🛬 Attending

<aside> 😎 Google Scholar

</aside>

<aside> 🤠 Research Gate

</aside>

<aside> 😎 ORCID

</aside>

<aside> 🧐 Academic CV

</aside>

Socials

<aside> 😝

Linkedin

</aside>

<aside> 😁 BlueSky

</aside>

<aside> 😆 X/Twitter

</aside>

I aim to improve AI agents by elevating their upper bound (Reasoning🧠→Smarter🤔) and lower bound (Alignment⚖️→Safer🫡) Specifically, my vertical research focus on improving faithful evaluation to better inform post-training (reasoning-driven RL and alignment) See my *slides for Synergizing Post-Training with The Science of Evaluation.* To that end, I often leverage horizontal methods from actionable interpretability (e.g. model diffing techniques incl. task vectors) and robustness probes (e.g. longitudinal analysis as contamination probe)

• Science of Evaluation:
  • I develop reasoning benchmarks that are often grounded in real-world scientific publications (PioneerPhysics, CauSciBench) or provably complex theories (Lean-based Synthesis+Theoretical Computer Science, Benchmarks&Agents for Mixed-Boolean-Arithmetic Expressions)
  • I study the cross-domain transferability/generalizability of reasoning capability, in pre-training as well as post-training.
  • I probe patterns of Contamination/Memorization in legacy benchmarks and propose to leverage perturbation to revive their utility to combat benchmark saturation.
• Science of Post-Training: by actionable interpretability and explore New Signal for RL (from internal feedback or multi-agent feedback)
  • AtomThink: Multimodal Slow Thinking with Atomic Step Reasoning
• Science of Alignment (See my slides for Emergent Scheming and Strategic Deception) What are the key triggers for models to deceive/scheme/persuade other agents in a covert objective?
  • Real vs. Unreal: Are models more inclined to do harm under more realistic or unrealistic scenario?

Education

<aside> 🐔 PhD

Advisor:

</aside>

<aside> 🐥 MSc. Interdisciplinary Science ETH (CS and Physics) ETH Zurich, Switzerland (2025)

Master Thesis: Prof. Zhijing Jin, Prof. Bernhard Schölkopf

</aside>

<aside> 🐣 BSc. Interdisciplinary Science ETH (CS, Physics and Chemistry) ETH Zurich, Switzerland (2025)

Bachelor Thesis: Prof. Mrinmaya Sachan

</aside>

Portfolio in Progress

Portfolio Items