TCP-MCP: Co-Evolution of Prompts and Communication Topologies for Multi-Agent Systems
A new framework called TCP-MCP (Topology-Coupled Prompting for Multi-Agent Collaborative Problem-Solving) has been introduced by researchers. This innovative approach simultaneously evolves agent prompts and communication structures as a single genome. It employs a landscape probe at the initialization stage to fine-tune early search behaviors and utilizes Pareto-front diagnostics to enhance exploration across three key objectives: task performance, token cost, and structural complexity. Built on DeepSeek-V3.2, TCP-MCP demonstrates impressive accuracy rates of 82.66%, 89.96%, and 96.61% on MMLU-Pro, MMLU, and GSM8K, respectively, consistently surpassing automated techniques in these evaluations.
Key facts
- TCP-MCP co-evolves prompts and communication topologies as a unified genome.
- Uses initialization-time landscape probe and Pareto-front diagnostics.
- Optimizes for task performance, token cost, and structural complexity.
- DeepSeek-V3.2 backbone used across all methods.
- Achieves 82.66% on MMLU-Pro, 89.96% on MMLU, 96.61% on GSM8K.
- Consistently outperforms automated methods on three benchmarks.
Entities
—