Junyang Wang (王俊阳)

Junyang Wang (王俊阳)

Peking University CS undergraduate
Research: embodied AI, multimodal agents, post-training

Home 主页

Peking University CS · embodied AI and agents

Junyang Wang / 王俊阳

I am a Computer Science undergraduate at Peking University. My current thread is post-training and evaluation for embodied and multimodal models: turning failure cases into data, rewards, eval scripts, and the next training signal.

I also like making agent ideas runnable. Recent public projects include CritPT-RL, a post-training/eval experiment for scientific Python-answer tasks, and Texas-Poker-Agents, a local table where a human plays with multiple LLM agents under imperfect information.

Selected Work Experience GitHub Email
Junyang Wang profile photo
Main Thread Embodied/multimodal post-training, reward design, benchmark-style evaluation, and failure feedback loops.
Public Evidence CritPT-RL documents a working pipeline and a useful negative result: cleaner format did not automatically improve official accuracy.
Builder Habit Agent systems should run, log decisions, expose visible information boundaries, and leave traces that can be replayed.

Selected Work

CritPT-RL Qwen-style GRPO post-training lab for scientific Python-answer tasks. The most useful result is a diagnosis of reward/eval mismatch. Texas-Poker-Agents A local no-limit Hold'em table for one human and multiple LLM seats, with logs, table talk, fallback actions, and replayable hands. Lyuyuan AI An LLM-driven narrative RPG prototype with character prompts, relationship state, event bus logging, and lightweight saves.

Research Snapshot

Post-training for embodied models

I work on the experimental loop after a multimodal embodied model fails: cleaning multi-source robot data, constructing SFT/CoT samples, filtering RL candidates, maintaining evaluation scripts, and feeding error cases back into the next round.

Center for Frontier Computing Research, Peking University

Video spatio-temporal grounding

In a research rotation, I studied a multi-agent debating framework for video grounding and helped with localization experiments, result logging, and error-case analysis.

Wangxuan Institute of Computer Technology, Peking University

Working Notes

For now, my most concrete notes live inside project READMEs and experiment logs. The notes I want to turn into public writeups are the failure analysis behind CritPT-RL, visible-information boundaries for poker agents, and small implementation notes from Mini-Lisp.

Contact

北京大学计算机 · 具身智能与 Agent

王俊阳 / Junyang Wang

我是北京大学计算机本科生，主要关注具身/多模态模型的后训练与评测：怎么把模型失败案例变成数据、reward、评测脚本和下一轮训练信号。相比只看最终 benchmark 分数，我更关心中间过程是否可复现、可检查、可诊断。

我也喜欢把 agent 想法做成能运行的系统。最近的公开项目包括 CritPT-RL，一个 scientific Python-answer 任务上的 post-training/eval 实验；以及 Texas-Poker-Agents，一个真人与多个 LLM agent 同桌对局、记录日志并复盘行为的本地实验台。

代表项目 经历 GitHub 邮箱
王俊阳头像
当前主线 具身/多模态后训练、reward 设计、benchmark-style evaluation，以及失败案例回流。
公开证据 CritPT-RL 记录了一个跑通的训练评测流程，也记录了一个有价值的负结果：格式更干净不等于 official accuracy 更高。
工程习惯 Agent 系统应该能运行、能记录决策、能限制可见信息边界，也能留下可复盘的行为痕迹。

代表项目

CritPT-RL 面向 scientific Python-answer 任务的 Qwen-style GRPO 后训练实验。最重要的结论是 reward 与目标评测之间的语义错位。 Texas-Poker-Agents 本地无限注德州扑克实验台：真人对战多个 LLM 座位，记录日志、牌桌话、fallback 行为和完整复盘。 Lyuyuan AI LLM 叙事 RPG 原型，包含角色 prompt、关系状态、事件总线日志和轻量存档。

科研切片

具身模型 post-training

我参与模型失败之后的实验闭环：多源机器人数据清洗、SFT/CoT 样本构造、RL 候选样本筛选、评测脚本维护，以及错误案例回流。

北京大学前沿计算研究中心

视频时空定位

在科研轮转中，我研究过 multi-agent debating 框架下的视频 grounding，并参与定位实验、结果记录和错误案例分析。

北京大学王选计算机研究所

技术笔记

目前最具体的笔记还在项目 README 和实验记录里。接下来最值得公开整理的是 CritPT-RL 的失败诊断、poker agent 的可见信息边界，以及 Mini-Lisp 里的 macro / quasiquote 实现笔记。

联系