Flagship project: Qwen-style GRPO post-training and official-style evaluation for scientific Python-answer tasks, with a clear negative result around reward/eval mismatch
旗舰项目：面向 scientific Python-answer 任务的 Qwen-style GRPO 后训练与 official-style evaluation，并记录了 reward/eval mismatch 的负结果
A local imperfect-information game lab where one human plays no-limit Texas Hold’em against multiple LLM agents, with visible-state prompts, logs, table talk, fallback actions, and replay
本地不完美信息博弈实验台：真人与多个 LLM agent 对战无限注德州扑克，包含可见状态 prompt、日志、牌桌话、fallback 行为和复盘
A public historical prototype for an LLM-driven narrative RPG with multi-character dialogue, relationship state, event bus logging, APIs, tests, and lightweight saves
公开历史原型：LLM 驱动的轻叙事 RPG，包含多角色对话、关系状态、事件总线日志、API、测试和轻量存档
Private prototype, public summary only: a multi-agent school-world simulation with personas, memory, relationships, dialogue, locations, and replayable JSONL events
私有原型，仅公开高层总结：多智能体校园世界仿真，包含人设、记忆、关系、对话、位置和可回放 JSONL 事件
Coursework / systems project: a C++ Mini-Lisp interpreter with tokenizer/parser, evaluator, REPL, scripts, special forms, macros, builtins, and CMake builds
课程 / 系统项目：C++ Mini-Lisp 解释器，覆盖词法/语法解析、求值、REPL、脚本、特殊形式、宏、内建过程和 CMake 构建