aiqbench analysis
This research challenges the prevailing assumption that rounding errors inevitably accumulate in looped models, demonstrating instead that stable loops are resilient to such shifts unless they exceed specific tolerance thresholds. The authors introduce a practical controller utilizing 8-bit weights for final inference stages, which significantly reduces memory traffic while maintaining or improving performance on complex reasoning tasks like Sudoku and mazes.
A Tilted Bowl Is Not a Slippery Slope: Compressing Looped Models
arXiv:2609.39277v1 Announce Type: cross Abstract: Looped models reason by applying the same block of weights many times, so compressing that block saves memory traffic on every loop. Compressed looped models, however, often collapse, and the collapse is usually blamed on rounding error that accumulates from loop to loop. In this work we test that account on more than 30 models from five families and find, to our surprise, that it holds only for loops that never settle. When a loop settles, a fixed rounding error does not accumulate. It moves the point where the loop settles, much as tilting a bowl moves where a ball comes to rest, and the answer is lost only when the shift is larger than the readout tolerates. This picture lets us predict which models fail from a single label-free measurement, and it tells us why failed models recover: their loops still settle, so a few final loops with 8-bit weights bring the answer back. Motivated by these findings, we build a controller that stops when the model's halting head fires and then finishes with 8-bit loops. On Sudoku-Extreme and Maze-Hard it beats fixed-depth inference by up to 15 points under a third of the weight traffic.
Abstract:Looped models reason by applying the same block of weights many times, so compressing that block saves memory traffic on every loop. Compressed looped models, however, often collapse, and the collapse is usually blamed on rounding error that accumulates from loop to loop. In this work we test that account on more than 30 models from five families and find, to our surprise, that it holds only for loops that never settle. When a loop settles, a fixed rounding error does not accumulate. It moves the point where the loop settles, much as tilting a bowl moves where a ball comes to rest, and the answer is lost only when the shift is larger than the readout tolerates. This picture lets us predict which models fail from a single label-free measurement, and it tells us why failed models recover: their loops still settle, so a few final loops with 8-bit weights bring the answer back. Motivated by these findings, we build a controller that stops when the model's halting head fires and then finishes with 8-bit loops. On Sudoku-Extreme and Maze-Hard it beats fixed-depth inference by up to 15 points under a third of the weight traffic.
Machine Learning (cs.LG); Computation and Language (cs.CL)
arXiv:2609.39277 [cs.LG]
(or arXiv:2609.39277v1 [cs.LG] for this version)
https://doi.org/10.48550/arXiv.2609.39277 Focus to learn more arXiv-issued DOI via DataCite (pending registration)
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Full text under the publisher's open licence · 295 words
Related coverage
Import AI 470: No rights for machines; automating environment generation with SPADE; and building better GPU kernels with Hawkeye
Import AI · Aug 24, 2026
A Key-Position Confound in no-cot-bench (and implications for interpreting looped transformers)
LessWrong · Oct 2, 2026
[AINews] Gemini 4 Argon: GDM’s answer to Astra/Fable, with 1M output
Latent Space · Oct 1, 2026
Expenditure Horizon: Measuring Optimization Ability, with an Application to NanoGPT
METR · Jul 21, 2026