New Recurrent Depth Architecture Enables Implicit Reasoning Without Chain-of-Thought

iScienceLuvr · x · 2026-09-02

A paper titled 'Scaling up Test-Time Compute with Latent Reasoning: A Recurrent Depth Approach' proposes a novel language model architecture. It iterates a recurrent block to unroll to arbitrary depth at test-time, performing implicit reasoning in latent space. This contrasts with mainstream reasoning models that scale compute by producing more tokens. The approach requires no specialized CoT training data, works with small context windows, and captures reasoning not easily represented in words. A proof-of-concept has been scaled to 3.5B parameters.

Original post →

More from Research

Research channel →