CRITIC: L

🌈 Abstract

The paper proposes a framework called CRITIC that allows large language models (LLMs) to self-verify and self-correct their outputs by interacting with external tools. Unlike traditional LLMs that sometimes exhibit inconsistencies and problematic behaviors, CRITIC empowers LLMs to validate and progressively amend their own outputs in a manner similar to human interaction with tools. The paper demonstrates the effectiveness of CRITIC across diverse tasks including free-form question answering, mathematical program synthesis, and toxicity reduction. The results highlight the crucial importance of external feedback in promoting the ongoing self-improvement of LLMs, as the authors find that exclusive reliance on self-correction without external feedback may yield modest improvements or even deteriorate performance.

🙋 Q&A

[01] Large Language Models and Their Limitations

1. What are some of the problematic behaviors that large language models sometimes exhibit? Large language models (LLMs) sometimes show inconsistencies and problematic behaviors, such as hallucinating facts, generating flawed code, or creating offensive and toxic content.

2. How do humans typically utilize external tools to refine their initial content? Humans typically utilize external tools to cross-check and refine their initial content, like using a search engine for fact-checking, or a code interpreter for debugging.

3. What is the key idea behind the CRITIC framework? The key idea behind CRITIC is to empower LLMs, which are essentially "black boxes", to validate and progressively amend their own outputs in a manner similar to human interaction with tools.

[02] The CRITIC Framework

1. What are the two main steps in the CRITIC framework? The CRITIC framework consists of two main steps: (1) verifying the output by interacting with external tools to generate critiques, and (2) correcting the output based on the received critiques.

2. How does CRITIC utilize in-context learning and tool interaction? CRITIC utilizes in-context learning with tool interaction to proficiently identify and rectify unsatisfactory behaviors using the LLM itself, without requiring expensive annotations or task-specific training.

3. What types of external tools does CRITIC interact with? CRITIC interacts with various external tools such as search engines, code interpreters, calculators, and text APIs to verify different aspects of the generated output.

[03] Experimental Evaluation

1. What are the key tasks used to evaluate CRITIC? CRITIC is evaluated on three distinct tasks: free-form question answering, mathematical program synthesis, and toxicity reduction.

2. How does CRITIC perform compared to previous techniques across these tasks? CRITIC consistently outperforms previous techniques, including self-consistency, ReAct, and rejection sampling, across the evaluated tasks and language models.

3. What is the key finding regarding the unreliability of LLMs in self-verification and self-correction? The results underscore the inadequacy of LLMs in self-verification and self-correction, highlighting the crucial importance of feedback from external tool interaction for consistent self-improvement of LLMs.