Large language models (LLMs) can reason in continuous space by recursively processing their own hidden states or by passing those states between agents. Conventional training supervises only the cross‑entropy (CE) of the final decoded answer, leaving the thought process unconstrained. Theoretical and empirical analyses reveal four failure modes of CE‑only training: collapse of thoughts across distinct questions, retention of irrelevant information, lack of causality in reasoning steps, and instability of representations, all of which reduce the probability of the correct answer.
We introduce REST (Representation‑Supervised Thoughts), a training objective that converts four desirable properties of a valid thought representation—causality, minimality, separability, and stability—into differentiable losses added to CE. REST can be instantiated in latent single‑agent and multi‑agent systems without architectural changes or extra inference parameters.
Across seven benchmarks covering mathematics, science, medicine, and code generation, using the same training data, compute budget, and latent budget, REST improves accuracy over CE‑only training by up to $7.5\%$ and speeds convergence to the final answer by about $30\%$, across various agent settings and model sizes. Moreover, REST thoughts encode more of the information required to reach the correct answer, and decoding them more faithfully recovers the agent’s intended output, making latent communication easier to interpret.
Project website: https://fard-lab.github.io/REST
Review