What it is
Frames orchestrator-worker LLM systems and reflection as a bilevel game, then adds verifier-gated memory updates with convergence guarantees and SWE-bench gains. Instead of unguided self-reflection or text-only judging, memory edits are accepted only when an environment-grounded verifier lowers risk. The paper also models orchestrator-worker coordination as a bilevel approximate potential game, tying decomposition quality to coordination slack.
Key results
- 72.2% resolved on 500 SWE-bench, vs 70.8% mini-SWE-agent v2
- 361/500 SWE-bench instances resolved by the full Kimi-based system
Shows grounded verification is necessary for reliable reflective memory improvement when correctness depends on external environment state.