LLMs can implicitly learn from mistakes in-context

We consider the scenario where an LLM outputs a corrective rationale for an erroneous answer, then uses it to improve its next answer akin to explicit learning in humans—a phenomenon whereby patterns and structure within a new piece of information are deliberately sought out and verbalised to aid reasoning and abstraction (Stadler, 1997). This active process differs from implicit learning, where complex skills are passively acquired simply by observing the environment, and any pattern detection occurs implicitly and automatically (Frensch and Rünger, 2003; Kaufman et al., 2010).

We generate the corrective rationales in Eexplicit following a strategy similar to that described in An et al. (2023): we prompt GPT-4 in a few-shot fashion, showing it questions with incorrect and correct answers, as well as rationales. Given a new question and pair of answers, we ask the model to identify the mistakes in the incorrect answer and explain how to correct them. We use the same few-shot examples as An et al. (2023), slightly reformatted for our task.

Discussion

Our results demonstrate that LLMs perform better across several mathematical reasoning tasks when they are prompted for implicit learning, even over CoT prompting and providing the models with additional information through rationales. To minimise any risk that spurious correlations may be influencing these results, here we provide further, in-depth analysis of our findings, their robustness and implications.