EVOCHAMBER: Multi-agent AI co-evolves teams, boosting math by 32%

Multi-agent AI systems traditionally evolve agents in isolation or broadcast updates symmetrically, both of which lose the specialization that makes collaboration effective. EVOCHAMBER, a new framework from Yaolun Zhang and five co-authors, solves this by implementing test-time evolution at three levels: individual memory, team structure, and population dynamics. At its heart lies CODREAM (Collaborative Dreaming), a post-task reflection protocol triggered by team failure or disagreement. Agents collaboratively distill insights and route them asymmetrically—from strong performers to weaker ones on specific niches—preserving specialization while filling knowledge gaps. Team-level operators dynamically assemble niche-conditioned teams, while population-level lifecycle operators fork, merge, prune, and seed agents based on performance pressure, all without any additional training.

Tested on three heterogeneous task streams using the Qwen3-8B model, EVOCHAMBER delivered impressive results: 63.9% on competition math, 75.7% on code, and 87.1% on multi-domain reasoning—a 32% relative improvement over the best baseline on math. Ablation studies confirmed that the asymmetric cross-agent transfer is the primary driver of these gains. Perhaps most strikingly, starting from several identical agents, the system spontaneously self-organized into four to five stable niche specialists, a structural signature of true multi-agent evolution. For tech professionals, EVOCHAMBER offers a practical, training-free way to make multi-agent systems self-optimize in real time, reducing the need for costly retraining and enabling adaptive collaboration across diverse tasks.