Faculty Profile: Prof. J. Zico Kolter — CMU MLD
Faculty Profile: Prof. J. Zico Kolter — CMU MLD
Position: Professor & MLD Director (Associate Department Head) Institution: Carnegie Mellon University, Machine Learning Department Website: https://zicokolter.com/ Report date: 2026-06-12
Research Focus
AI safety, adversarial robustness, deep learning theory, energy systems ML, OS-Harm (jailbreaking and harmlessness evaluation for LLMs), certified adversarial defenses.
Academic Profile
- PhD: Stanford University
- Part of Bosch Center for AI Research
- One of the founding organizers of NeurIPS AI Safety workshop
- Co-developed GCG (Greedy Coordinate Gradient) attack — seminal jailbreak paper
- OS-Harm evaluation framework for LLM harmlessness
- Very senior and well-networked in AI safety community
Key Publications
| Paper | Venue | Focus |
|---|---|---|
| GCG (Universal and Transferable Adversarial Attacks on LLMs) | 2023 | Jailbreaking via gradient-based suffix optimization |
| OS-Harm benchmark | 2024–2025 | Operating-system level harm evaluation for LLM agents |
| Deep learning theory (implicit regularization) | ICML/NeurIPS | Theoretical foundations of DL |
| Certified adversarial defenses | NeurIPS/ICML | Provable robustness |
Fit with Weijia Zhang
| Dimension | Assessment |
|---|---|
| LLM agent safety (OS-Harm) | ✅ Direct relevance if Weijia does agent safety |
| Adversarial robustness / jailbreaks | ✅ Core expertise |
| AI safety foundations | ✅ Senior voice in the field |
| NLP agents | ⚠️ Via OS-Harm agent evaluation |
| GUI / VLM agents | ⚠️ Adjacent via OS-Harm |
| RL post-training | ❌ Not primary |
Lab / Mentorship Notes
- Very senior professor; mentorship bandwidth may be limited
- Lab is large with multiple PhD students and postdocs
- Bosch partnership provides industry research access
- MLD Director role means significant admin load
Verdict
P3 套磁（若做 agent safety 或 jailbreak 方向）。 Kolter is extremely influential in AI safety, but very senior with high admin load. OS-Harm is directly relevant if Weijia’s GUI agent work touches safety evaluation. However, day-to-day mentorship may come from senior PhD students or postdocs rather than Kolter directly.