Faculty Profile: Prof. Niloofar Mireshghallah — CMU LTI + EPP
Faculty Profile: Prof. Niloofar Mireshghallah — CMU LTI + EPP
Position: Assistant Professor（joining Fall 2026，joint LTI + Engineering & Public Policy） Institution: Carnegie Mellon University; also core member of CyLab Website: https://mireshghallah.github.io/ Report date: 2026-06-12
Research Focus
Privacy and NLP, LLM safety/security, long-horizon and embodied agents (robotics), societal implications of ML. Intersection of privacy, NLP, and policy.
Academic Profile
- PhD: UC San Diego (2023)
- Research Scientist: Meta FAIR, Alignment group (2023–Nov 2025)
- NCWIT Collegiate Award (2020)
- Rising Star in Adversarial ML Award (2022)
- Qualcomm Innovation Fellowship finalist (2021)
- Joining CMU Fall 2026 — lab brand new
Key Publications (recent)
| Paper | Focus |
|---|---|
| PrivacyReasoner: Can LLM Emulate a Human-like Privacy Mind? (2026) | LLM privacy reasoning |
| Probabilistic Reasoning with LLMs for k-anonymity Estimation (2025) | Privacy in LLM output |
| Various works on memorization, unlearning, differential privacy for LLMs | LLM privacy/safety |
Stated Research Interests at CMU
From her CMU profile: privacy & security of LLMs, coding, long horizon & embodied agents (robotics)
Fit with Weijia Zhang
| Dimension | Assessment |
|---|---|
| Long-horizon agents | ✅ Explicitly stated as a CMU research focus |
| Embodied / agentic AI | ✅ Stated interest in embodied agents |
| LLM safety / alignment | ✅ Meta FAIR Alignment group background |
| Privacy / security of LLMs | ✅ Core expertise |
| RL / post-training | ⚠️ FAIR Alignment group suggests RLHF exposure |
| GUI / VLM agents | ⚠️ Adjacent via embodied agents |
Cold Email Strategy
- Reference: GUIAgentDebugger (long-horizon agent debugging, error taxonomy across perception/interaction/reasoning) → maps to her long-horizon agents interest
- Reference: OpenManus-RL (multi-agent collaboration, memory architecture) → maps to agentic + safety intersection
- Mention: Interested in safety and reliability of long-horizon LLM agents, combining her privacy/safety expertise with agent capability work
Verdict
P1 套磁。 Joining Fall 2026 from Meta FAIR Alignment, lab is brand new. Her explicitly stated interest in “long horizon & embodied agents” is a direct match to GUIAgentDebugger. Privacy × safety × long-horizon agents is a differentiating and fundable direction. Email now to express interest.