REST adds representation-focused losses to latent reasoning training

Researchers introduce REST, or REpresentation-Supervised Thoughts, a training objective for language models that reason in continuous space rather than decoded text. In the systems described, a model can recur over its own hidden states or pass those states between agents. The paper says training commonly supervises the final decoded answer with cross-entropy, without directly constraining the intermediate thought. The authors frame REST as a way to train those representations alongside the answer.

The paper identifies four shortcomings of training only on the final answer, including thoughts becoming too similar across distinct questions and retaining irrelevant information. The authors say these failures can lower the probability of a correct answer. REST adds differentiable losses to cross-entropy, based on four properties the researchers associate with valid thought representations: causality, minimality, separability, and stability. The objective therefore focuses on how intermediate representations relate to a task, not only whether the final decoded answer is correct.

The researchers apply REST to latent single-agent and multi-agent systems. They report that this requires no architectural changes and adds no inference-time parameters. Their evaluation spans seven benchmarks in mathematics, science, medicine, and code generation. The paper says comparisons used the same training data, compute, and latent budget, and covered different agent settings and model sizes. The source does not list individual benchmark results in its abstract.

Across those evaluations, the authors report accuracy improvements over cross-entropy-only training of up to 7.5 percentage points. They also report convergence on a final answer improved by up to 30%. The paper further says REST-trained thoughts encode more of what is needed to reach the correct answer, and that decoding those thoughts better recovers the agent’s intended output. The researchers link that recovery to easier interpretation of communication in latent systems.

For builders, the proposal is a training-objective change aimed at internal representations, while retaining the systems’ existing architectures and inference-time parameter counts. Its reported tests cover both single-agent and multi-agent settings, but the abstract does not provide enough detail to compare particular benchmarks or assess performance in deployment. The results are claims from the paper’s evaluation, rather than evidence here of how REST performs in other settings.