A new research paper posted on Hugging Face presents a theoretical framework for multi-agent large language model (LLM) systems, where an orchestrator decomposes tasks for a team of worker agents that improve through textual reflection. The authors argue that despite strong empirical results, such systems lack a unified account of coordination, memory improvement, and external verification.
The study formalizes orchestrator-worker interaction as a bilevel coordination game. Under bounded coupling, the workers' local-update game is shown to be an approximate potential game, with equilibrium slack controlled by decomposition quality. This provides a theoretical basis for how task decomposition affects overall system performance.
On the reflection side, the paper analyzes reflection as stochastic movement over semantic memory states. It proves an information-theoretic impossibility result: any gate that observes only the generated transcript cannot uniformly improve memory across text-indistinguishable environments, whereas an environment-grounded gate can. This separation motivates the introduction of Stochastic Reflective Memory Ascent (SRMA), which accepts a candidate memory only after a grounded evaluation risk strictly decreases.
SRMA is shown to converge exactly, with geometric or polynomial rates that are order-tight under calibration and non-degenerate corrective mass. The paper also provides confidence gating for stochastic evaluation and re-anchoring guarantees for piecewise-stationary environments. Experiments instantiate these objects with environment-grounded metrics.
On 500 SWE-bench instances, the complete Kimi-based system resolves 72.2% of tasks, compared to a 70.8% public mini-SWE-agent reference. The code is available on GitHub.