-
Background The paper discusses the reflective capacity of Large Language Models (LLMs). Existing LLMs optimize their responses through post-hoc prompting strategies such as reflection and self-refinement based on self-evaluated or external feedback. However, recent findings suggest that without external feedback, LLMs' intrinsic reflection is inconsistent. In standard self-reflection strategies, only 15.1% of incorrect responses are corrected after the first attempt, and the feedback is either overconfident (46.7%) or inconsistent (45.7%), significantly impairing reflection's effectiveness.
-
Existing Work The current research indicates that LLMs have difficulty self-correcting previous responses without external feedback. This suggests that self-correction is unreliable when relying solely on the LLM itself and simple post-hoc prompting strategies. LLMs often exhibit overconfidence or high randomness when self-evaluating, providing stubborn or inconsistent feedback, which results in poor reflection.
- Proposed 'Self-Contrast' Strategy
-
Challenge 1: Improving the quality of LLM's self-reflection and avoiding overconfidence and inconsistency To address the quality issues with self-evaluated feedback, researchers propose the "Self-Contrast" method. This method autonomously creates diverse solving perspectives, contrasts their differences, and summarizes these discrepancies into a checklist to re-examine and eliminate inconsistencies. This approach endows LLMs with diverse perspectives to mitigate stubborn biases and serves as a catalyst for more accurate and stable reflection.
-
Challenge 2: Ensuring the effectiveness and generality of the method The researchers conducted experiments using different LLMs on a series of reasoning and translation tasks to highlight the strategy's effectiveness and general applicability. They observed that contrasting LLMs' multiple solutions to find discrepancies typically points toward potential errors, details that are often overlooked, or pitfalls, thereby stimulating more accurate and stable self-correction through reflection on these differences.
-
The experiments demonstrate that the self-contrast method effectively increases the accuracy and consistency of LLMs in self-evaluation. The paper emphasizes that the self-correction success rate can be significantly improved by exploring various solving perspectives to contrast discrepancies and summarizing them into insightful checklists, as opposed to the existing standard reflection process that relies only on an initial response and subsequent self-evaluation. This approach offers a new strategy to enhance the introspection and self-correction capabilities of LLMs and validates its practical effectiveness and widespread applicability through a series of experiments.
This paper introduces a new strategy called "Self-Contrast" to address issues of stubbornness and inconsistency in reflection and self-correction processes within Large Language Models (LLMs). By creating diverse solving perspectives, contrasting different solutions, and summarizing disparities into a checklist, it enhances the quality of LLM reflection. The approach's effectiveness and broad applicability are validated through experiments.