DeepMind 駁斥「越多智能體越好」

🔑 增強重點摘要

• •The predictive model achieves cross-validated R²=0.524 using coordination metrics like efficiency, overhead, error amplification, and redundancy to forecast performance on unseen tasks.[1][2]
• •Independent agents amplify errors 17.2x compared to centralized coordination's 4.4x, highlighting topology-dependent error propagation.[1]
• •Out-of-sample validation on frontier models like GPT-5.2, Gemini-3.0 Pro, and Flash confirms four of five scaling principles with MAE=0.071-0.077.[1][2]

🛠️ 技術深入

• •Five canonical architectures: Single-Agent, Independent Multi-Agent, Centralized (80.8% gain on parallel tasks), Decentralized (+9.2% on web navigation), Hybrid.[1][2][4]
• •Controlled setup standardizes tools, prompts, and token budgets across 180 configs using three LLM families (GPT, Gemini, Claude) to isolate architecture effects.[1][3]
• •Three key effects: tool-coordination trade-off (tool-heavy tasks suffer multi-agent overhead), capability saturation (diminishing returns above ~45% single-agent baseline), topology-dependent error amplification.[1]

🔮 前景展望AI analysis grounded in cited sources

⏳ 時間線