First-year master's student at Sichuan University, working on LLM agents — tool use, verification, and post-training robustness.
0 CCF-A papers so far — in the writing phase.
- Evidence-grounded agent harnesses — separating LLM semantic decisions from a deterministic Python runtime, so tool calls, approvals, checkpoints, and failures are observable and verifiable (job-agent-harness).
- Post-training robustness — a controlled study of how an agent's structured decisions change when its skill prompt is incomplete or missing (skill-annealing).
| Project | What it does |
|---|---|
| job-agent-harness | Evidence-grounded agent harness with deterministic runtime contracts for tool use, approval, checkpointing, and verification. |
| skill-annealing | Controlled post-training study of agent robustness under reduced skill prompts. |
- Languages: Python
- Areas: LLM agents · tool use · agent evaluation & verification · post-training
Reach me on GitHub.
Riceff，四川大学计算机学院硕士研究生，研究方向为 LLM Agent——工具调用、验证与后训练鲁棒性。
已发0篇CCFA，备文期。
- 基于证据的 Agent Harness —— 将 LLM 语义决策与确定性 Python 运行时分离，让工具调用、审批、检查点与失败都可观测、可验证（job-agent-harness）。
- 后训练鲁棒性 —— 受控实验：当 Agent 的 skill prompt 不完整或缺失时，其结构化决策会如何变化（skill-annealing）。
| 项目 | 做什么 |
|---|---|
| job-agent-harness | 基于证据的 Agent Harness，为工具调用、审批、检查点与验证提供确定性运行时契约。 |
| skill-annealing | 缩减技能提示下 Agent 鲁棒性的受控后训练研究。 |
- 语言：Python
- 方向：LLM Agent · 工具调用 · Agent 评测与验证 · 后训练
欢迎通过 GitHub 联系我。