Jiayan Guo

Tencent · General-purpose agents
Model post-training

General-purpose agents &
Embodied intelligence

ZeroSearch Co-corresponding author RynnBrain One of the project leads

Explore selected work GitHub

GENERALIST AGENTSFIG. 01

Reasoning, acting, and learning

NEWS / UPDATES

News

September 2026
Three papers accepted at NeurIPS 2026.
Papers and acceptance details
- WorldVLA: A Unified Vision-Language-Action and World ModelNeurIPS 2026 · Poster
- Benchmarking Fine-Grained Spatio-Temporal Awareness in Embodied Brain ModelsNeurIPS 2026 · Evaluations and Datasets Track · Poster
- InterLV-Search: Benchmarking Interleaved Multimodal Agentic SearchNeurIPS 2026 · Evaluations and Datasets Track · Poster

01 / RESEARCH

Research directions

General-purpose agents

I study how to train agents to search, use tools, and reason through multi-step tasks, using reinforcement learning and feedback from interactions.

Selected work: ZeroSearch

Reasoning · Tool Use · Learning

II.

Embodied intelligence

My work on embodied intelligence covers perception, spatial reasoning, vision-language-action models, and world models.

Selected work: RynnBrain · WorldVLA · RynnVLA-001 / 002

Perception · Reasoning · Action

02 / FEATURED WORK

Selected work

GitHub

PHYSICAL INTELLIGENCE · EMBODIED FOUNDATION MODELS

RynnBrain

One of the project leads

An open-source embodied foundation model for egocentric perception, spatiotemporal grounding, physical reasoning, and planning. Contributions include pre-training, Chain-of-Point reasoning, VLA, and training infrastructure.

Spatial grounding and embodied reasoning
Pre-training and post-training

Contribution statement ↗Subsequent collaborative work: RynnBrain 1.1

Repository Paper

Official RynnBrain architecture: images, video, and instructions pass through a visual encoder and a Dense/MoE decoder to produce spatial grounding, reasoning, and action plans. — RynnBrain 1.0 · Model architectureView original image Official source

GENERALIST AGENTS · SEARCH & TOOL USE

ZeroSearch

Co-corresponding author

Trains search agents through simulated retrieval interactions and reinforcement learning.

Simulated retrieval and curriculum training
Search policies and reinforcement learning

Repository Paper

Official ZeroSearch method: interactions between a simulated retrieval model and a policy model, with PPO and GRPO reward and advantage computation. — ZeroSearch · PPO / GRPO Training frameworkView original image Official source

PHYSICAL INTELLIGENCE · WORLD MODELING

WorldVLA

Co-author

Jointly models robot actions and future visual states autoregressively, using action attention masking to reduce error accumulation.

Research progressionWorldVLA precedes RynnVLA-002. The original repository now redirects to the successor project, which shares the code entry point.

Paper Shared successor code Model

PHYSICAL INTELLIGENCE · VISION–LANGUAGE–ACTION

RynnVLA series

Co-author

RynnVLA-001 ICRA 2026

Uses first-person human manipulation videos for pre-training, with trajectory-aware modeling and ActionVAE to transfer demonstrations to robot control.

Paper Code Project overview

RynnVLA-002 Successor to WorldVLA

Combines a continuous Action Transformer, wrist-camera inputs, and robot states to jointly model actions and future visual states.

Paper Code Model

03 / ABOUT

About

Jiayan Guo

I work on post-training for general-purpose agents at Tencent. I focus on how agents organize tasks, use tools, and carry out multi-step workflows.

Previously, I worked on embodied intelligence at Alibaba DAMO Academy. I led the development of ZeroSearch, a training framework for search agents. I was also one of the project leads for RynnBrain, where I was responsible for model pre-training and post-training. My research covers search agents, multimodal learning, and embodied intelligence.

WSDM 2023 Best Paper Honorable Mention; National Scholarship for Doctoral Students; Beijing Outstanding Graduate. View papers ↘

Google Scholar GitHub / SpaceLearner

ADDITIONAL RESEARCH

Other research projects

Graph learning and recommender systems.

EARLIER WORK · GRAPH LEARNING & RECOMMENDATION

JGCF

Graph collaborative filtering with Jacobi polynomials to model signals between users and items.

Repository Paper

04 / PUBLICATIONS

Papers & Publications

Google Scholar

Published papers and preprints, organized by year27 records

2026

9 papers

arXiv · PreprintSELECTED

RynnBrain: Open Embodied Foundation Models

View authors

Ronghao Dang, Jiayan Guo, Bohan Hou, Sicong Leng, Kehan Li, Xin Li, Jiangpin Liu, Yunxuan Mao, Zhikai Wang, Yuqian Yuan, Minghao Zhu, Xiao Lin, Yang Bai, Qian Jiang, Yaxi Zhao, Minghua Zeng, Junlong Gao, Yuming Jiang, Jun Cen, Siteng Huang, Liuyi Wang, Wenqiao Zhang, Chengju Liu, Jianfei Yang, Shijian Lu, Deli Zhao

Embodied foundation models for the physical world.

Jiayan Guo · One of the project leads

arXiv Code

NeurIPS 2026 · Evaluations and Datasets Track · Poster