 57.2% -> 66.7% | Helped |
| Regenerate and compare steps instead of direct zero-shot error checking (ablation) | SelfCheck on MathQA with GPT-3.5 · MathQA | verification accuracy: 63.1% -> 66.7% | Helped |
Surprising findings
- Using GPT-3.5 to check GPT-4 generated solutions outperformed GPT-4 self-checking on GSM8K and MathQA.
- Directly asking the LLM to check whole reasoning chains yields near-100% true and false positive rates, providing no useful signal.
Limitations
- Ablations run on only 100 MathQA questions due to budget; main GPT-4 experiments use 500 sampled questions with only 2 solutions each.
- Relies on closed API models (GPT-3.5, GPT-4); cost of many LLM calls per step is high.
- Evaluated only on math reasoning; generality to other domains is claimed but not tested.
Experimental setup
Datasets: GSM8K MathQA MATH
Benchmarks: GSM8K MathQA MATH
Metrics: final answer accuracy verification accuracy TP/FP rates (ROC) accuracy gain over majority voting
Baselines: majority voting Self-Verification Deductive Verification Global Check Single-stage Check Error Check
Builds on: chain-of-thought prompting self-consistency decoding Self-Verification Deductive Verification
Closest papers by meaning
- Can LLMs Correct Themselves? A Benchmark of Self-Correction in LLMs (NeurIPS 2025 · similarity 0.90)
- Incentivizing LLMs to Self-Verify Their Answers (NeurIPS 2025 · similarity 0.90)
- S^3cMath: Spontaneous Step-Level Self-Correction Makes Large Language Models Better Mathematical Reasoners (AAAI 2025 · similarity 0.89)
- Diagnosing Multi-step Reasoning Failures in Black-box LLMs via Stepwise Confidence Attribution (ICML 2026 · similarity 0.88)
- CITE: Anytime Valid Statistical Inference in LLM Self-Consistency (NeurIPS 2026 · similarity 0.88)
- Sample Smart, Not Hard: Correctness-First Decoding for Better Reasoning in LLMs (ICLR 2026 · 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.