SELF-[IN]CORRECT: LLMs Struggle with Discriminating Self-Generated Responses

TL;DR

This paper investigates whether large language models (LLMs) can reliably improve their outputs by discriminating among their own previous responses, finding that they are not consistently better at discrimination than initial generation.

Contribution

The paper introduces a unified framework to compare generative and discriminative capabilities of LLMs across tasks and demonstrates that models do not outperform in discrimination tasks compared to initial response generation.

Findings

Models are not reliably better at discriminating than generating initial responses.

Discrimination ability of LLMs does not surpass their initial response quality.

Challenges the idea that LLMs can self-improve through discrimination.

Abstract

Can LLMs consistently improve their previous outputs for better results? For this to be true, LLMs would need to be better at discriminating among previously-generated alternatives, than generating initial responses. We explore the validity of this hypothesis in practice. We first formulate a unified framework that allows us to compare the generative and discriminative capability of any model on any task. In our resulting experimental analysis of several open-source and industrial LLMs, we observe that models are not reliably better at discriminating among previously-generated alternatives than generating initial responses. This finding challenges the notion that LLMs may be able to enhance their performance only through their own judgment.