When Do Biological Reasoning Models Use Their Biological Inputs?

When Do Biological Reasoning Models Use Their Biological Inputs?. By Ada Fang and five others, from Harvard University and the Massachusetts Institute of Technology. Published on the first of October. Its code repository is worked out: about one hundred and thirty source files, with a readme and a licence.

Biological reasoning models use post-training to connect a language model to two kinds of input: representations from a biological foundation model, and biological text. Their benchmark accuracy is taken as evidence that the language model reasons over those inputs. This paper tests that assumption, and its conclusion is that current post-training strategies do not ensure the foundation model representations actually contribute to task performance. In some models they barely do.

Six biological reasoning models are examined, across DNA, protein and single-cell tasks, with four kinds of test. One biological input is perturbed while the query and every other input are held fixed. Evidence conflicts are built, pairing the foundation model representation of one genome, protein or cell with the text of another. Linear probes are fitted to the representations the language model receives. And reasoning traces are checked against the biological inputs.

For BioReason and BioReason-Pro, the foundation models Evo2 and ESM3 contribute little on the evaluated tasks. Shuffling the DNA sequence barely changes BioReason's disease prediction accuracy. In evidence conflicts, the two models follow the text in ninety-seven point nine percent and ninety-nine point seven percent of cases. Yet the information is there: linear probes trained on the Evo2 and ESM3 representations do predict the task targets. The foundation models encode what the task needs, and still add limited overall performance.

The traces tell a similar story. BioReason traces misstate nucleotide changes, and under evidence conflicts BioReason-Pro traces describe functions that are then omitted from the final predictions. Training history does not rescue it either. Across supervised and reinforcement learning checkpoints of BioReason-Pro, and forty-two checkpoints of BioReason, rising accuracy does not imply a greater contribution from the biological inputs.

The finding is not that every such model ignores its inputs. In contrast to the two BioReason models, foundation model inputs do contribute to ChatNT, Prot2Text-V2 and CellWhisperer performance. For Cell2Sentence-Scale, differentially expressed genes in the gene sentence contribute to performance. The negative results are also tied to the evaluated tasks. What generalises is narrower and sharper: benchmark accuracy alone cannot tell you which case a given model is in.

A score measures the output, not the route to it. In two of the six models, the evidence conflicts show answers following the text almost every time, while probes show the foundation model representations hold information relevant to the task. That information is there, and it is largely not what drives performance. Nothing in the accuracy number would reveal this, and the checkpoint analysis shows that pushing accuracy higher does not fix it either. Perturbing the inputs, setting them against each other, probing what the language model receives and reading the traces are what expose it. The broader lesson the authors draw is about post-training itself: as currently practised, it does not ensure that the biological representations a model is given end up contributing to what it does.