The reflection capacity of Large Language Model (LLM) has garnered extensive
attention. A post-hoc prompting strategy, e.g., reflexion and self-refine,
refines LLM’s response based on self-evaluated or external feedback. However,
recent research indicates without external feedback, LLM’s intrinsic reflection
is unstable. Our investigation unveils that the key bottleneck is the quality
of the self-evaluated feedback. We find LLMs often exhibit overconfidence or
high randomness when self-evaluate, offering stubborn or inconsistent feedback,
which causes poor reflection. To remedy this, we advocate Self-Contrast: It
adaptively explores diverse solving perspectives tailored to the request,
contrasts the differences, and summarizes these discrepancies into a checklist
which could be used to re-examine and eliminate discrepancies. Our method
endows LLM with diverse perspectives to alleviate stubborn biases. Moreover,
their discrepancies indicate potential errors or inherent uncertainties that
LLM often overlooks. Reflecting upon these can catalyze more accurate and
stable reflection. Experiments conducted on a series of reasoning and
translation tasks with different LLMs serve to underscore the effectiveness and
generality of our strategy.
Improving Large Language Models’ Reflection Capacity through Self-Contrast
The reflection capacity of Large Language Models (LLM) has become a topic of extensive research and discussion. LLMs have the ability to generate responses based on self-evaluation or external feedback. However, recent studies have found that LLMs’ intrinsic reflection is often unstable in the absence of external feedback.
This instability can be attributed to the quality of self-evaluated feedback. LLMs tend to exhibit overconfidence or high randomness when evaluating their own responses, leading to stubborn or inconsistent feedback. This, in turn, hinders the LLM’s reflective abilities.
To address this issue, the authors propose a novel strategy called “Self-Contrast”. This strategy enables LLMs to adaptively explore diverse perspectives tailored to the given request. By contrasting the differences between these perspectives, the LLM generates a checklist that helps re-examine and eliminate discrepancies.
The Self-Contrast method provides LLMs with a range of perspectives, alleviating stubborn biases. It also helps identify potential errors or uncertainties that LLMs often overlook. Reflecting upon these discrepancies can enhance the accuracy and stability of the LLM’s reflection.
The authors conducted experiments on a series of reasoning and translation tasks using different LLMs to validate the effectiveness and general applicability of their self-contrast strategy.
This research highlights the multi-disciplinary nature of large language models. It combines insights from natural language processing, machine learning, and cognitive science to improve the reflective capacities of LLMs. By understanding and addressing the limitations of self-evaluated feedback, the study presents an innovative approach that can enhance the performance of LLMs in various tasks.