Nanjing Univ Proposes HCL: Continual Learning for Agents Without Model Tuning
机器之心 · wechat · 2026-08-28
Nanjing University and UOW propose Harness Continual Learning (HCL), shifting the object of continual learning from model parameters to the Agent's Harness (prompts, memory, skills, tools, etc.). With frozen model weights, the system evolves via a Proposal-Evaluation-Commitment loop. Experiments show significant performance gains in ALFWorld, Minecraft, and text reasoning tasks (e.g., GSM8K accuracy rose from 49.4% to 92%) while effectively mitigating forgetting.
More from coding & agent
- Nanjing Univ. Releases Procedura: Agentic 3D Modeling with Procedural Control — nanjinguniv · 2026-08-28
- Agent tool calls fail silently with no traces, breaking production pipelines — Icy-Weakness8310 · 2026-08-28
- Helm MCP: Give AI assistants access to real Helm chart data, stop hallucinations — modelcontextprotocol · 2026-08-28
- SymPy Sandbox MCP: Secure symbolic math computation for LLMs via SymPy — modelcontextprotocol · 2026-08-28
- Survey on LLM Agent Evaluation: Taxonomy and Enterprise Challenges — kalyan_kpl · 2026-08-28
- Building a Sci-Fi Movie RAG Agent in ~60 Lines of TypeScript — mastra_ai · 2026-08-28