In-Context Examples Suppress Scientific Knowledge Recall in LLMs

TL;DR

Adding in-context examples to large language models shifts their reasoning from relying on scientific knowledge to pattern fitting, which can affect accuracy in scientific tasks.

Contribution

This paper demonstrates that in-context examples can suppress scientific knowledge recall in LLMs, leading to a shift towards empirical pattern fitting across multiple domains.

Findings

In-context examples reduce models' reliance on pretrained scientific knowledge.

The shift towards pattern fitting varies in accuracy impact depending on the task.

Displacement of knowledge occurs consistently across different scientific domains.

Abstract

Scientific reasoning rarely stops at what is directly observable; it often requires uncovering hidden structure from data. From estimating reaction constants in chemistry to inferring demand elasticities in economics, this latent structure recovery is what distinguishes scientific reasoning from curve fitting. Large language models (LLMs) can often recall and apply relevant scientific formulas, but we show that this ability is surprisingly easy to suppress. We show that adding in-context examples makes models rely less on pretrained domain knowledge, even when those examples are generated by the very same formula. Rather than reinforcing knowledge-driven derivation, examples shift computation toward empirical pattern fitting. We document this knowledge displacement on 60 latent structure recovery tasks across five scientific domains, 6,000 trials, and four models. This displacement is consistent across domains, but its accuracy consequences depend on how the displaced strategy compares to the one that replaces it: the same shift can lower accuracy, leave it unchanged, or appear to improve it. In all cases, however, the model shifts away from knowledge-driven reasoning. For practitioners deploying LLMs on scientific tasks, the message is cautionary: in-context examples may displace, rather than reinforce, the knowledge they are intended to support.