This paper from researchers at the University of Maryland and Lawrence Livermore National Laboratory introduces a fundamentally different approach to test-time scaling. Rather than having models "think longer" by generating more chain-of-thought tokens (the approach used by o1, DeepSeek-R1, and similar systems), they propose a recurrent depth architecture where a single block is iterated repeatedly at inference time. The model can unroll to arbitrary depth, spending more compute without producing any additional output tokens. Their proof-of-concept reaches 3.5B parameters trained on 800B tokens.
The dominant paradigm for reasoning models right now — generate more tokens, think out loud — has real limitations. It requires specialized training data with reasoning traces, burns through context window budget, and fundamentally can only capture reasoning patterns that can be expressed in natural language. This latent reasoning approach sidesteps all three constraints. There's no need for curated chain-of-thought datasets, context window is not consumed, and the model can potentially internalize reasoning patterns that are genuinely hard to verbalize. The result: on some reasoning benchmarks, the 3.5B model performs equivalently to a 50B parameter model when given sufficient compute budget at inference time.
If this approach scales well, it represents a third path for reasoning AI — alongside pure scale (bigger models) and explicit chain-of-thought (more tokens). Latent reasoning could be particularly valuable in constrained environments: embedded systems, real-time applications, or situations where verbose chain-of-thought output is undesirable. It also raises interesting questions about interpretability — when reasoning happens entirely in latent space, there's no scratchpad to inspect, which could complicate alignment and safety work.
The paper is an explicit proof-of-concept, and the authors are candid about this. Key unknowns include whether the approach scales to larger base models, whether it can match the raw benchmark performance of dedicated reasoning models like o1 or R1, and whether the latent representations developed during iterative unrolling are stable and predictable. The fact that they released model weights and code suggests strong community confidence — we'll likely see follow-up scaling studies from independent researchers soon.