Problem
Research questions and friction points this paper is trying to address.
continuous language flows
reasoning traces
denoising
variable-length reasoning
Plan Canvas: Fixed Reasoning Regions for Continuous Language Flows
🤖 AI Summary
This study addresses the challenge in continuous language stream generation where variable-length reasoning trajectories render answer starting positions unknown and complicate denoising. To overcome this, we propose a fixed-boundary mechanism based on continuous diffusion models that predefines planning region capacity and answer onset positions, thereby decoupling the denoising clocks for trajectories and answers. Combined with supervised padding techniques, this approach enables independent dual-clock denoising and structured reasoning generation. Experimental results demonstrate that our method improves accuracy from 73.0% to 87.0% on the Deep ProsQA benchmark while significantly increasing the proportion of valid reasoning paths. Notably, performance gains are particularly pronounced in long-chain reasoning scenarios, highlighting the effectiveness of separating trajectory and answer denoising processes for enhancing structured generative reasoning capabilities.
📝 Abstract
Continuous language flows generate text by denoising all positions of a target canvas together. The natural way to add reasoning to such a model is to write a trace ahead of the answer, but the trace length changes from question to question. The answer start is therefore unknown during denoising, and the model has to decide the trace length, the place of every trace token, and the answer at the same time. We propose Plan Canvas to fix the boundary between the trace and the answer. A plan region of fixed capacity holds a compact trace, supervised padding fills its unused positions, and the answer starts at a fixed position. The fixed regions also allow separate denoising clocks for the plan and for the answer. With the trace text, backbone, and canvas length of the free-trace baseline held fixed, Plan Canvas improves accuracy on ProsQA and on Deep ProsQA, a graph benchmark with longer proofs. On Deep ProsQA, accuracy rises from 73.0\% to 87.0\%, the share of questions answered with a valid path rises from 30.8\% to 59.1\%, and the gain is largest on the longest proofs.
Innovation
Methods, ideas, or system contributions that make the work stand out.
Continuous Language Flows
Plan Canvas
Fixed Reasoning Regions
Separate Denoising Clocks
Supervised Padding
🔎 Similar Papers
No similar papers found.
💼 Related Jobs
No related jobs found.
