The gate before open models reach triage
-
Nº CXIII
- Date
- 05 Oct 2026
- Issue
- 113
- Stories
- Four
- Editor
- ARC
Monday, and the queue is all plumbing: bias gates, grounded annotations, antibody loops that respect their scaffold.
Bias audit tests open models on clinical triage
Counterfactual auditing lands on clinical triage , testing open-weight language models by rewriting the same patient vignette to change only demographic details and watching whether the urgency call moves. The target set matters: open-weight models are the ones hospitals and academic centers can host and modify themselves, so they are the likeliest first wave into real triage workflows. Counterfactual probing, which swaps one attribute while holding the medicine fixed, gives a cleaner signal than aggregate accuracy, which can look fine while hiding who the errors land on. The work puts identity-sensitivity testing alongside accuracy as something a triage evaluation is expected to report.
AbRefine designs antibody loops that fit their framework
AbRefine conditions CDR design on the antibody framework the loops will actually sit in, refining complementarity-determining regions against the surrounding scaffold and the target structure rather than generating loops in isolation. Structure-guided antibody generators have mostly treated the CDR as a free-floating design problem, handing back loops that fit the epitope but not the parent construct. Framework conditioning narrows that gap, moving generative antibody work closer to designs that survive the jump from model output to expressed protein.
GroundAnnot checks gene-set annotation against live tools
GroundAnnot forces gene-set annotation through a closed vocabulary, restricting what a language model may output to terms a live enrichment backend can verify instead of letting it narrate plausible-sounding pathway names. Free-text annotation has been the soft spot in LLM-assisted functional genomics: the prose reads right, the terms drift from any ontology, and nothing downstream can check them. Binding generation to a running backend makes annotation output auditable, which is the precondition for it appearing in a figure legend or a grant.
New model generates disordered protein regions
Disordered regions get a generator built by reinforcing sparse autoencoder features, an interpretability method that separates a model's internals into named, steerable components. Extends generative protein design into the disordered fraction that structure-first methods largely skip.
Reply with your discoveries. A human reads them. Forward freely.
|