Benchmarks and Evals
Benchmarks, environments, and evaluation methodology for coding, tool use, computer use, and expert-level professional work.
My team builds benchmarks and environments for coding agents, tool and MCP use, computer use, and expert-level professional work, and develops the evaluation methodology behind Scale’s public leaderboards.
Recent papers include READY (Chatrath et al., 2026), a framework for qualifying agents for reliable enterprise deployment under human oversight and cost constraints; ChainWorld (Siu et al., 2026), which composes atomic OSWorld tasks into long-horizon desktop workloads and shows that current computer-use agents complete at most 31% of them; and DrugDiscoveryBench (Akyürek et al., 2026), an expert-curated benchmark of early-stage drug-discovery tasks for coding agents (see also AI for Science).