Building Foundation Models for System 2 Reasoning.
I build Self-Evolving Scientific Agents—systems that move beyond generative fluency to rigorous, System 2 understanding.
My research lies at the intersection of Agentic Reasoning, RL Post-Training, and Generative Planning. By aligning LLMs with structured verification loops and non-autoregressive mechanisms (Diffusion), I aim to enable AI systems to perform reliable, multi-step deduction in open-ended domains.
Defining and measuring rigorous thinking. I build comprehensive benchmarks (e.g., MolPuzzle) and evaluation protocols to stress-test LLMs on multi-step deduction, identifying the gap between fluency and true reasoning.
Orchestrating self-correcting agents. I combine Retrieval-Augmented Generation (RAG) with Reinforcement Learning (RL) post-training to align agentic workflows with structured knowledge, ensuring verifiable and robust planning.
Beyond autoregressive limits. I explore Diffusion Models as non-autoregressive planners for structured generation. This enables global context modeling and precise controllability in complex design spaces.
"Adaptive reasoning to enable more flexible thinking."
"The first multimodal benchmark evaluating if LLMs can deduce molecular structures from spectroscopic data."
"A massive community benchmark establishing the new standard for AI in Science."
"Developed a multi-agent orchestration framework for chemical synthesis planning."
"A comprehensive survey on the state of AI for spectroscopic analysis."
"One of the earliest comprehensive benchmarks for LLMs in the chemistry domain."