Evaluate post-training objectives that reward correct use of biological inputs

Determine whether post-training objectives that reward correct use of biological inputs improve the contribution of those inputs to biological reasoning task performance.

Background

The paper evaluates six biological reasoning models spanning genomic, protein, and single-cell tasks, focusing on whether foundation-model representations and biological text actually influence model predictions. Its perturbation and evidence-conflict experiments show that some models achieve high benchmark accuracy while making limited use of predictive biological representations, particularly BioReason with Evo2 representations and BioReason-Pro with ESM3 representations.

The authors further find that standard supervised fine-tuning and reinforcement learning can improve overall accuracy without consistently increasing the performance contribution of biological inputs. They therefore identify the design of post-training objectives that explicitly reward correct responses to changes in biological inputs as a direction requiring evaluation, with the unresolved issue being whether such objectives would cause models to use those inputs more effectively.