BLOG / RESEARCH LOG
Working notes on trustworthy AI - from optimization and post-training to agentic systems, recursive self-improvement, and formal verification.
GEAR makes global search scale in two ways: GPU batches reduce the effective cost of processing a domain, while reusable constraint-aware bounds reduce how many domains the solver needs to visit.
Exact equality elimination compresses discretized PDE optimization from thousands of variables to single digits, turning global certification into a neural verification problem - and moving the bottleneck to the output head.