Principled Thoughts for Latent Recursive LLM Systems

Principled Thoughts for Latent Recursive LLM Systems. By Fahd Seddik and one other, from the University of British Columbia. Published on the twenty-eighth of September. Its code repository is public and real: about twenty-five source files.

Large language models can reason in continuous space rather than in decoded text, either by recurring on their own hidden states or by passing those states between agents. Training such a system usually supervises only one thing: the Cross-Entropy, CE, of the final decoded answer. The thought itself, the latent state in between, is left unconstrained. The authors establish, in theory and then empirically, four failures of training with CE alone, each lowering the probability of the correct answer. Two examples are thoughts collapsing together across distinct questions, and thoughts holding on to irrelevant information.

REST, short for REpresentation-Supervised Thoughts, supervises the thought directly. It takes four properties a valid thought representation should have, causality, minimality, separability and stability, and turns each into a differentiable loss that is added to CE. Instead of hoping a good thought emerges from answer supervision alone, the objective says what a good thought is. Each of the four is a demand on the thought rather than on the answer, which is precisely the part that CE-only training left free.

REST is instantiated in latent systems with a single agent and with multiple agents. It needs no architectural changes, and adds no parameters at inference. The comparison is kept tight: the same training data, the same compute, and the same latent budget as CE-only training. Whatever difference appears comes from the extra losses, not from a bigger model or more training.

The evaluation spans seven benchmarks across mathematics, science, medicine and code generation. Across agent settings and model sizes, REST raises accuracy over CE-only training by up to seven point five percentage points, and convergence on a final answer by thirty percent. The thoughts themselves also change. REST thoughts encode more of what is required to reach the correct answer, and decoding them better recovers the output the agent intended.

That last result carries a second benefit. When agents talk to each other in latent space, a person cannot easily read what passed between them. Thoughts that decode back to their intended meaning make that communication easier to interpret, so the same losses that improve accuracy also open a window onto what the system is thinking.