Memdoor Multi-Agent Team Tops LiveCodeBench Hard

A collaborative harness coordinating five instances of Qwen3.8-27B surpassed frontier proprietary systems on LiveCodeBench Hard. The run demonstrates that structured role separation beats single-model prompt chaining.

Technical Breakdown: The Memdoor architecture partitions problem solving into distinct worker nodes. Five 27B parameter models run concurrent phases: specification extraction, algorithmic search, code generation, adversarial test synthesis, and static trace execution. A centralized shared-state memory arbiter synchronizes state transitions. When tests fail, trace artifacts route back solely to the algorithmic revision node.

The framework achieved an 8.4 percent accuracy gain over monolithic runs while cutting inference context requirements. Rather than filling a single 128k context window with speculative scratchpads, each agent operates inside bounded 8k context windows. This topology bounds KV cache memory consumption on inference nodes.

Why This Matters & Why Now: Token costs favor running distributed smaller models over frontier models. Splitting reasoning across small weights drops total execution cost by over 60 percent. Engineering teams must design orchestration meshes instead of waiting for bigger monolithic models.

Related Topic Hub: Autonomous Agents