The paper identifies a credit-assignment flaw in RL training for tool-calling LLM agents: because agent outputs mix structured tool calls with natural-language summaries, standard algorithms like GRPO apply a uniform advantage signal across all tokens, letting noise from summary generation contaminate the gradient for tool-decision tokens. The authors propose SLCA-GRPO, which performs segment-level credit assignment via a Schema-Guided LLM Simulator and hierarchical rewards that route advantages separately to tool-call versus summary tokens. Using a 7B model under matched training budgets, it improves over baselines by 2.53 points in-domain, 1.36 points on the Berkeley Function-Calling Leaderboard, and 9.15 points on tau-squared-Bench.
