Research

I work on systems that turn open-ended work into trainable environments, verifiers, rewards, and model-improvement loops.

Current threads

Training scientific judgment

Verified campaign environments convert search, trust, escalation, and revision into a multi-turn RL problem with physics-grounded oracle reward.

post code model

Scaling test-time verification for novel materials

Probe-gradient guidance extracts band-gap signal from an unconditional crystal diffusion model and steers sampling without retraining.

post code model

Self-improving pretraining

Ordinary pretraining chunks become continuation-choice data, interleaved-thought SFT examples, and RL mid-training environments.

post code model dataset

Do language models know when to change their mind?

Signal detection measures whether models revise for valid critique while resisting invalid critique across architecture, domain, and source pressure.

post code

Papers

RLVR / Code

Surprisal-Guided Selection: Compute-Optimal Test-Time Strategies for Execution-Grounded Code Generation

RL-trained 120B KernelBench policy plus verifier-filtered selection under fixed inference budgets.

paper code model

Security agents / Calibration

OpenSec: Measuring Incident Response Agent Calibration Under Adversarial Evidence

Dual-control incident-response environment measures whether agents verify evidence before containment under adversarial context.

paper code

Agents / Continual learning

Continual Learning, Not Training: Online Adaptation For Agents

ATLAS frames agent improvement as online adaptation from trajectories rather than static fine-tuning alone.

paper code docs stateful agents on-policy distillation

Production and OSS infrastructure

SGLang

Model-serving optimization contributions in the open-source inference stack.

Slime

Multi-turn RL training cookbook contributions for agentic post-training recipes.

CL-Bench

Evaluation harness for benchmarking continually learning agents.