One Readout, Many Repairs: Diffusion-Guided Hierarchical Search for Tool-Agent Repair

2026-09-28

AI Summary

This study addresses the high computational cost of sequence regeneration and redundant operation selection in failure repair for tool-use agents by proposing the ReCommit framework. The method introduces an "operation support set" that amortizes operation-level computational overhead through a single parallel read. By employing a training-free masked diffusion language model coupled with a hierarchical search algorithm, ReCommit jointly optimizes the operation support set and its concrete implementation while reusing operation-type scores to prevent redundant exploration. Evaluated on the Agent-Diff benchmark, the proposed framework improves the relative recovery rate by 75.9% and reduces the average repair time by 61.3%, demonstrating significant gains in agent failure recovery efficiency.

Abstract

Tool agents use large language models to act through external tools, yet successfully executed calls can still leave user requests unfulfilled. Tool-agent repair seeks alternative call sequences that execute successfully and fulfill the original requests. However, repair requires exploring both operation choices and their concrete realizations, making complete-sequence regeneration costly. Moreover, regeneration repeats operation selection even when failure arises from how those operations are realized. The resulting challenge is to reduce this repetition while preserving exploration of alternative operations and realizations. Therefore, we formulate repair as hierarchical search over operation supports, which we introduce as sets of permitted operation types that define reusable search regions for concrete tool-call sequences. We propose ReCommit, a training-free, diffusion-guided framework for improving tool-agent failure recovery while reducing repair computation. ReCommit amortizes operation-level proposal computation across repair trials by reusing operation-type scores from a single parallel readout of a masked diffusion language model. These scores guide search across supports, while realization search explores alternative entity bindings, arguments, and action composition within each support. Experiments on real failures across four enterprise services in the Agent-Diff benchmark show 75.9\% and 63.2\% relative recovery gains with 61.3\% and 51.3\% reductions in mean full-budget repair time at repair budgets $B=3$ and $B=13$, respectively, over the strongest evaluated 8B comparison method. ReCommit achieves a favorable recovery--cost trade-off, including in comparisons with the evaluated 32B models.

Problem

• Tool-agent repair
• Hierarchical search
• Failure recovery
• Diffusion language model
• Repair computation

Innovation

• Diffusion-Guided Repair
• Hierarchical Search
• Tool-Agent
• Operation Supports
• Masked Diffusion Language Model