A new research paper investigates whether biological reasoning models effectively utilize their biological inputs. The study tested six models across DNA, protein, and single-cell tasks by perturbing biological inputs and analyzing model outputs. Findings indicate that some models, like Evo2 and ESM3, contribute minimally to performance, with models often relying more on text than provided biological data. However, other models such as ChatNT, Prot2Text-V2, and CellWhisperer showed that foundation model inputs do contribute to their performance. AI
IMPACT Raises questions about the effectiveness of current training strategies for biological reasoning models.
RANK_REASON Research paper analyzing model behavior and input utilization. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →