Continuous Interaction Diffusion: A New Runtime for Tool-Augmented Reasoning in Diffusion LLMs
A new research paper introduces Continuous Interaction Diffusion (CID), a model-runtime architecture designed to integrate tool interaction into the iterative denoising process of diffusion language models (dLLMs). The paper, available on arXiv (2608.10438), argues that traditional autoregressive models naturally fit tool use by emitting a tool call, waiting for the result, and continuing generation. However, diffusion models, which refine multiple parts of their output in parallel, face restrictions with this stop-and-resume pattern. This can force premature tool decisions, delay observations, and introduce redundant refinement and tool execution, potentially harming accuracy and efficiency. CID addresses these issues by separating a model-read-only fact channel, allowing the model to interact with tools during denoising. The architecture aims to improve both task accuracy and inference efficiency for tool-augmented reasoning in diffusion models.
Key facts
- Paper introduces Continuous Interaction Diffusion (CID) for diffusion language models.
- CID integrates tool interaction into iterative denoising.
- Traditional autoregressive models use a stop-and-resume pattern for tool calls.
- Diffusion models refine output in parallel, making stop-and-resume restrictive.
- CID separates a model-read-only fact channel.
- Paper available on arXiv with ID 2608.10438.
- Aims to improve task accuracy and inference efficiency.
- Addresses redundant refinement and tool execution.
Entities
Institutions
- arXiv