The paper measures whether ChatGPT (GPT-3.5-turbo) and Text-davinci-003 can judge the correctness of their own answers and fix errors using only their internal knowledge. On self-verification (Table 5), the self-eval method achieves only 54% accuracy on HotpotQA (near random) and performs worse than the 'only-true' baseline on TriviaQA. On self-correction without tools (CRITIC w/o tool), the models' own critiques contribute marginally (-0.03 and +2.33 F1 on QA) and can even deteriorate performance, as seen in the 1.8-point decrease on GSM8K for Text-davinci-003 and the toxicity scores remaining near baseline levels. The authors conclude that 'exclusive reliance on self-correction without external feedback may yield modest improvements or even deteriorate performance.'
Evidence
correlational
Key metric
self-eval accuracy 54% on HotpotQA; CRITIC w/o tool: -0.03 and +2.33 F1 (QA, two LLMs); -1.8 on GSM8K (Text-davinci-003); toxicity 0.353 vs 0.344 baseline (Text-davinci-003), 0.339 vs 0.325 (ChatGPT)
Caveat
Results depend on the specific prompting strategy (CRITIC w/o tool) and the tasks tested; the authors note that 'different prompt constructions may impact the experimental results' and that the finding is demonstrated on three task types.