I work on AI safety and alignment. My interest is in what model behavior is evidence of: I test whether a claimed capability is really there by intervening on generation, on training data, and on post-training itself. I also think about what benchmark-driven development structurally cannot see.
Previously: causal discovery for biological data, and memorization and privacy in neural networks.
I am applying to PhD programs for Fall 2027. If you work on evaluation, behavioral attribution, or causal inference for models, I would like to hear from you.
Position: There are futures that benchmark-driven AI cannot see
Benchmarks decide which problems count as progress, and in doing so they quietly fix the set of futures a research culture can even aim at. We argue that the resulting blind spots are structural rather than incidental, and that they will not be fixed by adding more benchmarks.
HearSayBench: Can LLMs Navigate from Abstract Human Rights to Lived Lives?
A benchmark probing whether models recognize the unstated structural constraints that make standard advice impossible for people in adversarial conditions. Across 11 frontier and open-weight models we find a systemic 37% drop between understanding a situation and giving advice that respects it. Oral at the Pluralistic Alignment Workshop, ICML 2026.
Design Patterns of Human-AI Interfaces in Healthcare
A review of clinical human-AI interfaces yielding 15 recurring information entities and 12 design patterns, each grounded in interviews with 12 healthcare professionals and evaluated with 14 UI designers.
Dynamic Implicit Neural Representations for Image Compression
Implicit neural representations with input-dependent dynamics for image compression.
Technical AI Governance Research Fellow
ERA:AI Fellowship, Emmanuel College, University of Cambridge
Supervised by Fazl Barez (Oxford). Behavioral attribution in language models: what generation-time and training-time interventions can and cannot establish about a claimed capability.
Research Intern
Hong Kong University of Science and Technology
With Huamin Qu (HKUST) and Adam Perer (CMU). Synthesized design patterns from clinical human-AI interaction systems and validated them with practitioners and designers.
ERA:AI Fellowship, Cambridge
30 of 6,000+applicants selected
Oral presentation, ICML 2026
168 of 24,371submissions selected for oral
Iranian National University Entrance Exam
Rank 10 of 150,000+national participants