Published in Annual Meeting of the Association for Computational Linguistics (ACL), 2026
A task-agnostic evolutionary pipeline that jointly synthesizes problems, candidate solutions, and verification artifacts for verifiable learning.
Published in International Conference on Machine Learning (ICML), 2026
TESSY interleaves teacher and student models to synthesize SFT data that keeps the teacher’s reasoning ability while matching the student’s style.
Published in IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2026
Confidence-as-a-Reward (CRew): a training-free reward signal built from a model’s token-level confidence in its final answer.
Published in Findings of the Association for Computational Linguistics: ACL 2025, 2025
An open-source framework for GitHub issue resolution built from a retrieval module and a code editing module, trained on 110K real issues.
Published in International Conference on Learning Representations (ICLR), 2025
ProverGen combines LLM generation with symbolic provers to build ProverQA, a scalable and diverse first-order logic reasoning benchmark.
Published in International Conference on Computational Linguistics (COLING), 2025
A benchmark evaluating LLMs across the entire software development lifecycle: design, environment setup, implementation, acceptance testing, and unit testing.
Published in Deep Learning for Code (DL4C) Workshop at ICML 2026, 2026
A framework for high-performance GPU kernel generation that pairs an evaluation-driven evolutionary agent with an evolution-oriented post-training recipe.