#19 · ML / AI methods

LLM agents and biological models are being tested and constrained against explicit…

The common thread is constraining and auditing language-model-based systems in biology against explicit evidence rather than fluency. EDB-Bio (96k episodes from drug transcriptomics, CRISPR screens and GWAS) tests whether LMs weigh biological evidence. Two agent preprints locate performance in scaffolding: a prompting harness, not model tier, drives hypothesis quality in ABCD-Stack, and DTDR scaffolding lifts a bioinformatics sandbox agent from 0% to 9.09% accuracy while cutting steps 66%. AutoSPECTRA uses a coding-agent protocol to audit generalizability of foundation models across protein, perturbation and pathology tasks. PPI-Caliper adds conformal calibration to protein language model interaction scores, controlling false discovery rate without retraining across five species. New here is pairing evaluation with explicit procedural and statistical controls.