TxBench-OD is a verifiable benchmark of 113 of AI agent evaluations assessing oligonucleotide discovery ability based on published ASO and siRNA studies. This is Latch's the third focused slice of a broader benchmarking effort across drug-discovery stages and therapeutic modalities, after TxBench-PP and TxBench-OD. The manuscript can be found at latch.bio/txbench-oligo and the live scoreboard is on benchmarks.bio/txbench-od.
evals/: seven public example evaluations with prompt/config metadata.trajectories/: public trajectories for selected model-harness configurations for the public evaluations.paper/: folder containing paper and scripts to generate the figures