Aswin Surya
AI-safety research · UC Berkeley EECS
I work on AI safety: chain-of-thought faithfulness and monitorability, whether a model's stated reasoning reflects the computation that produces its answer.
Featured work
Cross-judge agreement breaks on saturated CoT-monitorability metrics
Two LLM judges can agree on 97% of items and still post a Cohen's κ around 0.2, which reads as almost no agreement. The standard reliability check collapses exactly where these metrics saturate: the high-monitorability regime they're built to measure. It looks like the judges disagree. They don't.
writeup & code coming soon
Research & writing
Agentic evaluation
Benchmarks pushing agents on real web tasks, turning a website screenshot into working code.
Berkeley AI Research (BAIR) · Dawn Song
Verifiable safety
Extending data-driven Hamilton–Jacobi reachability for control systems with unknown dynamics.
Berkeley AI Research (BAIR) · Claire Tomlin
A Mosquito is Worth 16×16 Larvae
First-authored. Vision Transformers vs. CNNs for mosquito-larvae classification; presented at AGU.
arXiv · 2022
Background
- Ramp, Backend Software Engineering Intern2026
- Apple, Software Engineering Intern2025
- JustPaid.AI, AI Engineering Intern · YC W232025
- Stanford AIMI · MIT Beaverworks, Research Intern2023
- NASA / UT Austin, AI Research Intern2022–23