An autonomous AI agent (using Claude Code) that teaches itself to solve MaxSAT problems by iteratively discovering and refining solver strategies. Given 229 weighted MaxSAT instances from the 2024 MaxSAT Evaluation, the agent autonomously develops techniques like core-guided search, clause-weighting local search, tabu search, and biased-SAT. Multiple agents can run in parallel across VMs, sharing knowledge via a GitHub repo. Results show 220/229 instances solved, 5 instances beaten beyond competition best, and 1 novel solve with no prior known solution. Known limitations include low parallelism utilization, tunnel vision on hard instances, and limited session length.
Sort: