能力Expertise
分两条轴呈现：已验证的能力（有论文/落地产出，证明研究实力）与兴趣与探索方向（我想深耕、也在寻找合作与机会的方向）。等级依据文末的评定标准给出，尽量有据可依。Two axes: track record (published/shipped output that proves capability) and interests I'm pursuing (where I want to go deep and find collaboration). Levels follow the rubric below.
✅ 已验证的能力✅ Track record
一作提出 LAMIC（AAAI 2026），并在 Onestory 实习中落地进核心产品；目前在可灵从事视频生成方向。微调过 T2I/I2I 生成模型（LoRA / Flux-Kontext / EasyControl）并上线。First-authored LAMIC (AAAI 2026), shipped into Onestory products; now working on video generation at Kling. Fine-tuned and deployed T2I/I2I models (LoRA / Flux-Kontext / EasyControl).
在 LAMIC / TAG-WM 中深度使用多模态扩散 Transformer 与扩散反演；熟悉 DDPM/DDIM、Flow Matching 等采样与训练机制。Heavily used multimodal diffusion transformers and diffusion inversion in LAMIC / TAG-WM; familiar with DDPM/DDIM and flow matching.
一作提出 TAG-WM（ICCV 2025）与 Flow of Truth（面向图像到视频的时序取证），研究生成图像/视频的水印、版权保护与篡改定位。First-authored TAG-WM (ICCV 2025) and Flow of Truth (temporal forensics for I2V).
🧭 兴趣与探索方向🧭 Interests I'm pursuing
我最想深耕、也在寻找合作与机会的方向。Where I want to go deep and am looking for collaboration & opportunities.
多模态生成Multimodal Generation
让 AI 跨模态地生成内容，从当前的图像/视频走向更统一的多模态生成——这是我最想深耕的方向。AI that generates across modalities, moving from image/video toward more unified multimodal generation — the direction I most want to pursue.
多模态理解Multimodal Understanding
让 AI 跨模态地理解世界；在可灵接触统一「理解-生成」，持续跟进 MLLM / VLM 与视觉 tokenizer。AI that understands the world across modalities; explored unified understanding-and-generation at Kling, tracking MLLM / VLM and visual tokenizers.
机器人 · 具身智能Robotics · Embodied AI
系统性阅读 VLA、世界-动作模型、模仿学习（占我近期阅读很大比重），正从阅读走向实践。Systematic reading on VLA, world-action models and imitation learning (a large part of my reading base); moving toward practice.
仿真 · 游戏（可交互世界）Simulation · Games (interactive worlds)
关注可交互世界、游戏环境与 sim-to-real，作为让 AI 生成并栖居于世界的试验场。Interested in interactive worlds, game environments and sim-to-real as testbeds for AI that generates and inhabits worlds.
强化学习与决策Reinforcement Learning & Decision Making
关注 RL / RLHF 与决策智能，作为连接「生成」与「交互」的桥梁。Following RL / RLHF and decision making as the bridge between generation and interaction.
📏 评定标准📏 How levels are decided
「已验证的能力」等级并非凭感觉，而是综合：① 论文产出（是否一作/顶会） ② 工程落地（是否上线/产品化） ③ 复现与实践经验 ④ 文献阅读广度。Track-record levels are not vibes — they combine: (1) publications (first-author / top venue), (2) shipped/productionized work, (3) reproduction & hands-on practice, and (4) reading breadth.
有一作论文 / 落地产出，能独立从想法到实现完成完整研究或工程。First-author papers or shipped systems; can drive a project end-to-end.
有较多复现与实践经验，掌握主流方法，能快速上手推进。Solid hands-on and reproduction experience; can ramp up quickly.
系统性阅读 + 初步实验/复现，正在积累。Systematic reading plus early experiments; actively building up.
广泛涉猎，理解核心概念与前沿脉络。Broad exposure; understand core concepts and the research landscape.