DecompGym tests coding agents on decomposition solvers
The point
DecompGym benchmarks coding agents on designing decomposition algorithms. Agents write solvers across 81 instances from nine problem families, and each solver must run in 60 seconds, pass a method audit, and land within 1 percent of the optimal objective.
Why it matters
For builders, output checks alone miss bad methods, so agent feedback needs to name the exact defect.
Key details
- Evaluated 12 GPT, Claude, and Gemini models in coding agent tools
- Reflection turned 66 percent of retries into successes, versus 39 percent for regeneration
Original ~6 min → 25 sec here Source: arXiv cs.AI, Oct 30, 2026