GSA 数字员工GSA Digital Worker
工作项目 · 中国科学院北京基因组研究所At Beijing Institute of Genomics, CAS
面向科研数据审编与文献取证，独立搭建支持长任务执行、跨会话记忆与权限治理的 Agent 系统。An agent system I independently built for research-data curation and literature evidence gathering, with long-running tasks, cross-session memory and access controls.
以 SOP 状态机与 Agent Loop 协同执行，接入版本化 Skills、MCP 工具与生命周期 Hooks。在 60 个中断与超时用例中，任务恢复成功率达 95%。Combines an SOP state machine with an agent loop, versioned skills, MCP tools and lifecycle hooks. Task recovery succeeded in 95% of 60 interruption and timeout test cases.
Agent HarnessHooksSkills · MCPMemoryAccess Control
GSA 多跳检索与 Search Agent 后训练GSA Multi-hop Retrieval & Search Agent Post-training
工作项目 · 中国科学院北京基因组研究所At Beijing Institute of Genomics, CAS
围绕文献—数据集关系核验，结合混合检索、多跳证据验证与模型后训练，向数字员工提供可定位原文的证据包。Hybrid retrieval, multi-hop evidence verification and model post-training for literature–dataset relationship checks, returning evidence packages traceable to source passages.
构建 1,350 条逐跳证据 QA，并按论文与研究项目隔离数据。在固定 200 条领域测试题、相同工具与调用预算下，Qwen3-4B 经 LoRA SFT / GRPO 的答案正确率为 54% → 68% → 76%。Built 1,350 evidence-grounded QA examples, split by paper and research project. On a fixed 200-question domain test set with the same tools and call budget, LoRA SFT and GRPO improved Qwen3-4B answer accuracy from 54% → 68% → 76%.
Agentic RAGHybrid RetrievalQwen3-4BLoRA SFTGRPO · verl
EvoWork · 自进化 Agent HarnessEvoWork · An Evolving Agent Harness
个人项目 · 独立设计与实现Independent project · Design & implementation
从零实现 Runtime、Memory、Skills 自进化、Eval 与安全五个子系统，探索让 Agent 从任务失败中学习，并通过回归验证后集成技能。Built runtime, memory, skill evolution, evaluation and safety subsystems from scratch, exploring how agents can learn from failures and integrate skills after regression validation.
打通“轨迹分析 → 失败归因 → 技能提案 → 沙箱回归 → 确认集成”。经 6 轮迭代，固定 60 Case 回归集成功率从 52% 提升至 80%。Connects trace analysis, failure attribution, skill proposals, sandbox regression and confirmed integration. Across six iterations, success on a fixed 60-case regression suite rose from 52% to 80%.
RuntimeLayered MemoryContext EngineeringSkill EvolutionEvaluation