Welcome to the TMGNLP group! We are part of the TMG Group at the Harbin Institute of Technology, Shenzhen. TMGNLP focuses on developing advanced agentic systems powered by large language models, with an emphasis on three core research directions:
-
Large-scale Model-based Reinforcement Learning: Building scalable RL frameworks that support massive environment interactions (Envs), multi-turn actions, and multi-tool model invocations. Our goal is to enable efficient training and iteration of large models through hundreds of tool extensions and thousands of concurrent environment interactions.
-
Model Architecture and Inference Optimization: Addressing challenges such as high latency, low throughput, and excessive memory usage in large model inference. We explore sparsity-based acceleration techniques, testing-time scaling architectures, and memory-enhanced model designs tailored for intelligent agents.
-
Continual Learning and Memory Management: Developing hierarchical memory systems for general-purpose agents using graph-based enhancements. By leveraging graph storage and analytics, we aim to improve reasoning, memory management, and tool utilization in agents, while mitigating catastrophic forgetting and supporting lifelong learning.
These efforts integrate cutting-edge techniques such as LLM-based tool calling, code generation, memory-augmented architectures, and interactive reasoning systems to advance the capabilities of language model agents in complex, dynamic environments.
