Bilevel Coordinated Reflection: A Game-Theoretic Approach to Multi-Agent LLM Systems
The paper argues that reflection in multi-agent LLMs only reliably helps when checked against the environment, not just the transcript.
The authors model orchestrator-worker systems as a bilevel coordination game, tying worker equilibrium quality to how well tasks are decomposed. They formalize reflection as movement through memory states and show limits for free-form transcript-only reflection. Their proposed SRMA method accepts memory updates only when grounded evaluation risk falls. In SWE-bench tests, the full Kimi-based system resolves 72.2% of 500 instances, compared with 70.8% for the public mini-SWE-agent reference. HF Daily Papers' note
The authors model orchestrator-worker systems as a bilevel coordination game, tying worker equilibrium quality to how well tasks are decomposed. They formalize reflection as movement through memory states and show limits for free-form transcript-only reflection. Their proposed SRMA method accepts memory updates only when grounded evaluation risk falls. In SWE-bench tests, the full Kimi-based system resolves 72.2% of 500 instances, compared with 70.8% for the public mini-SWE-agent reference. HF Daily Papers' note
score 4