Latent Recurrent Thoughts: Recurrent Refinement of Proposed Latents for Reasoning with Frozen LLMs
A small recurrent module generates and refines continuous “thought” vectors while the main LLM stays frozen.
The paper frames LRT as reasoning outside token-by-token chain-of-thought, using latent vectors instead of written intermediate steps. A task proposer creates base latents, a recurrent reasoner iteratively adjusts them, and the frozen decoder produces the final answer. The authors report gains over prior frozen-decoder latent-reasoning methods on symbolic, coding, and QA benchmarks under matched settings. They also say it beats non-thinking-mode chain-of-thought prompting on the same backbone with much less inference compute. ArXiv · AI/CL/LG's note
The paper frames LRT as reasoning outside token-by-token chain-of-thought, using latent vectors instead of written intermediate steps. A task proposer creates base latents, a recurrent reasoner iteratively adjusts them, and the frozen decoder produces the final answer. The authors report gains over prior frozen-decoder latent-reasoning methods on symbolic, coding, and QA benchmarks under matched settings. They also say it beats non-thinking-mode chain-of-thought prompting on the same backbone with much less inference compute. ArXiv · AI/CL/LG's note
score 5