CANTANTE: Optimizing Agentic Systems via Contrastive Credit Attribution [R]

In three linesCANTANTE solves credit assignment in LLM-based multi-agent systems by decomposing global rewards into per-agent optimization signals. Evaluated on MBPP, GSM8K, and HotpotQA, it outperforms GEPA and MIPROv2 (+18.9 pts MBPP, +12.5 pts GSM8K) with no inference overhead.

Summary generated by Claude — human-verified