Scaling up Test-Time Compute with Latent Reasoning: A Recurrent Depth Approach
akhaliq
Abstract
We study a novel language model architecture that is capable of scaling test-time computation by implicitly reasoning in latent space. Our model works by iterating a recurrent block, thereby unrolling to arbitrary depth at test-time. This stands in contrast to mainstream reasoning models that scale up compute by producing more tokens. Unlike approaches based on chain-of-thought, our approach does not require any specialized training data, can work with small context windows, and can capture types of reasoning that are not easily represented in words. We scale a proof-of-concept model to 3.5 billion parameters and 800 billion tokens. We show that the resulting model can improve its performance on reasoning benchmarks, sometimes dramatically, up to a computation load equivalent to 50 billion parameters.
Community
Very cool paper! This is another empirical evidence of "latent reasoning". Just like COCONUT from Meta AI 🚀
Code and more info can be found here: https://github.com/seal-rg/recurrent-pretraining
The model collection is here: https://huggingface.co/collections/tomg-group-umd/recurrent-models-6793c4ae556d3843c5496572
Feel free to ask any questions here as well!
Question for the authors, how do you see this approach compared with Deep Equilibrium Models? Did you test using a deq model as the intermidiate layers? I think the overall approach would benefit a lot to have a objective that tries to "optimize" the latent states in a stable embedding.
Hi! The difference to an equilibrium model is in the training objective, this model is trained with truncated backpropagation through the unrolled recurrence, whereas DEQ would differentiate through the fixed point. Both approaches are related to optimization in intermediate layers, the internal "energy" function is just not learned explicitly.
One advantage that we see with unrolling over fixed point differentation is that we do not necessarily want to require the operator to converge to a fixed point, with unrolling we leave this open. Simply training with sufficient scale, we see that the model finds a solution that does converge in a pretty straightforward many for most tokens, but it also implements orbits and other structures which seem helpful (See section 7).
This is an automated message from the Librarian Bot. I found the following papers similar to this paper.
The following papers were recommended by the Semantic Scholar API
- Token Assorted: Mixing Latent and Text Tokens for Improved Language Model Reasoning (2025)
- Deliberation in Latent Space via Differentiable Cache Augmentation (2024)
- Byte Latent Transformer: Patches Scale Better Than Tokens (2024)
- Long-Range Tasks Using Short-Context LLMs: Incremental Reasoning With Structured Memories (2024)
- Efficient Reasoning with Hidden Thinking (2025)
- Sample, Scrutinize and Scale: Effective Inference-Time Search by Scaling Verification (2025)
- Compressed Chain of Thought: Efficient Reasoning Through Dense Representations (2024)
Please give a thumbs up to this comment if you found it helpful!
If you want recommendations for any Paper on Hugging Face checkout this Space
You can directly ask Librarian Bot for paper recommendations by tagging it in a comment:
@librarian-bot
recommend
Models citing this paper 1
Datasets citing this paper 0
No dataset linking this paper
Spaces citing this paper 0
No Space linking this paper