Search papers, labs, and topics across Lattice.
This paper introduces DiG-Plan, a novel framework that addresses the issue of early commitment in tool-graph planning by employing a diffusion-based proposer to generate diverse tool sets, followed by an autoregressive refiner for dependency prediction. The study reveals that this approach significantly enhances solution coverage, achieving a Pass@10 score of 0.943 compared to 0.320 with standard autoregressive sampling. Notably, DiG-Plan demonstrates a 10% relative improvement over autoregressive baselines on TaskBench, particularly excelling in complex compositional tasks, indicating its robustness across various domains.
Early commitment in tool-graph planning can be mitigated, boosting solution coverage from 32% to nearly 94% with a novel diffusion-based approach.
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.
