Looped Language Models (LoopLMs) introduce a new way for AI systems to reason - silently refining their internal representations instead of generating step-by-step explanations in natural language. By running the same transformer block multiple times, these models achieve deeper, more efficient reasoning without relying on token-heavy chain-of-thought outputs. (ubos.tech)

However, this shift creates a major challenge: traditional reinforcement learning methods fail because they only reward the final output, ignoring the model’s internal reasoning process. (huggingface.co)

A new approach, Reward Latent Thought Trajectories (RLTT), solves this by distributing rewards across the entire reasoning trajectory, enabling more effective learning in latent space. (arXiv)

This leads to significant improvements in reasoning performance across both mathematical and general tasks, even with smaller models. (arXiv)

Interestingly, models trained this way also become more concise, producing shorter and more confident answers without explicit optimization for brevity.

At the same time, LoopLMs connect closely to emerging ideas like representation recycling and adaptive computation, pointing toward a future where models dynamically adjust how deeply they “think.”

But this progress comes with a critical tradeoff.

As reasoning moves fully into latent space, models become more powerful, but also far less transparent.

Unlike chain-of-thought reasoning, there are no intermediate steps to inspect, monitor, or audit.

This challenges one of the few practical tools we currently have for AI safety: observing how models think through problems.

RLTT further intensifies this shift by directly optimizing internal reasoning processes that remain invisible to humans.

The result is a growing gap between capability and interpretability.

We gain efficiency, performance, and scalability, but lose insight into how decisions are made.

This raises a fundamental question for the future of AI: should we prioritize smarter models or more transparent ones?

And is it even possible to have both?

Read the full article here: http://apolo.us/blog-posts/thinking-in-silence-how-looped-language-models-learn-to-reason-without-words

‍