Reproduce “A Model with No Head and Many Thoughts” 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

Nikita Koriagin, Yaroslav Aksenov, George Bredis, Gleb Gerasimov, Nikita Balagansky, Daniil Gavrilov

Soft Latent Thinking (SLT) enables large language models to execute intermediate reasoning steps through a compact latent projector, replacing the computationally intensive full vocabulary head. This design improves pass@k accuracy on mathematical reasoning benchmarks, enhances token efficiency, and significantly reduces the computational overhead of reasoning operations.

Key insights

  • Decoupling the internal reasoning process from the explicit full vocabulary projection via a compact latent projector reduces computational costs and allows for more flexible, continuous state transitions.
  • The combination of a lower-dimensional latent space and Gumbel-Softmax sampling promotes greater diversity across reasoning rollouts, which contributes to improved multi-sample accuracy (pass@k).
  • A modular training strategy involving a specialized latent projector and LoRA adapters on a frozen backbone enables efficient adaptation for specific reasoning domains and flexible deployment, including an option to bypass the projector.

Summary by alphaXiv.

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