Llms struggle to reason about hidden chemical structure changes
Beyond Benchmark Accuracy: A Diagnostic Test for LLM Chemical Prediction
Computation and Language
Summary
Large language models (LLMs) are good at predicting chemical properties when they can see the names or identities of molecules. But it’s unclear if they truly understand how changes in molecular structure affect those properties or just remember facts linked to molecule names. The authors created a testing method that hides the identity of target molecules and instead describes them as changes made to known molecules. They found that LLMs often rely heavily on molecule labels rather than reasoning about the structural changes themselves. This means the high scores on usual tests may overestimate how well these models understand chemistry.
What this means in practice
- •For chemical informatics teams: Evaluate chemical prediction models more reliably by testing their reasoning about molecular changes rather than memorizing molecule names.
- •For machine learning engineers: Develop LLM-based tools that better understand chemical transformations by applying identity-blinded tests during model development.
Authors
Yanyi Su, Yu Zeng, Hongshuai Wang, Chenyan Wang, Zhifeng Gao, Xiaohui Yang
Abstract
LLMs have achieved strong performance on chemical prediction benchmarks. However, existing evaluation methods typically provide target entity identity information directly, making it difficult to determine whether model predictions arise from transferable structure-property reasoning capabilities or rely on target-related contextual information or memory retrieval. We propose the Identity-Blinded Reasoning Diagnostic framework to evaluate chemical prediction models when target entity identity information is hidden. In this framework, the target is no longer directly defined by its entity identity, but instead by a known source entity and an explicit structural edit operation. This setting preserves the information required to determine the target while reducing the opportunity for models to access target identity cues. We evaluate this framework across six tasks spanning odor prediction, molecular property prediction, retrosynthesis, and reaction yield prediction. The results show that models generally achieve better prediction performance when source and target entity properties are consistent. Providing explicit source entity labels further strengthens this dependence. Moreover, while keeping the target entity, structural change, and ground-truth label unchanged, replacing only the source entity label causes some model predictions to systematically shift toward the substituted source labels. These results indicate that high benchmark accuracy in chemical prediction tasks does not necessarily imply that LLMs possess reliable structure-property reasoning capabilities. Our diagnostic framework provides a controlled evaluation approach for testing whether models can genuinely update predictions according to entity changes. The framework can be extended to other AI for Science tasks, particularly those where targets can be defined through transformations of known entities.