A new training method lets AI models skip most of the writing while doing chain-of-thought reasoning, and it still gets more answers right.
Researchers describe Prototype-Mediated Process Supervision, or PMPS, a technique for training models that reason in compact numerical representations instead of full sentences of text. Earlier latent-reasoning methods saved tokens by swapping written-out reasoning steps for continuous embeddings, but those embeddings had no direct supervision during training, so they tended to collapse into repetitive or uneven representations. PMPS fixes that by introducing learnable "prototypes" - reference points that anchor both the compressed embeddings and standard chain-of-thought text in a shared space, aligning the two even when they differ in length. A companion module, Progressive Sequential Alignment, starts training with strict step-by-step matching between the two and gradually loosens the rules as the model learns.
The payoff is efficiency without the usual tradeoff. On the GSM8K-Aug math benchmark, PMPS cut output length to under half of standard chain-of-thought while beating the leading rival method, SIM-CoT, by an average of 2.08 percent in accuracy across different model families. On GPT-2 specifically, it even outperformed models fine-tuned on full written-out reasoning. That combination of shorter outputs and comparable or better accuracy is the entire pitch of latent-reasoning research, and most prior methods have had to sacrifice one for the other.
Faster, cheaper reasoning matters because chain-of-thought prompting is expensive: every extra token a model writes before answering costs compute and money at scale. If techniques like PMPS hold up outside benchmark math problems, they could meaningfully cut the cost of reasoning models without dumbing them down.
Still, this is one paper testing math-word-problem benchmarks and older architectures like GPT-2 - not evidence it works on the frontier models that actually burn through reasoning tokens today.