#19 · ML / AI methods

AI agents tested on single-cell, genomics, and published-study reproduction tasks

The dominant thread is evaluation of LLM agents that carry out multistep biological analysis. scBench-Long assembles 21 long-horizon single-cell evaluations in which agents must recover scientific claims from raw data. GeneBench-Pro benchmarks 129 multistage statistical analysis problems across genomics, quantitative biology, and translational biomedicine, with a best reported pass rate of 31.5%. A separate preprint reports that five AI research agents reproduce three published studies, showing promise but also hallucination and variation across runs. Octopus, an LLM-swarm pipeline, represents the more discovery-oriented case: it nominates IGF2 as a 5-FU resistance vulnerability in colorectal cancer, checked against PDX and survival data. What is new is assessment under verifiable or published-study conditions rather than demonstration alone.