Research map / Large language models / Agents, code and math

CRITIC: Large Language Models Can Self-Correct with Tool-Interactive Critiquing

ICLR 2024 · Large language models › Agents, code and math · approach: Execution-based code verification · 57 citations (OpenAlex)

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.

Introduces CRITIC, a training-free framework where frozen LLMs verify their outputs by calling external tools (Google search, Python interpreter, Perspective API) and iteratively correct them via in-context learning.

Problem

LLMs produce hallucinated facts, flawed code, and toxic text, and existing fixes need costly human annotation or task-specific training; self-correction without external feedback is unreliable.

Approach

Interleave LLM-generated verification queries with tool API results (search engine, code interpreter, toxicity scorer) in few-shot prompts, so the model critiques its own initial output using tool feedback and then revises it, repeating verify-then-correct loops without any weight updates.

Key results

• On ChatGPT, CRITIC improves F1 by 7.7 on QA tasks over CoT (CRITIC* +12.4) and reduces toxicity probability from 0.192 to 0.040 on RealToxicityPrompts.
• On math program synthesis with ChatGPT, CRITIC raises GSM8k accuracy from 72.5 (PoT) to 78.2 and TabMWP from 75.0 to 89.0.
• Without tool feedback (CRITIC w/o Tool), self-correction gives marginal or negative changes, e.g. -1.8 on text-davinci-003 GSM8k.

Measured effects reported in the paper

Surprising findings

• Self-correction using only the LLM's own critiques yields modest gains or even degrades performance across tasks, showing LLMs are unreliable self-verifiers.
• Many 'errors' on HotpotQA under automatic F1/EM are label ambiguity or outdated labels rather than model mistakes (49% of CRITIC error samples).

Limitations

• Inference latency increases linearly with the number of verify-correct iterations.
• Results depend on manually crafted in-context demonstrations and prompt construction.
• Evaluated only on text tasks; tool choice is pre-specified per task rather than automatically selected.

Experimental setup

Datasets: AmbigNQ TriviaQA HotpotQA GSM8k SVAMP TabMWP RealToxicityPrompts

Benchmarks: AmbigNQ TriviaQA HotpotQA GSM8k SVAMP TabMWP RealToxicityPrompts

Metrics: EM F1 exact match maximum toxicity toxicity probability perplexity dist-2 dist-3 AUROC

Baselines: Vanilla prompting CoT Self-Consistency ReAct Rejection Sampling PoT Self-Refine PPLM GeDi DExperts PPO Quark

Builds on: ReAct Program-of-Thought Chain-of-thought prompting Self-Consistency Self-Correct

Closest papers by meaning

• CriticEval: Evaluating Large-scale Language Model as Critic (NeurIPS 2024 · similarity 0.91)
• MINT: Evaluating LLMs in Multi-turn Interaction with Tools and Language Feedback (ICLR 2024 · similarity 0.89)
• Large Language Models Cannot Self-Correct Reasoning Yet (ICLR 2024 · similarity 0.89)
• Toward Self-Improvement of LLMs via Imagination, Searching, and Criticizing (NeurIPS 2024 · similarity 0.89)
• Accelerating Unbiased LLM Evaluation via Synthetic Feedback (ICML 2025 · similarity 0.89)
• Teaching Language Models to Critique via Reinforcement Learning (ICML 2025 · similarity 0.89)