LLM coding agents write from-scratch LP solvers averaging ~10x faster than general-purpose ones
sh_reya · x · 2026-11-03
A blog post argues the future of optimization solvers is instance-specific specialization: LLM coding agents can write specialized solvers from scratch, averaging 10x speedups.
- Benchmark of 5 LP instance families (Stochastic STORM, set covering, p-median, facility location, QAP relaxation), split into dev and held-out instances; validity checked before speed
- Claude Opus 5.5 in Claude Code and GPT-6 Astra in Codex developed solvers in a sandbox with Python/C++/CUDA/Rust, given only family names
- Held-out speedups vs. fastest general-purpose solver: 4.8x/4.4x, 7.5x/5.2x, 65x/36x, 53x/17x (GPU), 11x/5.7x (GPU)
- Conclusion: LLM tools make auto-detecting and exploiting instance structure cheap, suggesting a new solver-building paradigm
More from coding & agent
- Filmmaker shares 5 prompt techniques for realistic AI dialogue — bodily actions as punctuation gave the biggest boost — Eric520CC · 2026-11-03
- Claude Code v2.1.317 adds email-exfiltration confirmation and gateway cost report API — ashwin-ant · 2026-11-03
- Codex now defaults to cloud environments, ignoring the machine you assigned — sloppenheimer · 2026-11-03
- One Commit Touches 1241 Files, +25066/-9 Lines: A 'Yolo' AI Coding Masterpiece — tomjohndesign · 2026-11-03
- Dev shares 6-step privacy-safe analytics setup for a WhatsApp agent using SHA256-hashed IDs — petrusenko_max · 2026-11-03
- Dev embeds in-context feedback to kickoff background agents fixing his app in a couple clicks — SubtleGradient · 2026-11-03