ICTNLP/UMA-4B
ICTNLP/UMA-4B is the 4 billion parameter Generalist checkpoint of the Unified Memory Agent (UMA) developed by ICTNLP. This model is designed for long-context reasoning by incrementally maintaining a compact core summary and a structured key-value Memory Bank. It excels at proactive structured memory construction and downstream question answering through explicit memory and retrieval operations, built upon the Qwen3-4B-Instruct-2507 base model.
Loading preview...
UMA-4B: Generalist Unified Memory Agent
UMA-4B is the Generalist checkpoint of the Unified Memory Agent (UMA) framework, developed by ICTNLP. This 4 billion parameter model is built on Qwen/Qwen3-4B-Instruct-2507 and is specifically designed for long-context reasoning by managing a dynamic, structured memory. It was trained using end-to-end reinforcement learning with Task-Stratified GRPO on datasets like HotpotQA and the Mem-alpha corpus.
Key Capabilities
- Proactive Structured Memory: Incrementally builds and maintains a compact core summary and a key-value Memory Bank.
- Tool-Using Agent: Performs memory construction and question answering through explicit memory and retrieval operations.
- Long-Context Reasoning: Addresses challenges in extended conversational or document contexts by efficiently managing information.
- Research Focus: Intended for research into streaming memory agents, memory maintenance with explicit tool calls, and evaluation of the UMA framework.
Intended Use Cases
- Research on long-context and streaming memory agents.
- Development of systems requiring proactive structured memory construction.
- Applications involving downstream question answering over reusable memory.
- Evaluation and extension of the UMA framework.
It's important to note that UMA-4B is a research model and its agent behavior depends on the UMA prompt templates, tool implementations, and retrieval backend. For full memory-agent inference, users should refer to the official repository.