Reproduce “STITCH: Simultaneous Thinking and Talking with Chunked Reasoning for Spoken Language Models” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Cheng-Han Chiang, Xiaofei Wang, Linjie Li, Chung-Ching Lin, Kevin Lin, Shujie Liu, Zhendong Wang, Zhengyuan Yang and 2 more
STITCH (Simultaneous Thinking and Talking with Chunked Reasoning) enables Spoken Language Models (SLMs) to perform internal reasoning while generating spoken responses. This approach significantly improved accuracy on math reasoning tasks by over 15% compared to non-reasoning baselines, with the STITCH-S variant achieving these gains without introducing additional first-packet latency.
Key insights
- Interleaving unspoken reasoning with spoken output allows SLMs to perform complex reasoning while minimizing latency, especially with the STITCH-S variant's zero additional first-packet latency.
- The ability to adjust reasoning chunk length at inference time provides flexibility to balance computational resources, latency, and desired reasoning depth without retraining.
- STITCH can effectively integrate reasoning from external models, enabling SLMs to leverage stronger, specialized reasoning components to further enhance their spoken responses.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.