Research map / Large language models / Reasoning and chain of thought
Once-More: Continuous Self-Correction for Large Language Models via Perplexity-Guided Intervention
Jiaxun Gao, Him Wai Ng, Z. Jane Wang
Read the paperCheck your idea against 137,217 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 Once-More, a method for continuous self-correction of LLMs during long generation, guided by perplexity-based intervention at inference time.
Problem
Compounding errors during long LLM text generation, where early mistakes propagate into drift, faulty reasoning, or repetition, and self-correction behaviour is unpredictable at inference time.
Approach
Uses perplexity as a guide to intervene during generation so the model continuously self-corrects rather than relying on unpredictable emergent self-correction from scale; details beyond the abstract are not available.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| apply Once-More continuous self-correction | Qwen3-14B · AIME24 | accuracy: 26.7 -> 36.7 | Improved |
| apply Once-More | Qwen3-4B · LiveBench (reasoning) | accuracy: 20.3 -> 33.0 | Improved |
| apply Once-More | Qwen3-4B · GPQA Diamond | accuracy: CRITIC 48.9 vs Once-More 47.5 | Mixed |
| apply Self-Refine | Qwen3-8B · AIME25 | accuracy: 16.7 -> 15.5 | Hurt |
| add verifier feedback on top of logit redistribution (ablation) | Once-More Qwen3-14B · GPQA Diamond | accuracy: 48.4 -> 55.6 | Improved |
| add logit redistribution on top of verifier feedback (ablation) | Once-More Qwen3-14B · AIME24 | accuracy: 33.3 -> 36.7 | Improved |
| increase unit length from 1 to 128 sentences (ablation) | Once-More Qwen3-14B · AIME24 | accuracy: 36.6 -> 26.7 | Hurt |
| vary unit length among 1, 2, 4 sentences (ablation) | Once-More Qwen3-14B · AIME24 | accuracy: 36.6 / 34.3 / 35.5 | No clear effect |
| increase suppression strength alpha from 0.1 to 1 (ablation) | Once-More Qwen3-14B · AIME24 | accuracy: 30.0 -> 36.6 | Improved |
| use position-only suppression (tau=0.01) instead of distance decay tau=1 (ablation) | Once-More Qwen3-14B · AIME24 | accuracy: 36.6 -> 23.3 | Hurt |
| use stronger 14B verifier with 4B producer instead of 4B verifier (ablation) | Once-More Qwen3-4B · AIME24 | accuracy: 16.7 -> 33.3 | Improved |
| apply Once-More vs token usage of Self-Refine | Qwen3-14B · AIME | token usage: 17-21% fewer tokens | 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.
- CRITIC — CRITIC: Large Language Models Can Self-Correct with Tool-Interactive Critiquing (ICLR 2024) · better on 8 of 8 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
- Large Language Models Cannot Self-Correct Reasoning Yet (ICLR 2024 · similarity 0.93)
- Mind the Gap: Examining the Self-Improvement Capabilities of Large Language Models (ICLR 2025 · similarity 0.91)
- Small Language Model Can Self-Correct (AAAI 2024 · similarity 0.91)
- Can LLMs Correct Themselves? A Benchmark of Self-Correction in LLMs (NeurIPS 2025 · similarity 0.90)
- Training Language Models to Self-Correct via Reinforcement Learning (ICLR 2025 · similarity 0.90)
- A Theoretical Understanding of Self-Correction through In-context Alignment (NeurIPS 2024 · 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.
