Research
Papers
-
PIPer: On-Device Environment Setup via Online Reinforcement LearningNeurIPS 2025 Workshop on Deep Learning for Code
An 8B model trained with SFT and RL from verifiable rewards that matches GPT-4o and Qwen3-32B at environment setup.
-
EnvBench: A Benchmark for Automated Development Environment SetupICLR 2025 Workshop on Deep Learning for Code
The largest dataset to date for environment setup, with a framework for building and evaluating LLM agents on it.
-
Long Code Arena: A Set of Benchmarks for Long-Context Code ModelsarXiv preprint
Six benchmarks for code models whose context spans a whole repository.
-
On the Importance of Reasoning for Context Retrieval in Repository-Level Code EditingarXiv preprint
More deliberate reasoning helps coding agents find the code that matters, though not decide when they have found enough.
-
Beyond Dynamics: Learning to Discover Conservation PrinciplesMachine Learning: Science and Technology
Representation learning and topology that recover conservation laws from data, including in quantum systems.




