Boran Han

Research

Models that learn, reason, and adapt.

My work connects reasoning, reinforcement learning, autonomous agents, foundation-model pretraining, and efficient systems. These programs share one goal: useful intelligence that remains adaptive, grounded, and practical.

Research programs / 01-05

Five connected lines of inquiry.

01 / Reinforcement learning

Reinforcement learning

Reconciling skill discovery with policy optimization and making large-scale LLM reinforcement learning efficient across heterogeneous compute.

ReSkill framework for skill creation and policy optimization in agentic reinforcement learning 01 ReSkillReconciling Skill Creation with Policy Optimization in Agentic RL 2026 HetRL reinforcement learning framework for heterogeneous compute environments 02 HetRLEfficient Reinforcement Learning for LLMs in Heterogeneous Environments MLSys 2026
Skill creation / heterogeneous RL

02 / Pretraining

Pretraining & foundation models

Learning broad priors for time series, tabular data, and multisensor Earth observation so one model can transfer across data regimes and downstream tasks.

Chronos-2 architecture for universal time-series forecasting 01 Chronos-2From Univariate to Universal Forecasting 2025 Mitra mixed synthetic priors and tabular foundation model architecture 02 MitraMixed Synthetic Priors for Enhancing Tabular Foundation Models NeurIPS 2025 msGFM cross-sensor pretraining and downstream transfer architecture 03 msGFMBridging Remote Sensors with Multisensor Geospatial Foundation Models CVPR 2024
Universal forecasting / synthetic priors / multisensor pretraining

03 / Reasoning

Reasoning

Injecting specialist knowledge, coordinating multiple reasoning paths, and grounding multimodal models in the context needed to solve complex problems.

SenTSR-Bench knowledge injection and time-series reasoning framework 01 SenTSR-BenchThinking with Injected Knowledge for Time-Series Reasoning AISTATS 2026 CaMML examples of context-aware multimodal reasoning 02 CaMMLContext-Aware Multimodal Learner for Large Models ACL 2024 CoMM collaborative multi-agent and multi-reasoning-path prompting method 03 CoMMCollaborative Multi-Agent, Multi-Reasoning-Path Prompting for Complex Problem Solving NAACL 2024 Examples of incorrect citation and omitted information when language models reason over tables 04 Table DREsWhen LLMs Read Tables Carelessly: Measuring and Reducing Data Referencing Errors 2026
Knowledge injection / multimodal context / collaborative reasoning / table reliability

04 / Agents

Agents

Building autonomous systems that create reusable capabilities, coordinate specialized reasoning roles, and automate end-to-end machine learning workflows.

MLZero multi-agent system for end-to-end machine learning automation 01 MLZeroA Multi-Agent System for End-to-end Machine Learning Automation NeurIPS 2025 CoMM collaborative multi-agent and multi-reasoning-path prompting method 02 CoMMCollaborative Multi-Agent, Multi-Reasoning-Path Prompting NAACL 2024 ReSkill agent creating and evaluating reusable skills during policy optimization 03 ReSkillSkill creation and policy optimization for adaptive agents 2026
Workflow automation / multi-agent collaboration / reusable skills

05 / Systems

Systems

Co-designing algorithms and infrastructure to keep reinforcement learning and retrieval-augmented generation efficient across real hardware environments.

HetRL reinforcement learning framework for heterogeneous compute environments 01 HetRLEfficient Reinforcement Learning for LLMs in Heterogeneous Environments MLSys 2026 PipeRAG pipeline parallel retrieval and generation system 02 PipeRAGFast Retrieval-Augmented Generation via Algorithm-System Co-design KDD 2025
Heterogeneous training / pipelined retrieval and generation