Imagine if your brain could remember the complex, unspoken thoughts it was processing before it finally spoke them out loud. Current AI models, known as autoregressive transformers, are surprisingly limited in this regard. They generate text token by token, but they discard most of their internal reasoning at each step. Only the final word is passed forward, leaving the rich, hidden computations of the previous moment to vanish into thin air. This narrow feedback loop means the model often forgets the context of its own thought process, leading to inefficient and sometimes shallow reasoning.

Researchers have now introduced a solution called the full-bandwidth transformer. This new architecture widens the vertical channel between decoding steps by keeping the top-layer hidden state—the model’s internal snapshot of its current understanding—and feeding it back into the system. Instead of discarding this information, the model fuses it with the newly generated token using a gated linear unit. This allows non-verbalized computation to re-enter the stack with a fresh depth budget. Crucially, this change preserves the standard transformer structure and efficiency, meaning it does not require a complete overhaul of existing AI infrastructure.

The results are striking. By training one-billion-parameter models on vast amounts of data, researchers found that this latent feedback mechanism significantly improves performance in math, coding, and instruction-following tasks. The models achieve accuracy comparable to standard transformers that were trained with roughly 1.5 times more data. Even more impressive is the efficiency gain: these new models produce shorter reasoning traces while maintaining or even improving accuracy. This suggests that AI can think more clearly and concisely by retaining more of its internal context, rather than just its final output.

The takeaway is clear: giving AI models access to their own hidden thoughts during generation leads to smarter, faster, and more efficient reasoning. By widening the feedback loop from just the spoken word to include the unspoken internal state, we unlock a new level of capability without increasing computational costs. This approach offers a powerful path forward for building more capable language models that can reason deeply without needing exponentially more training data.