IAAILoopSLM
Thinking in Depth, Speaking Directly: Recurrent Latent Reasoning for Paralinguistically Grounded Spoken Dialogue
A spoken dialogue model that reuses a decoder block for latent reasoning with acoustic grounding at every pass, trained in two stages that separate learning to reason from learning to respond so it can reply directly without chain-of-thought.
Inside the method
Speech features → [ Decoder block + acoustic grounding ] × r → reply
Simplified conceptual schematic. Consult the paper for the complete architecture.
- Recurrence family
- Core recurrence
- Depth control
- Fixed latent passes
- KV / state strategy
- See paper
Reading note
On EchoMind it beats Qwen2.5-Omni-7B; versus CoT-SFT it gains over 20 points in reasoning accuracy with 64.5% fewer tokens at half the latency.
Sources checked 2026-10-28. This catalog does not imply independent reproduction.
Related works
Cite this work
@misc{loopslm2026,
title = {Thinking in Depth, Speaking Directly: Recurrent Latent Reasoning for Paralinguistically Grounded Spoken Dialogue},
author = {Shengbo Cai and Yuxiang Wang and Jingran Xie and Zhisheng Zhang and Shun Lei and Di Cao and Teddy Sun and Zhiyong Wu},
year = {2026},
eprint = {2609.37818},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2609.37818}
}