Reproduce “Unlocking the Working Memory of Large Language Models for Latent Reasoning” 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
Lukas Aichberger, Sepp Hochreiter
Researchers developed Reasoning in Memory (RiM), a method that equips large language models with a dedicated internal working memory for latent computation, decoupling internal processing from autoregressive generation. This approach, using fixed memory blocks, enables parallel intermediate computation, leading to competitive reasoning performance on math tasks with substantially reduced inference latency.
Key insights
- LLMs possess an intrinsic capacity for latent reasoning that can be effectively utilized by providing structured internal workspaces like memory blocks, decoupling internal computation from explicit generation.
- A carefully designed two-stage curriculum, involving initial grounding with intermediate reasoning step supervision followed by final answer refinement, is effective in teaching LLMs to leverage these non-autoregressive memory blocks.
- Latent representations within memory blocks evolve systematically, becoming structured and input-dependent, indicating that the model learns to store and transform task-relevant information for intermediate computation.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.