Publications

87 expert tasks with deterministic verifiers and paired with/without-skill runs: curated skills lift agent pass rates by +16.6 pp, and focused skills beat exhaustive ones.

Stateful mock Gmail, Slack, Calendar, Docs, and Drive: top agents succeed on 39-64% of tasks yet take unsafe actions 7-33% of the time.

Reward-integrity layer for agent benchmarks: static taint analysis finds hacking paths before a run, and runtime evidence detects hacking at 96% accuracy across 31k+ runs on SkillsBench, ClawsBench, and Terminal-Bench 3.

A desiderata for hallucination benchmarks and TRIVIA+, a long-context RAG benchmark with realistic label noise; LLM-as-a-Judge stays competitive.

Stop parallel reasoning early once the majority vote is provably safe: 25-47% fewer self-consistency tokens at full-budget accuracy.

A stochastic flow-matching policy in latent space, with a solver guaranteeing feasible actions: +20.6% over state-of-the-art combinatorial RL.