Recent work on recurrent-depth language models

Two recent papers explore how recurrent latent reasoning can become more adaptive and more stable.

🔭 Paper scouting · Trending papers
on Scaling up Test-Time Compute with Latent Reasoning: A Recurrent Depth Approach
by Chanoch Baranes · November 7, 2026 · edit this page

The paper of this week (3) shows how Huginn scales test-time computation by repeatedly applying a shared recurrent block. I found two papers that explore how this fixed recurrent process can be made more adaptive or more stable. To be honest I am not sure if the term “trending” is the most fitting. I enjoyed those papers and they both appeared at major conferences so I’m going to say it’s okay.

MoDr architecture. Source: Zhang et al., Mixture-of-Depth-Recurrent Transformers for Test-Time Reasoning.

MoDr: Mixture-of-Depth-Recurrent Transformers for Test-Time Reasoning replaces Huginn’s single recurrent chain with a learnable hard-gated router over multiple LoRA-based branches, allowing different inputs to follow different latent reasoning paths.

STARS: Stabilizing Recurrent Dynamics for Test-Time Scalable Latent Reasoning keeps the recurrent-depth setup but changes the training objective by regularizing the latent dynamics, making additional recurrences more stable and reducing the performance collapse that can occur at large test-time depths.



Enjoy Reading This Article?

Here are some more articles you might like to read next:

  • Critique: Parameter-efficient, but is it compute-efficient?
  • Reproducing the exit rule and the latent trajectories on Huginn-0125
  • Learning about Information Bottleneck
  • Key points regarding Latenct reasoning
  • A better graph instead of Table 4