Paperscope

The breadth is striking, but the authors explicitly acknowledge that they did not perform dedicated entity-level filtering between continual pretraining and downstream evaluation data. That does not prove any particular score is inflated; it means the scores cannot cleanly establish generalisation to entirely unseen biological entities. “Competitive or leading” also bundles different outcomes into one headline count. A second limitation is structural: the molecule and protein tokenisers use separate geometric reference frames, so the model cannot natively represent their complex in a shared pose. Broad modality coverage therefore does not yet imply native docking capability.

The benchmark design deserves credit for addressing sequence similarity and time, rather than relying on a convenient random split. The remaining boundary is functional validation. The design evaluation uses computational measures including predicted structural confidence, sequence novelty and annotation recovery. These measure useful properties, but a plausible fold and familiar functional signature do not guarantee activity, stability or expression in a real experiment. Outperforming general language models is also a less demanding comparison than replacing the best specialised tool for each biological task. The paper makes those specialist comparisons; the conclusions should retain their mixed, task-specific character.

Inference-time control is attractive, but “eradication” is much broader than the evidence. A more balanced distribution for selected attributes does not establish fairness across identities, contexts or combinations of people. It also leaves a normative choice: should outputs reflect population frequencies, equal representation or the user’s explicit request? The paper acknowledges attribute-binding failures and mismatches between occupational stereotypes and the model’s existing tendencies. Those are not minor edge cases; they show why a fixed direction of correction can interact unpredictably with context. The useful contribution is a controllable mitigation, not a declaration that the system is unbiased.

The distinction between removing a concept and removing an unwanted association is important. Models may need knowledge of a harmful behaviour without being encouraged to perform it. The scope of the result remains bounded by the theoretical assumptions and the three experimental constructions. A factor that is prominent in a low-rank update may be identifiable here while a messier mixture of correlated factors is not. Likewise, eliminating measured misalignment in one setting is a statement about those evaluations, not every possible harmful behaviour. Preserving pretrained content is valuable, but it also means the method is not intended to repair every problem already present in the base model.

Intermediate questions reveal errors that a single final label can hide. They do not automatically reveal the model’s actual causal reasoning: a model can answer each stage using separate shortcuts, while a later error can also be inherited from an earlier one. The large QA count needs to be interpreted alongside shared images and derived variants rather than as independent clinical encounters. The authors explicitly acknowledge missing patient context and the selection of informative 2D slices from volumetric data. Those choices make a useful controlled benchmark, while reducing the search and contextual integration required in clinical imaging.

The temporal split is a meaningful safeguard, and pairwise notes offer more focused assistance than generic disease descriptions. But a diagnostic memory is also a place where an incorrect generalisation can persist. A rule that separates two diseases in curated cases may fail when prevalence, imaging quality or patient presentation changes. Better average accuracy does not tell us whether the memory increases confidence on exactly those exceptions. The analogy with clinical experience should therefore be handled carefully: retrospective feedback with known answers is a cleaner learning signal than the delayed, ambiguous outcomes clinicians encounter.