Heng Zhou
NeoteAI
I am a research intern at NeoteAI, working on the N0 series of tactile-centric foundation models for robot manipulation.
Before that I spent a year and a quarter at the Shanghai Artificial Intelligence Laboratory with Lei Bai, working on multi-agent systems, agentic reinforcement learning, and embodied spatial intelligence.
Embodied Foundation Models · World Models · Agents
Education
Experience
Honours & Awards
News
Selected Publications view all →
Building agents that reason in language, coordinate as teams, and act in the physical world.
Agents that Reason
Search, self-organize, and learn to think. 3
LiveSearchBench: An Automatically Constructed Benchmark for Retrieval and Reasoning over Dynamic Knowledge
preprint
An automatically constructed benchmark that mines deltas between successive Wikidata snapshots into SPARQL-verified questions at three reasoning depths, exposing how sharply models fail on post-pretraining facts.
ReSo: A Reward-driven Self-organizing LLM-based Multi-Agent System for Reasoning Tasks
EMNLP 2025 main Oral paper, SAC Highlight Award, (Top 1%)
A reward-driven multi-agent framework that pairs task-graph generation with two-stage agent selection guided by a Collaborative Reward Model, plus an annotation-free pipeline for synthesizing multi-agent benchmarks.
Agents that Collaborate
From one policy to many bodies in one world. 3CoEnv: Driving Embodied Multi-Agent Collaboration via Compositional Environment
preprint
A compositional environment framework that pairs real-to-sim scene reconstruction with VLM-driven action synthesis and collision-checked sim-to-real transfer for safe multi-arm robotic collaboration.
Ego to World: Collaborative Spatial Reasoning in Embodied Systems via Reinforcement Learning
preprint
A cross-view benchmark plus a two-stage supervised-then-RL framework whose Cross-View Spatial Reward ties reasoning steps to visual evidence, fusing ego-centric observations into world-centric scene understanding.
VIKI-R: Coordinating Embodied Multi-Agent Cooperation via Reinforcement Learning
Annual Conference on Neural Information Processing Systems (NeurIPS) 2025
A hierarchical benchmark for embodied multi-agent cooperation spanning agent activation, task planning, and trajectory perception, paired with a chain-of-thought fine-tuning plus multi-level reinforcement learning framework.
Agents that Act
Grounding intelligence in perception and contact. 2N0-TWAM: Scaling Tactile-Native World-Action Model for Contact-Rich Manipulation
Technical Report
A tactile-native world-action model that predicts future vision, contact, and action using a unified force-based tactile representation and an asymmetric Mixture-of-Transformers for real-time manipulation.
N0-VTLA: Scaling Vision-Tactile-Language-Action Model with Latent Tactile Tokens
Technical Report
A vision-tactile-language-action foundation model pretrained on large-scale visuo-tactile robot data, adding a predictive tactile pathway and advantage-conditioned offline reinforcement learning for contact-rich manipulation.