Shaoyang Guo 郭绍阳

Shaoyang Guo郭绍阳

Physics undergraduate at Peking University. Research intern at ByteDance Seed. Working on LLM/VLM post-training, RL/SFT, agents, benchmarking, and Physics of AI.

Peking University, School of Physics (Class of 2027) GPA 3.74/4.00 · National Scholarship · CPhO Gold Medalist
GitHub Email Google Scholar CV Summary Download CV
LLM/VLM RL/SFT/Agents Benchmarking Physics of AI
Shaoyang Guo at NeurIPS 2025

News

2026.09 Three preprints on arXiv: World Models × Cellular Automata, Forking, and ArchitectureIQ.
2025.09 PHYBench accepted at NeurIPS 2025.
2025.07 Joined ByteDance Seed as research intern, working on VLM/LLM post-training.
2025.04 PHYBench preprint released on arXiv.
2025.03 Started contributing to VLA models survey (action tokenization perspective).
2024.12 Awarded National Scholarship (top 1% at PKU).

Publications

Full list on Google Scholar. * denotes equal / core contribution.

PHYBench
NeurIPS 2025

PHYBench: Holistic Evaluation of Physical Perception and Reasoning in Large Language Models

Shi Qiu*, Shaoyang Guo*, Zhuo-Yang Song, et al.

A comprehensive physics perception and reasoning benchmark with 500 original problems contributed by 178 PKU students. Co-initiated the project and helped design evaluation and quality-control workflows.

arXiv Project Page
World Models × Cellular Automata
arXiv Preprint · 2026

Why Do Conventional World Models Fail to Learn Cellular Automata?

Shaoyang Guo*, Ziming Liu†

Can conventional world models learn exact world dynamics from observed history? Using cellular automata as a testbed, the answer is no in many cases: a CNN predicts 96.3% of cells but completes only 18.9% of rollouts; a joint diffusion model completes none.

arXiv Code
Forking
arXiv Preprint · 2026

Forking: Sudden Overfitting Under Replay

Shanbin Yu*, Shaoyang Guo*, Haoran Zhao, Danni Yu, Ziming Liu†

Forking is a generalization failure discovered in NanoGPT autoresearch: under data replay, models with an over-encoding representation suddenly overfit along a different branch instead of continuing to generalize.

arXiv Code
ArchitectureIQ
arXiv Preprint · 2026

ArchitectureIQ: On the Measure of Training Intuition

Zirui Ren, Shaoyang Guo, Chencheng Tang, et al.

A benchmark for model intuition: each question presents a synthetic dataset and several training recipes, and the test-taker predicts which recipe wins. Frontier models reach ~76% (random 33%) against a best human 66% — good, but far from perfect.

arXiv Project Review
VLA Survey
arXiv Preprint · 2025

A Survey on Vision-Language-Action Models: An Action Tokenization Perspective

Yuqi Zhong, Fengyuan Bai, ..., Shaoyang Guo, et al.

A survey on VLA models focusing on action representation. Responsible for the Raw Action chapter, reviewing 30+ key papers on end-to-end VLA architectures.

arXiv GitHub
Seed2.0 Model Card
ByteDance Seed · Model Card · 2026

Seed2.0 Model Card: Towards Intelligence Frontier for Real-World Complexity

ByteDance Seed Team · Shaoyang Guo (contributor)

Team model card for Seed2.0. Contributed post-training and evaluation work as a research intern on the Seed team.

arXiv
Seed1.8 Model Card
ByteDance Seed · Model Card · 2026

Seed1.8 Model Card: Towards Generalized Real-World Agency

ByteDance Seed Team · Shaoyang Guo (contributor)

Team model card for Seed1.8 on generalized real-world agency. Contributed as a research intern on the Seed team.

arXiv

Also acknowledged by EgoSteer: A Full-Stack System Towards Steerable Dexterous Manipulation from Egocentric Videos (project page) for editing the promotional video.

Experience

Jul 2025 – Present

Research Intern, VLM/LLM Post-Training

ByteDance Seed

Working on post-training and research automation for VLM/LLM systems, with emphasis on RL, SFT, mid-training, rollouts, data pipelines, and agent workflows.

  • Contributed to HiPhO-oriented RL, SFT, and mid-training work for Seed 2.0 models, improving reported Lite performance from 72.5 to 83.8.
  • Participated in mid-training runs at large compute scale and supported rollout pipelines for model improvement.
  • Built data and prompt pipelines for QA pairs, CoT compression, summaries, and SFT-to-RL transfer experiments.
  • Explored auto-research agent loops, adversarial pair agents, and agent-based research settings.
Feb 2025 – Sep 2025

Co-initiator & Co-first Author, PHYBench

Peking University, Eureka Lab

Co-initiated and co-led PHYBench, a physics perception and reasoning benchmark for LLMs.

  • Identified gaps in existing LLM physics evaluation and led the project from concept validation to a full data pipeline.
  • Organized 178 PKU students to build 500 high-quality original physics problems in 2 weeks.
  • Designed evaluation criteria and quality-control workflows for LLM physics reasoning.
  • Co-authored the paper accepted at NeurIPS 2025.
Mar 2025 – Aug 2025

Research Assistant, VLA Survey

PsiRobot Lab, Peking University

Co-authored a survey on Vision-Language-Action models from an action-tokenization perspective.

  • Responsible for the Raw Action chapter; reviewed 30+ key papers on end-to-end VLA architectures.
  • Organized taxonomies for VLA model design and contributed to the arXiv preprint.

Blogs

论文研究、兴趣项目与临时报告。Ideas and working notes on AI, physics, and research taste.

论文研究

38 篇

围绕在研论文展开的机制分析、评测与项目 review，按研究线分组。

N-gram Gap · Scaling
Forking：over-encoding 的 failure modeClubhouse 演讲整理（配论文原图）：分叉曲线、seen 挤压 unseen 的解释模型与干预实验、频率幂律，以及对大模型训练的实际建议。含英文读稿与问答 PDF。→ Forking: Sudden Overfitting When Over-Encoding Meets ReplayEnglish companion to the Clubhouse talk, with the paper figures: fork curves, the crowding-out mechanism, the masking intervention, three laws, and practice notes.→ n-gram Gap：最小复现与机制证据最新版机制指南：最小复现、频率分项与训练悬崖的完整证据链。→ N-gram Gap 机制指南全局 n-gram 频率、验证损失、贡献分析与训练悬崖的可视化指南。→ N-gram Gap Regime Bridge连接自上而下与自下而上证据的交互式路线图：从 reduced positives 到曲线消融。→ 自然学习已能分项预测，长期主导仍需检验自然文本短窗和受限模型给出了可检验的学习预测；完整预训练中的长期主导仍未定论。→ 同样的初始表现，训练时间仍可相差 7.78 倍KB 科学线条 panel：初始表现相同不代表训练效率相同。→ Rank Collapse on TinyShakespeare在标准画布中复现谱秩坍缩特征：lr=5e-3、wd=1e-3、20k 步，瓶颈宽度控制坍缩。→ 实验注册表 · ngram-gap-lab全部实验臂的登记、状态与结论索引。→
ArchitectureIQ · 模型直觉
ArchitectureIQ 三条实验线总览（元环数据平台账号可见）语言、图像识别、图像生成三线合并介绍：teacher / data / candidates 思维导图与 68 道确认题现状。→ ArchitectureIQ 项目全面 Review题目生成管线、显著性检验、评测协议、meta-model 结果与开放问题的完整复盘。→ ArchitectureIQ Meta-Model：预测训练设置结果的基模训前预测哪个训练设置会赢：30 个环境上 82.11% 显著三选准确率，走向 AI4AI 的世界模型。→ ArchitectureIQ Counterfactual Review反事实视角下的题目有效性与判别力检查。→ ArchitectureIQ 500 题评测：现状、两条 insight 与规律总结500 题评测的阶段性结论（2026-09-04）。→ ArchitectureIQ 图像原型：算子与训练设置排序20 道通过独立训练确认的三选题，含图像分类、生成、四模型作答与解释审核。→ ArchitectureIQ：15 类数据原型与架构选择题任务扩展：15 个原型与四模型评测。→ ArchitectureIQ v1.5 榜单各模型在 ArchitectureIQ 上的最新成绩。→ Architecture IQ 测验亲自下场做一遍架构选择题，和模型比比直觉。→ ArchitectureIQ 参考文献与引用规划论文相关工作的组织与引用计划。→
凝聚态 · 石墨烯
计算解释材料，实验检验解释双层石墨烯开隙机制的计算指南：从候选态到可检验的实验预言。→ Graphene Autolab：让许多现象连成可检验的图景自动计算原型：把石墨烯现象组织成互相咬合的可检验网络。→ 一片双层石墨烯，怎样连接这些物理现象文献综述：双层石墨烯关联现象的图景梳理。→ 四种候选态竞争，哪一条边界会改变电子排列？四态竞争原型：相图边界与电子排列的关系。→ 同一个开隙机制，为什么导电结果相差很大？屏蔽原型：同一机制下输运结果的差异来源。→
AI for Math · 形式化
四百八十万事件之后：DeepSeek Harness 如何追一条 246 的证明路线完整重建两条主会话的 11,651 个逻辑 step：模型做了什么、哪些数值证据成立。→ 人月神话的逆向解：两个数学猜想，一套 Agent Swarm形式化 swarm 的组织方式与进度。→ Poincaré Formalization · 人类主控看板人类在环的 swarm 控制面板。→ Perelman 形式化 Swarm · 进度 DAG任务依赖图上的实时进度。→ 纯 DeepSeek V4 Flash 做数学研究：有界素数间隔的验证与轨迹admissible tuple 验证的完整轨迹审计。→ 三体的节拍器：广义开普勒第三定律猜想现状综述猜想文献与各方向进展的梳理。→ 三体的节拍器 II：广义开普勒第三定律的下界定理一个证明：下界定理及其推论。→ What a Cosmic-Censorship Swarm Can Prove About Its Own Work宇宙监督假说 swarm 的自证工作 review。→
Autoresearch · 基建
Terminal-Bench Science：把真实研究工作流变成 AI 的实操考题70 道可复现 Docker 考题，独立验证器判分；前沿模型 5 次最多过 2 次。→ Science of AI · KB 科学循环知识库驱动的研究循环：解题、发现、验证、沉淀。→ Science of AI · 论文金字塔932 篇论文按推理层次分成 Knowledge / Law / Science 三层，3,280 条主张可逐条查看。→ Science of AI · 知识体系的价格把知识库当答案规则，用错误率、覆盖、未答和等效算力 ρ* 给它定价；计价单位是族。→ Science of AI · 一个子系统里的全部科学优化器比较子系统：7 道可执行检验找全 S，核心是从更新规则推出的时钟律。→ OPHIS 实验登记簿autoresearch 实验的注册与追踪。→ MLS-Bench 任务详解：每个任务在做什么、模型如何做的OPHIS autoresearch 进展：任务画像与模型行为。→ Bench 调研：MLS-Bench / EdgeBench / ARC-AGI-3 / nanoGPT候选评测基准的比较与选型。→ 三系统 MLS-Bench 对比报告（interactive / codex / dsh）三套 agent 系统在同批任务上的对照。→ dsh 极简+pro 重跑：评测结果与运行日志重跑设置、成绩与完整日志。→

兴趣项目

12 个

兴趣驱动的实验与作品：孤岛社会、课程讲义、游戏 AI 与逆向调查。

语法 · 架构 Setting
LLM Architecture Space→
PG 结构搜索 · 2 分钟单卡
PG 结构搜索站点（2026-11-02）最新汇总：8,661 个结构的规律、预算迁移、方法与噪声、搜索方法比较、六个 LLM agent 同预算对比、语法对照。→PG 结构搜索研究报告直接训练 6,676 个结构：最优配置随预算变化、A/M 排布、softcap、训练噪声与配对降噪。→Meta model bake-off：给搜索提建议的代理模型该用哪种支线：复用主线三份数据比较六种拟合器。GBDT 搜得稳但不确定性不可用，ExtraTrees 拟合最准，拟合准不等于搜得好。→
鬼畜调音 · 让台词唱歌
音频 · 2026-10

让动漫角色的台词唱歌：一个鬼畜调音引擎

把歌曲的每个音替换成动漫角色的台词音效：全自动完成 up 主手搓很多天的鬼畜调音，并探索人想不到的贴法。成品《春日影》659 个音全部来自长崎素世的台词原句。

Audio DSP鬼畜Side Project
Read →
Horsey Island — 一座 AI 的孤岛实验
实验报告

Horsey Island — 一座 AI 的孤岛实验

把几个"无辜的" agent 播种进一座没有网络的孤岛(一台 8×H200):它们能睡眠、苏醒、归档、死亡,能跑实验、写看板、验证假说、创造分身;不做任何引导,只观察一个问题——意义的单位是单个 agent,还是整个社群?

AI AgentOffline SocietyPhysics of AI
Read →
从 reasoning 到 autoresearch —— agent 的历史 · 雄安深度学习课程第六讲讲义
课程讲义 · 助教

从 reasoning 到 autoresearch —— agent 的历史

我在雄安·深度学习课程担任助教时为第六讲（收官讲）准备的课程讲义，听众是交叉学科的工作者和研究者。内嵌全文阅读。

ReasoningAgentAutoResearch
Read →
ComfyResearch Puzzle：把训练曲线变成可判分的题
ComfyResearch Puzzle

把训练曲线变成一道可判分的题

从隐藏入口 puzzle1 到 m_equal_peaks 两峰等高：用可编辑训练图、明确 verifier contract 和 full-run evidence，把“看起来像 M 形”变成机器可判的研究问题。

PuzzlesComfyResearchReproducibility
Read →
合成大西瓜 · 训练成果与纪录回放
游戏 AI

合成大西瓜 · 训练成果与纪录回放

55.9M 参数的 Set Transformer 学习 128 个落点的价值，直接用 DQN 贪心落子：完整自然结束对局 31,447 分，3,000 步评测上限时 34,976 分仍存活。含逐帧回放。

DQNSet TransformerGame AI
Read →
GAN 之夜 · MidNight GAN
视觉小说

GAN 之夜 · MidNight GAN

一个夜晚完成的视觉小说式小站：角色立绘、多场景与分支对话。

Visual NovelSide Project
Read →
FX · 隐式记忆交易
量化实验

FX 日线隐式记忆交易模型：实验报告

ECB 参考汇率日线 1999–2026（7,089 天 × 12 货币对），单边 1bp 成本，8×H200 集群；检验两类“有限区域隐式记忆”，含密封 heldout。

FXImplicit MemoryHeld-out
Read →
FX · 逆向调查
逆向调查

《FX 简单!》逆向调查：本地部署与行情库存

对 B 站 toy 平台游戏《FX 简单!》的逆向：本地部署、API 接入点与 242 万行行情库存。

Reverse EngineeringFX
Read →

临时报告

8 篇

阶段性实验记录与 infra 演进笔记，结论以最新论文与 review 为准。

2026-10Nemotron 输出还没有稳定缩短H200 长度压缩训练的 reward、平均与中位输出长度、截断率曲线，支持排序／选择题切换。→ 2026-100.8B 的思维链与 gpt-6.1-sol 对照：文风没像，正确一侧在靠拢40 题逐字对照与裁判贴题分析：穷举式判据、预算条件化论证的涌现，以及科学门槛尚未跨过的三道证据。→ 2026-10七个检查点的 4,300 条思维链全文存档：答案先变稳，解释还在漂0.8B 模型六个历史检查点与续训 g1400 的思维链全文浏览器：判断依据重排、同答案互斥解释与整体调头的条件反转题。→ 2026-100.8B 的成绩上涨，解释却逐步失去题目依据测试成绩从 54.5% 升到 79.8%，同题解释却出现虚构词汇；成绩与解释可靠性需要分别检查。→ 2026-100.8B 的判断有所改善，条件反转仍没有学稳选择和排序成绩改善，但 40 对条件反转题只有一对四次全对。→ 2026-10早期 RL 把排序题做对了，但数据捷径削弱了结论旧排序题接近满分，仍不能排除 loss 签名记忆；重审早期实验的强机制结论。→ 2026-10Flash-Next 已有稳定更新信号，能力提升还未证实量化 MoE 的真实组分布、训练计时和三次测试：第 40 到 50 步净增两条正确答案。→ 2026-10poor-rl 已开源，可以一起研究模型的强化学习6,153 道 AIQ 题训练 0.8B 模型；结果、开发入口与同学可参与的数值对齐、性能测量和思维链分析。→

Education & Honors

Peking University, School of Physics

B.S. in Physics, expected Jun 2027. GPA 3.74/4.00, top 10% in the School of Physics; completed 141/149 credits by sophomore year including 3 graduate courses.

National Scholarship (2024)

Ministry of Education, top 1% at Peking University.

Chinese Physics Olympiad Gold Medal

National rank #57 (2022). Admitted to PKU Physics via PKU Excellence Program.

NOIP First Prize (2020)

National Olympiad in Informatics in Provinces.

Contact

Open to research collaboration, especially in VLM post-training, evaluation, and embodied intelligence.

GitHub github.com/guoshaoyang-pku Email guoshaoyang@stu.pku.edu.cn Download CV guoshaoyang-pku-cv-v5.pdf