Reproduce “LLM Latent Reasoning as Chain of Superposition” 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
Jingcheng Deng, Liang Pang, Zihao Wei, Shicheng Xu, Zenghao Duan, Kun Xu, Yang Song, Huawei Shen and 1 more
Researchers introduce Latent-SFT, a framework enabling large language models to perform latent reasoning by considering a superposition of multiple logical paths. The method reduces reasoning chain length by 2.7x to 5.5x while improving accuracy by up to 2.0% on mathematical reasoning benchmarks compared to explicit chain-of-thought approaches.
Key insights
- Raw hidden states are distributionally misaligned with a model's token embedding space, making direct use as latent tokens unstable and semantically imprecise.
- LLMs can leverage a compact latent space to explore and compress multiple reasoning paths simultaneously, operating in a 'superposition' state rather than a single explicit sequence.
- Stochasticity in optimization, specifically via Gumbel-noise injection, acts as an implicit Hessian regularizer, preventing overfitting and improving the generalization of latent reasoning trajectories.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.