PhD candidate @ Nankai University, College of Computer Science | Research on retrieval-augmented LLMs, knowledge-conflict control, long-context memory, and the agent-harness engineering substrate that makes LLM agents finish long-horizon tasks reliably.
| Area | Project | Stars | Notes |
|---|---|---|---|
| Multimodal parametric RAG | SCoRAG | ★ 11 | A multimodal parametric RAG framework that compiles retrieved evidence into modality-specific LoRA adapters and routes them to compatible module slots. |
| RAG / knowledge-conflict control | Swin-VIB | ★ 46 | Source code for the AAAI 2026 paper Accommodate Knowledge Conflicts in Retrieval-augmented LLMs: Towards Robust Response Generation in the Wild. |
| Multi-view representation | DistilMVC | ★ 69 | Source code for the TNNLS 2024 paper Towards Generalized Multi-stage Clustering: Multi-view Self-distillation. |
| Research feed | MyArxiv | ★ 0 | Personal arXiv tracking and digest setup. |
| Project | Stars | PR | What I Did |
|---|---|---|---|
| openclaw | ★ 386k | #78288 | Show target node name in exec-tool transparency messages so multi-agent traces stay readable when several agents share an exec channel. |
| openclaw | ★ 386k | #113560 | Prevent same-named generated files from overwriting earlier SharePoint uploads so Teams shows the correct current file and retains previous content. |
| letta-ai/letta-agent-sdk | ★ 81 | #249 | Normalize chronological cursors for descending conversation-message pagination, preventing overlapping pages and duplicate messages. |
| letta-ai/letta-agent-sdk | ★ 81 | #250 | Decode file URLs before spawning the MCP test fixture, fixing failures in checkout paths containing spaces or non-ASCII characters. |
| copilot-money-mcp | ★ 69 | #619 | Make privacy comment stripping syntax-aware so comment-like delimiters inside strings, templates, and regular expressions are preserved. |
Retrieval-augmented generation under context distortion · knowledge-conflict control · learnable long-context compression · agent context-management policy evaluation.
南开大学计算机学院在读博士 | 研究方向：检索增强大模型、知识冲突控制、长上下文记忆，以及让 LLM Agent 稳定完成长跨度任务的 agent harness 工程底座。
| 方向 | 项目 | Stars | 简介 |
|---|---|---|---|
| 多模态参数化 RAG | SCoRAG | ★ 11 | 将检索证据编译为模态专用 LoRA 适配器，并通过槽位路由将其注入兼容模块的多模态参数化 RAG 框架。 |
| RAG / 知识冲突控制 | Swin-VIB | ★ 46 | AAAI 2026 论文《Accommodate Knowledge Conflicts in Retrieval-augmented LLMs: Towards Robust Response Generation in the Wild》源码实现。 |
| 多视图表征 | DistilMVC | ★ 69 | TNNLS 2024 论文《Towards Generalized Multi-stage Clustering: Multi-view Self-distillation》源码实现。 |
| 研究追踪 | MyArxiv | ★ 0 | 个人 arXiv 跟踪与摘要工具。 |
| 项目 | Stars | PR | 修了啥 |
|---|---|---|---|
| openclaw | ★ 386k | #78288 | 在 exec 工具的透传消息中展示目标节点名，让多个 agent 共享 exec 通道时的执行轨迹依然可读。 |
| openclaw | ★ 386k | #113560 | 避免同名生成文件覆盖已有的 SharePoint 上传，使 Teams 显示正确的当前文件并保留先前内容。 |
| letta-ai/letta-agent-sdk | ★ 81 | #249 | 规范降序消息分页的时间游标，避免翻页时出现页面重叠和重复消息。 |
| letta-ai/letta-agent-sdk | ★ 81 | #250 | 启动 MCP 测试夹具前正确解码文件 URL，修复检出路径包含空格或非 ASCII 字符时的失败。 |
| copilot-money-mcp | ★ 69 | #619 | 让隐私扫描中的注释移除具备语法感知能力，保留字符串、模板和正则表达式中的类注释分隔符。 |
检索增强生成下的上下文失真 · 知识冲突控制 · 可学习的长上下文压缩 · Agent 上下文管理策略评测。