Reproduce “Latent Recurrent Thoughts: Recurrent Refinement of Proposed Latents for Reasoning with Frozen LLMs” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Zhaoliang Chen, Jie Fu
LRT adds a task-trained computation module before a frozen large language model. A proposer creates continuous latent vectors, and a recurrent refiner updates them before the decoder produces an answer. On Countdown-4, exact solve rate reaches 56.7% ± 1.9% versus 8.4% ± 0.1% for EBM-CoT under a controlled setting. The modules are trained per task family, so this is not zero-shot.
Key insights
- Effective latent-space reasoning in frozen LLMs requires both a task-dedicated proposer to generate relevant initial latents and a recurrent refiner capable of multi-step iterative computation.
- Complex reasoning can be achieved by synergistically combining the specialized, iterative computation of a small recurrent module with the broad language understanding of a generalist frozen LLM.
- Deep computational processes can be performed with a small, fixed parameter budget by leveraging recurrence, and training efficiency for these recurrent modules can be maintained through truncated-gradient unrolling.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.