Reproduce “iCLP: Large Language Model Reasoning with Implicit Cognition Latent Planning” 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

Sijia Chen, Di Niu

The iCLP framework enables large language models to perform complex reasoning by generating adaptive "latent plans" in a compact hidden space, which then guide natural language Chain-of-Thought (CoT) generation. This approach yields an average accuracy improvement of 11.5% over baselines, reduces token cost, and shows robust cross-domain generalization while maintaining reasoning interpretability.

Key insights

  • Explicit reasoning plans exhibit inherent clustering within task categories and overlap across them, indicating the potential for reusable, abstract patterns.
  • Modeling LLM planning after human implicit cognition, which relies on subconscious, compact, and generalized patterns, can overcome limitations of explicit textual instructions.
  • A hybrid approach of planning in a compact, latent space and reasoning in interpretable natural language effectively balances computational efficiency, generalization, and user understanding.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.