Last updated: June 19, 2026.

• Actively developing an LLM multi-agent system for large-scale, automated curation of DNA methylation datasets.
  • Designed a fully automated and generalizable pipeline that adapts to unseen datasets and aligns with the workflow needs of biological researchers.
  • The pipeline scales to thousands of publicly available datasets and reduces curation workloads that previously required years of manual effort by multiple biologists to just a few hours.
  • Designed for deployment as a continuous system that automatically downloads, curates, and updates relevant public datasets, providing biologists with clean, standardized data to accelerate downstream research.
  • Developing a RAG system for agent memory storage and retrieval, enabling more accurate generation and allowing agents to self-evolve by learning from past experience.
• Developing a post-training method for LLM agents that replaces sparse final-outcome rewards with dense, verifiable step-level supervision for multi-step reasoning and tool calling.
  • Designing a framework that extracts verifiable intermediate rewards from environmental responses without requiring human-annotated labels or pretrained reward models.
  • Training Qwen-2.5 7B with GRPO on interactive benchmarks (ALFWorld, ScienceWorld) for stable multi-step policy optimization.

• Developed Soft-Rank Diffusion, the first discrete diffusion framework that lifts permutations into a continuous soft-rank space, enabling scalable generative modeling over the combinatorially intractable symmetric group.
  • Designed an innovative forward process that replaces abrupt shuffle-based corruption with smooth reflected diffusion in continuous soft-rank space, yielding more structured trajectories that scale gracefully with sequence length.
  • Derived a hybrid reverse sampler that augments intractable discrete steps in permutation space with tractable continuous updates in soft-rank space, enabling efficient and principled generation.
  • Demonstrated superior performance on sorting and combinatorial optimization benchmarks, with particularly strong gains in the long-sequence regime where prior methods degrade.
  • Accepted at ICML 2026.
• Developed CaDDi, a Non-Markovian Discrete Diffusion framework that unifies discrete diffusion with causal LLMs—two core paradigms in natural language generation previously viewed as orthogonal.
  • By breaking the longstanding Markovian assumption in discrete diffusion models, CaDDi achieves more coherent text generation—demonstrated by lower generative perplexity compared to prior discrete diffusion approaches.
  • Finetuned a Qwen-3 1.5B LLM to perform diffusion-style denoising, achieving improved performance on multiple reasoning benchmarks.
  • Published as a co-first author at NeurIPS 2025.
• Developed CaLMFlow, a flow-matching paradigm that reformulates the problem from solving an ODE to a Volterra integral equation, mitigating stiffness issues commonly encountered in ODE-based formulations.
  • Applied CaLMFlow to single-cell perturbation response prediction, demonstrating superior ability to model the underlying data distribution and significantly improved extrapolation to OOD perturbations compared to specialized baselines.