Research
Methods, findings, and resources for my work on reinforcement learning and language-model agents.
CoSIR · Preprint 2026
Post-train LLMs for collaboration with surprisal-reduction rewards in a verifiable toy environment, improving performance on out-of-distribution collaboration benchmarks.
Project pageMeshHeal · Preprint 2026
Peer review catches unreliable agent outputs, while longer-term monitoring guides exclusion and recovery.
Project pageIDS · NeurIPS 2026
Direct online exploration toward the uncertainty that remains after learning from offline data.
Project pageHIPO · EMNLP 2026
Train language models to help users while treating system-instruction compliance as an explicit constraint.
Project pageMarvel · TMLR 2026
Align value estimates and adapt constraint penalties to safely fine-tune offline policies online.
Project page