Research
I work on systems that turn scientific judgment under uncertainty into replayable environments, verifiers, rewards, and model-improvement loops.
Current threads
Scientific exploration
SDL-2 measures whether AI agents can decide, from experiments they choose, whether a magnet process will reliably meet a stated performance requirement. Frontier models show an emerging ability to establish what an unfamiliar physical system can reliably do. The findings implicate reasoning under variation as a central capability for scientific exploration.
Scientific simulation
Reinforcement learning against a physics verifier lifts a 9B model from 9.0% to 55.7% pass@1 on ten unseen equation families, mostly by making correct numerical methods reliable.
Scientific autoresearch
The open-source Dynamical CLI is a unified interface for simulation and physical execution. In water electrolysis, agents selected the best catalyst 1.5× as often with Dynamical, while broader investigations traced how evidence shaped their scientific decisions.
Scientific post-training
Post-training Qwen3.6-35B-A3B on recorded materials experiments, combined with evidence guidance and test-time scaling, produced 23% lower average log loss than the base system. Its judgments of experimental outcomes and confidence better matched the recorded results.
Scientific evaluation
SDL-1 turns recorded materials experiments into controlled tests of scientific learning. Across six agent systems, understanding an inspected experiment, improving predictions about other experiments, and choosing useful experiments were distinct capabilities.
Test-time scaling
Proprio turns instrument corrections into skills a new agent can pick up. New GPT-5.6 Luna agents given repaired skills succeeded in 6 of 6 simulated sessions against 4 of 6 with the originals. On a robot workstation, a new GPT-6 Astra agent given skills placed a charger in 2 minutes rather than 14.
Verification
We studied how property estimates can guide crystal generation with Crystalite and MatterGen. With Crystalite's generator fixed, a trained probe raised predicted band-gap targeting from 0% to 24%; a separately trained checkpoint reached 43%.
Post-training
Ordinary pretraining chunks become continuation-choice data, interleaved-thought SFT examples, and RL mid-training environments, and a controlled follow-up shows that teacher thoughts get their value from reading ahead.
Interpretability
Signal detection measures whether models revise for valid critique while resisting invalid critique across architecture, domain, and source pressure.
Papers
Scientific agents / Benchmark construction
A benchmark-construction method that turns disagreement between public scientific models into verifier-backed agent tasks. The first release is 165 regulatory-genomics triage tasks where 6 frontier agents reach an 11.3% aggregate strict solve rate and none exceed 15%.
Post-training / Trajectory data
An incentive-aligned agent arena becomes a post-training trajectory substrate. A structural-quality filter turns raw ShoppingBench subnet traces into a trainable corpus that lifts Qwen3-4B from 18.0% to 42.7% strict solve rate, training on a fraction of a single day of arena output.
RLVR / Code
RL-trained 120B KernelBench policy plus verifier-filtered selection under fixed inference budgets.
Security agents / Calibration
Dual-control incident-response benchmark measuring whether agents verify evidence before containment under adversarial context, with deterministic oracle scoring and published traces.
Agents / Continual learning
ATLAS modeled the advisor/executor pattern as a continual-learning system. A stronger teacher guided a lower-cost production agent, while reward-scored trajectories were stored as persistent learning memory and reused for inference-time adaptation, RL, and on-policy distillation.