01Act
Agent systems
Agents that plan, use tools, and improve through interaction instead of staying fixed after deployment.
Research at SJTU-MARL · Shanghai
Agents that remember,
adapt, and improve
from experience.
I am Shengtao Zhang. I study self-evolving agents at the intersection of reinforcement learning, memory, and efficient post-training. My research explores how agents learn from interaction and develop reusable capabilities through distillation and training.
SJTU-MARLAgent · RL · Memory
Current focus
How can agents turn experience into reliable, transferable skills?
Research thesis
From experience to capability.
My work connects three layers of adaptive intelligence: acting in an environment, learning from feedback, and preserving useful experience.
02Adapt
RL & post-training
Reinforcement learning, distillation, and efficient post-training for agents that learn from experience.
03Remember
Memory
Episodic and value-aware memory that retrieves experience for utility, provenance, and continual refinement.
Featured work
Memory and agent learning.
TRACE
Temporal memory admission helps returning agents retain valid information and reject stale state as multi-agent collaborations evolve.
When teammates change shared plans, a returning agent must preserve valid preferences while updating outdated bookings and unresolved tasks.
Spatial Memory Agent
Experience-grounded procedural memory lets frozen vision-language agents improve spatial reasoning without parameter updates or expert tools at deployment.
SMA turns verifier-scored spatial experience into transferable memory cards, then retrieves reliable procedures through semantic filtering and TRS-aware ranking while keeping the VLM frozen.
MemRL
A non-parametric route to runtime self-improvement: keep the reasoner stable, let episodic memory learn utility from feedback.
MemRL first recalls similar episodes, then selects them by learned Q-values. The frozen LLM acts with the retrieved context, while environmental rewards update each memory item's utility.
MemQ
Credit assignment for agent memory, propagated through the provenance chains that make later memories possible.
Each interaction retrieves memories, creates a new episode, and records its provenance. TD feedback then travels backward through the DAG to update the memories that enabled success.
EvoKernel
Value-guided experience reuse turns sparse NPU feedback into a continual drafting-and-refinement loop.
Cold-start drafting retrieves transferable experience for an initial kernel. Verification rewards update the shared memory, which later reuses successful traces for continual latency refinement.
Selected publications
Research index.
Work on self-evolving agents, reinforcement learning, and memory, alongside graph learning, language, and multimodal reasoning.
Google Scholar146 citationsh-index 3i10-index 1Checked
TRACE: Governing Memory Validity in Evolving Multi-Agent Systems
Temporal memory admission helps returning agents retain valid information and reject stale state as multi-agent collaborations evolve.
Spatial Memory Agent: Experience-Grounded Procedural Memory for Spatial Intelligence
Experience-grounded procedural memory lets frozen vision-language agents improve spatial reasoning without parameter updates or expert tools at deployment.
MemRL: Self-Evolving Agents via Runtime Reinforcement Learning on Episodic Memory
A non-parametric route to runtime self-improvement: keep the reasoner stable, let episodic memory learn utility from feedback.
MemQ: Integrating Q-Learning into Self-Evolving Memory Agents over Provenance DAGs
Credit assignment for agent memory, propagated through the provenance chains that make later memories possible.
Towards Cold-Start Drafting and Continual Refining: A Value-Driven Memory Approach with Application to NPU Kernel Synthesis
Value-guided experience reuse turns sparse NPU feedback into a continual drafting-and-refinement loop.
A Chain-of-thought Reasoning Breast Ultrasound Dataset Covering All Histopathology Categories
A reasoning-rich breast ultrasound resource that connects images, pathology coverage, and clinically grounded chains of thought.
PTCL: Pseudo-Label Temporal Curriculum Learning for Label-Limited Dynamic Graph
Temporal curriculum learning makes final-timestamp labels useful across the full evolution of a dynamic graph.
Unveiling the Hidden: Movie Genre and User Bias in Spoiler Detection
Genre structure and user history expose spoiler patterns that text-only classifiers tend to miss.
Recent notes
News.
TRACE releasedMemory validity for agents returning to evolving multi-agent collaborations. Paper and code are available.↗Spatial Memory Agent releasedExperience-grounded procedural memory for parameter-update-free spatial self-evolution.↗MemQ releasedCredit assignment for self-evolving memory agents over provenance DAGs.↗EvoKernel accepted at ICML 2026Value-driven memory for cold-start NPU kernel synthesis and continual refinement.↗MemRL releasedRuntime reinforcement learning on episodic memory for self-evolving agents.↗Unveiling the Hidden at ECML PKDDGenre- and user-aware modeling for spoiler detection.↗
About
Building learning loops for agents that operate in the real world.
I work on agent learning and memory with SJTU-MARL. My interests span reinforcement learning, memory, knowledge distillation, and efficient post-training. I study how agents acquire, retain, and generalize useful skills from experience, connecting runtime adaptation with lasting improvements in model capabilities.
Previously, I studied Artificial Intelligence at Xi'an Jiaotong University and worked on dynamic graph learning, language understanding, and reliable decision systems.
Z. S. T.
Zeal stirs tides.「热望所至，潮涌不息。」