CMU Faculty Summary — PhD Advisor Research
CMU Faculty Summary — PhD Advisor Research
For: Weijia Zhang (Yale M.S. CS 2026–2028 → PhD 2028+) Target: CMU (LTI / MLD / CSD) Report date: 2026-06-12
Strategic Context
CMU is the primary PhD target alongside UIUC. The two-phase plan:
- Yale M.S. (2026–2028): Arman Cohan as advisor → publish NLP/agent papers → build connections
- CMU PhD (2028+): Apply with Yale publications + warm intros
CMU LTI and MLD are the strongest departments globally for NLP + agentic AI. Cold-emailing now (2026) builds name recognition before 2028 applications.
Weijia’s research direction: NLP agents, GUI agents (GUIAgentDebugger), SFT/RL post-training, VLM — targeting industry-research roles at frontier labs (OpenAI, Anthropic, DeepMind, FAIR, MSR).
Priority P1 — 强烈推荐陶瓷
| Professor | Dept | Position | Research Focus | Why P1 | Profile |
|---|---|---|---|---|---|
| Aviral Kumar | MLD + CS | AP (joined Fall 2024) | Offline RL → LLM reasoning; SCoRe (multi-turn self-correction RL); ArCHer (hierarchical RL for LLM agents) | SCoRe = NeurIPS’24; direct overlap with Weijia’s RL+agent work; new lab likely recruiting; AI2050 Fellow | profile |
| Akari Asai | LTI | AP (joining Fall 2026) | Self-RAG, DR Tulu (RL deep research agent), A-RAG; retrieval + agent training | Officially recruiting Fall 2026 PhD; MIT TR 35 (2025); research agents = direct overlap | profile |
| Graham Neubig | LTI | Assoc. Prof | NeuLab; LLMs, agents, evaluation, code; organized CMU LLM Agent Workshop | Google DeepMind ×7 alumni; strongest frontier placement in LTI; NeuLab = hub for agent research | profile |
| Niloofar Mireshghallah | LTI + EPP | AP (joining Fall 2026) | Privacy + LLM safety + long-horizon & embodied agents; ex-Meta FAIR Alignment | Explicitly recruiting; mentions agent work; coming from Meta FAIR = frontier network | profile |
Priority P2 — 值得陶瓷
| Professor | Dept | Position | Research Focus | Why P2 | Profile |
|---|---|---|---|---|---|
| Daniel Fried | LTI | AP (joined Aug 2022) | Code generation, language agents for SE, SWE-bench Multimodal, COFFE | Recruiting Summer/Fall 2026; SE agent + VLM overlap; code agents = adjacent to GUI agents | profile |
| Sean Welleck | LTI | AP | L3 Lab; AlphaVerus (verified code gen), L1 (reasoning length control); formal verification + LLM | Recruiting Fall 2026 PhD; if Weijia wants to add verification angle to agent work | profile |
| Maarten Sap | LTI | AP | SOTOPIA (ICLR’24), SOTOPIA-π (ACL’24); social NLP + agent evaluation frameworks | SOTOPIA = reusable agent eval framework; 2025 Packard Fellow; good for agent eval direction | profile |
| Yonatan Bisk | LTI + Robotics | AP | REAL Center; embodied AI, VLM, multimodal grounding, DARPA ISAT | If Weijia’s VLM/GUI work evolves toward embodied agents; Amazon Research Award 2025 | profile |
Priority P3 — 可选陶瓷（方向匹配时）
| Professor | Dept | Position | Research Focus | Condition for P3 | Profile |
|---|---|---|---|---|---|
| Tim Dettmers | MLD + CSD | AP | QLoRA (NeurIPS’23), bitsandbytes; LLM efficiency/accessibility; AI2050 Fellow | Only if Weijia works on efficient SFT / on-device agents | profile |
| Lei Li | LTI | Assoc. Prof | LLM generation, multilingual NLP, text watermarking | Only if pivoting to generation quality or multilingual | profile |
| Daphne Ippolito | LTI | AP | LLM safety, data poisoning, AI-generated text detection | Only if pivoting to data integrity / safety | profile |
| Aditi Raghunathan | MLD | AP | ML robustness, certified defenses, distribution shift, LLM safety | Only if adding certified robustness theory angle | profile |
| J. Zico Kolter | MLD | Professor & Director | AI safety, GCG jailbreak, OS-Harm benchmark, adversarial defenses | Only if Weijia does agent safety evaluation; senior = limited bandwidth | profile |
2026 New Joiners (特别关注)
这两位 Fall 2026 才正式入职，lab 尚无 PhD 学生，现在陶瓷最有竞争力：
| Professor | Joining | Priority | Note |
|---|---|---|---|
| Akari Asai | Fall 2026 | ⭐⭐⭐⭐⭐ P1 | 明确招生；自 RAG→RL agent 方向完全对口 |
| Niloofar Mireshghallah | Fall 2026 | ⭐⭐⭐⭐ P1 | 明确招生；来自 Meta FAIR Alignment；提到 embodied agents |
CMU 冷邮件策略
第一梯队（立即发邮件，2026 年暑期）
- Akari Asai — 新加入 LTI，明确招生。重点提：Self-RAG 与你的 RAG 经验，GUIAgentDebugger 与 agent 研究的连接，以及 Yale Cohan 的研究方向。
- Niloofar Mireshghallah — 新加入 LTI，明确招生。重点提：long-horizon agent 与你的 MSRA/GUIAgentDebugger 经验。
- Aviral Kumar — Fall 2024 入职，lab 早期。重点提：SCoRe 与多轮 RL 优化的理解，ArCHer 对 hierarchical agent 的启发。
第二梯队（2027 年，Yale M.S. 有成果后）
- Graham Neubig — 资深，需要有论文成果再陶瓷。重点提：已发表的 agent 方向论文 + LTI 具体研究方向。
- Daniel Fried — 2026 年夏/秋招生。可先跟进，有代码 agent 工作时再联系。
邮件模板要素
- 具体提到一篇他们的论文（精读，提出有深度的问题）
- 简述 GUIAgentDebugger + MSRA 经历（1–2 句）
- 提到 Yale 导师 Arman Cohan（建立信任）
- 说明 2028 PhD 目标（不需要他们立刻回应）
- 简短（< 200 字）
CMU vs UIUC 对比
| 维度 | CMU | UIUC |
|---|---|---|
| NLP/Agent 顶尖 AP 数量 | ⭐⭐⭐⭐⭐ 更多 | ⭐⭐⭐⭐ 强 |
| Frontier 实验室 placement | ⭐⭐⭐⭐⭐ Neubig lab = DeepMind×7 | ⭐⭐⭐⭐ Heng Ji = DeepMind×2, Anthropic×1 |
| 2026 新入职教授 | Asai + Mireshghallah（P1×2） | 无新入职 AP |
| Heng Ji 特殊桥梁 | ❌ 无 Yale 直连 | ✅ Weijia→Cohan→Tang→Wang→Ji |
| RL post-training 专家 | ✅ Aviral Kumar (P1) | ⚠️ 无直接专家 |
| GUI/VLM agents | ✅ Bisk（embodied）, Fried（SWE-bench Multimodal） | ⚠️ 较弱 |
推荐： CMU 和 UIUC 并行陶瓷。CMU 优先 Asai/Kumar/Mireshghallah（2026），UIUC 优先 Heng Ji（通过 Yale 桥梁）。