CANTANTE, Agent Optimization

It presents an approach that solves credit assignment, the core bottleneck of LLM-based multi-agent systems, by decomposing each agent's contribution from the global score for training.

• A local optimizer proposes configurations such as prompts, and compares multiple rollouts for the same query to collect traceable reasoning traces and system scores.
• The attributer contrasts the rollouts to assign credit to each agent, and passes that signal back to the local optimizer.
• The experiments used the existing CAPO prompt optimizer.

The evaluation targets were MBPP, GSM8K, and HotpotQA, and it recorded a better average rank than DSPy-solutions GEPA and MIPROv2.

It also showed improvements of +18.9 points on MBPP and +12.5 points on GSM8K, while maintaining inference-time cost compared to non-optimized prompts.