Reinforcement Learning
Efficient decision-making
Offline RL, meta-RL, diffusion planning, policy transfer, multi-task learning, and compact symbolic policies.
Institute of Computing Technology · Chinese Academy of Sciences
Jiaming
Guo
Researching reinforcement learning, embodied intelligence, and large language models — with an emphasis on efficient decision-making, generalizable agents, and systems that connect reasoning with action.
01ABOUT
About
Learning systems that
reason, decide, and act.
I am a researcher at the Institute of Computing Technology, Chinese Academy of Sciences. My work studies how learning-based agents can make decisions more efficiently, transfer across tasks and environments, and connect high-level reasoning with low-level interaction.
My recent research spans reinforcement learning and meta-reinforcement learning, diffusion-based planning, embodied agents, vision-language-action systems, and large-language-model agents.
Ph.D.Institute of Computing Technology, CAS
Bachelor'sTsinghua University
02RESEARCH
Research directions
Three threads,
one goal.
Build agents that can learn reusable structure, reason over long horizons, and remain practical to deploy.
Embodied Intelligence
Generalizable embodied agents
Reusable motion primitives, VLA models, vision-and-language navigation, open-ended agents, and autonomous verification.
Large Language Models
Reasoning grounded in action
LLM agents, grounded skill learning, code-driven planning, long-context RL, and evaluation of model capabilities.
03PUBLICATIONS
Selected publications
Research,
at a glance.
Efficient Diffusion Planning with Temporal Diffusion
Temporal Diffusion Planner reuses and progressively updates plans to substantially improve decision frequency while preserving planning performance.
DASA: Distribution-Aware Sparse Attention for Accelerating Diffusion Transformer
Distribution-aware sparse attention and algorithm-hardware co-design for efficient Diffusion Transformer inference.
QiMeng-CRUX: Narrowing the Gap Between Natural Language and Verilog via Core Refined Understanding eXpression
A structured CRUX intermediate space and two-stage training framework for more precise natural-language-to-Verilog generation.
QiMeng-SALV: Signal-Aware Learning for Verilog Code Generation
Signal-aware verification, code extraction, and DPO turn partially correct Verilog generations into fine-grained functional learning signals.
Policy Constraint by Only Support Constraint for Offline Reinforcement Learning
Only Support Constraint (OSC) restricts the learned policy to behavior-policy support while avoiding unnecessary constraints within the support.
Ex³: Automatic Novel Writing by Extracting, Excelsior and Expanding
ACL 2024 Long Paper. † Equal contribution; * Corresponding author.
Efficient Symbolic Policy Learning with Differentiable Symbolic Expression
End-to-end differentiable symbolic policies for efficient, compact, and interpretable reinforcement learning.
Hindsight Value Function for Variance Reduction in Stochastic Dynamic Environment
A hindsight-conditioned value baseline that leverages future information to reduce policy-gradient variance in stochastic environments.
04CONTACT