IAAIAdaLoop
AdaLoop: Adaptive-Depth Latent Reasoning for Audio Language Models
A lightweight plug-in for audio language models: a shared transformer block iterates over the audio representation under question guidance, and a learned halting mechanism exits once the representation is ready, so perception-heavy questions get more steps.
Inside the method
Audio encoder → [ Shared block + question ] × r (halting) → LLM
Simplified conceptual schematic. Consult the paper for the complete architecture.
- Recurrence family
- Adaptive depth
- Depth control
- Learned halting per audio–question pair
- KV / state strategy
- Not applicable (encoder-side)
Reading note
Adds under 3% parameters without modifying encoder or LM; raises average accuracy by 2.9–3.8 points across three models on MMSU, MMAU-Pro and MMAR.
Sources checked 2026-10-28. This catalog does not imply independent reproduction.
Related works
Cite this work
@misc{adaloop2026,
title = {AdaLoop: Adaptive-Depth Latent Reasoning for Audio Language Models},
author = {Lee Seung-woo and Bowen Qi},
year = {2026},
eprint = {2610.06949},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2610.06949}
}