Track papers across labs, topics, and trends.
Search papers, labs, and topics across Lattice.
This paper introduces Latent Recurrent Thoughts (LRT), a novel approach that utilizes a frozen large language model (LLM) in conjunction with a small recurrent reasoner to refine continuous latent representations for improved reasoning. By decoupling the depth of computation from model size, LRT allows for iterative processing of latent states, leading to enhanced performance on tasks such as symbolic reasoning and natural-language reasoning. The results demonstrate that LRT significantly outperforms existing continuous-space reasoning methods and traditional chain-of-thought prompting while maintaining a lower inference compute cost.
Iterative refinement of latent states using a small recurrent reasoner can drastically enhance reasoning performance in frozen LLMs, outperforming traditional methods at a fraction of the compute cost.
Chain-of-thought reasoning unfolds in discrete token space: each step is committed as text, errors propagate, and eliciting good traces presupposes traces to imitate. Reasoning instead in a model's continuous representation space - where intermediate states are vectors rather than words - sidesteps these constraints, but leaves open how those latent states should be computed. We approach this along two axes. First, we keep a large language model (LLM) frozen and use it for what it is already good at - modeling and decoding sequences - while a small auxiliary network supplies continuous latent thoughts as input. Second, we produce those latents by recurrence: a tiny recurrent reasoner refines them over many steps, decoupling the depth of computation from the size of the model, so that the latents are a product of iterative processing rather than a single forward pass. We instantiate this as Latent Recurrent Thoughts (LRT): a task-dedicated proposer supplies base latents, a recurrent reasoner refines them through bounded residual corrections, and the frozen LLM decodes the answer. On symbolic reasoning with answer supervision but no reasoning traces (Countdown-4, Sudoku) and on natural-language reasoning (HumanEval, MBPP, StrategyQA), LRT substantially outperforms prior frozen-decoder continuous-space reasoning methods under an identical decoder, prompt, data, and training budget, and outperforms non-thinking-mode chain-of-thought prompting on the same backbone at a small fraction of its inference compute.