PRFC: Learning Progressive Recovery from Function Calling Errors for Tool-Using LLMs
Abstract
AbstractTool-using large language models (LLMs) can make multiple attempts to complete a task, using execution feedback from failed tool calls to identify errors and revise subsequent actions. However, their recovery performance often plateaus in later attempts, as models may repeat similar tool-call errors instead of making effective corrections. To address this limitation, we introduce PRFC, a reinforcement learning method that models error recovery as a sequential decision-making process. PRFC uses a progress-aware reward that assigns credit only when a later attempt improves upon previous failures, encouraging steady improvement across the recovery trajectory. Across four benchmarks with a Qwen3-4B backbone, PRFC improves multi-turn recovery on BFCL-v3 by an absolute 6.4% and achieves the highest mean score across three single-step tasks (API-Bank, T-Eval, and Nestful). These results show that modeling progress across attempts provides an effective learning signal for multi-attempt tool-use recovery. These results suggest that explicitly modeling progress across attempts provides a useful learning signal for tool-use recovery.