DiG-Plan: Mitigating Early Commitment for Tool-Graph Planning via Diffusion Guidance

2026-06-04

AI Summary

This work addresses the limitations of autoregressive decoding in tool planning, which suffers from early commitment and struggles to explore the combinatorially explosive solution space. The authors propose DiG-Plan, a novel framework that introduces diffusion mechanisms into tool graph planning for the first time. DiG-Plan employs a two-stage architecture that decouples combinatorial exploration from structural refinement: it first leverages diffusion-guided generation to produce diverse tool sets, then applies an autoregressive model to predict inter-tool dependencies. By integrating iterative denoising, masked reconstruction, and dependency modeling, the method substantially enhances coverage of the solution space. Evaluated on TaskBench, DiG-Plan outperforms autoregressive baselines by 10% (with the largest gains on complex tasks) and demonstrates strong cross-domain effectiveness on API-Bank, improving Pass@10 coverage from 0.320 to 0.943.

Abstract

Generating executable tool plans requires selecting appropriate subsets from tool libraries, a combinatorial search problem with an exponentially large solution space. However, we identify a critical misalignment in predominant approaches: standard autoregressive (AR) decoding suffers from early commitment, where initial token choices rigidly constrain the search trajectory. A controlled study shows that masked denoising raises Pass@10 solution coverage from 0.320 to 0.943 over AR sampling under matched compute. Motivated by this, we propose DiG-Plan, a framework that decouples combinatorial exploration from structural refinement. DiG-Plan employs a diffusion-based proposer to generate diverse tool sets via iterative refinement, followed by an AR refiner for dependency prediction. On TaskBench, DiG-Plan improves over AR baselines by a 10% relative margin, with the largest gains on complex compositional tasks; API-Bank results show that the propose-refine-select design remains effective across domains. Code is available at https://github.com/puddingyeah/DiG-Plan.

Problem

• tool-graph planning
• early commitment
• combinatorial search
• autoregressive decoding
• solution space

Innovation

• diffusion guidance
• tool-graph planning
• early commitment
• combinatorial exploration
• propose-refine framework