IC-852Intrinsic self-correction without external feedback consistently degrades reasoning accuracy across GPT-3.5-turbo, GPT-4, GPT-4-turbo, and LLaMA-2-70B-chat

Jie Huang, Xinyun Chen, Swaroop Mishra, Huaixiu Steven Zheng, Adams Wei Yu, Xinying Song, Denny Zhou

SourceLarge Language Models Cannot Self-Correct Reasoning Yet

The paper evaluates whether LLMs can improve their reasoning by reviewing and revising their own answers without any oracle label or external feedback. Across GSM8K, CommonsenseQA, and HotpotQA, every model tested shows a drop in accuracy after one or two rounds of self-correction. The degradation is most severe for LLaMA-2 (GSM8K drops from 62.0 to 36.5) and GPT-3.5 on CommonsenseQA (75.8 to 38.1). The authors attribute the drop to the model's inability to judge the correctness of its own reasoning, combined with the self-correction prompt biasing the model away from its initial (often correct) answer.

Evidence
correlational
Key metric
GPT-3.5: GSM8K 75.9→75.1→74.7, CommonsenseQA 75.8→38.1→41.8, HotpotQA 26.0→25.0→25.0; GPT-4: GSM8K 95.5→91.5→89.0, CommonsenseQA 82.0→79.5→80.0, HotpotQA 49.0→49.0→43.0; GPT-4-turbo: GSM8K 91.5→88.0→90.0, CommonsenseQA 84.0→81.5→83.0; LLaMA-2: GSM8K 62.0→43.5→36.5, CommonsenseQA 64.0→37.5→36.5
Caveat
The paper focuses on reasoning tasks; the authors note that self-correction may still be effective in other domains such as style or safety alignment. Sample sizes for GPT-4-turbo and LLaMA-2 are 200 questions per dataset (100 for HotpotQA) rather than the full test set.
Model
GPT-3.5 / ChatGPT-3.5 GPT-3.5-turbo, GPT-4 / ChatGPT4 / GPT-4 Code Interpreter / GPT-4 Technical Report GPT-4 Turbo, Llama 2 / Llama 2 base
Concepts
Failure mode
Datasets
GSM8K [eval], CommonsenseQA [eval], HotpotQA [eval]
Methods
Self-Consistency / Self-consistency prompting / Wang et al. 2023 (self-consistency) / Wang et al. 2023b (Self-Consistency) [compared-to]
Related work
Self-Refine [compared-to]
Related findings
IC-853, IC-854
Extraction
automatic-extraction