https://huskydoge.github.io/husky-blog/tags/recursive-models/ Recent content in Recursive Models on Husky'Log https://huskydoge.github.io/avatar.png https://huskydoge.github.io/avatar.png Hugo -- 0.144.2 en Sun, 19 Apr 2026 12:00:00 -0400 https://huskydoge.github.io/husky-blog/posts/recursive_models/loop-cost/ Sun, 19 Apr 2026 12:00:00 -0400 https://huskydoge.github.io/husky-blog/posts/recursive_models/loop-cost/ <figure class="align-center align-center"><img loading="lazy" src="https://huskydoge.github.io/husky-blog/posts/recursive_models/loop-cost/teaser.png#center" alt="Loop-model FLOPs and memory teaser schematic"/> </figure> <h2 id="introduction">Introduction</h2> <p>Loop models are becoming popular lately, with exciting results <span class="hugo-cite-intext" itemprop="citation"> <span class="hugo-cite-group" style="display: inline-block;"> <a href="#" style="text-decoration: none; border-bottom: 1px dotted #ccc;"> [1,2,3,4,5] </a> <span class="hugo-cite-citation"><span class="hugo-cite-citation-entry"><strong>Less is More: Recursive Reasoning with Tiny Networks</strong><br>A. Jolicoeur-Martineau, (2025)<br><a href="https://arxiv.org/abs/2510.04871" target="_blank">Link</a></span><span class="hugo-cite-citation-entry"><strong>Scaling up Test-Time Compute with Latent Reasoning: A Recurrent Depth Approach</strong><br>J. Geiping, S. McLeish, N. Jain, J. Kirchenbauer, S. Singh, B. Bartoldson, B. Kailkhura, A. Bhatele, T. Goldstein, (2025)<br><a href="https://arxiv.org/abs/2502.05171" target="_blank">Link</a></span><span class="hugo-cite-citation-entry"><strong>Parcae: Scaling Laws For Stable Looped Language Models</strong><br>H. Prairie, Z. Novack, T. Berg-Kirkpatrick, D. Fu, (2026)<br><a href="https://arxiv.org/abs/2604.12946" target="_blank">Link</a></span><span class="hugo-cite-citation-entry"><strong>Scaling latent reasoning via looped language models</strong><br>R. Zhu, Z. Wang, K. Hua, T. Zhang, Z. Li, H. Que, B. Wei, Z. Wen, F. Yin, H. Xing, others, (2025)<br></span><span class="hugo-cite-citation-entry"><strong>Hierarchical Reasoning Model</strong><br>G. Wang, J. Li, Y. Sun, X. Chen, C. Liu, Y. Wu, M. Lu, Y. Yadkori, (2025)<br><a href="https://arxiv.org/abs/2506.21734" target="_blank">Link</a></span></span> </span> </span>&#32;. Once we decide to reuse the same block across multiple layers, however, one practical question becomes unavoidable: <strong>what does the loop cost during training?</strong></p>