INTACT: Inducing Latent Reasoning on a Frozen Backbone

Key points

• Latent reasoning replaces natural-language chains of thought with compact continuous "soft-token" representations, reducing the number of intermediate tokens that must be decoded autoregressively.1
• The encoder iteratively maps the backbone's hidden states to new latent blocks, constructing a latent reasoning prefix, after which the frozen model writes its remaining reasoning and its answer as text.1
• Our method, INTACT, trains this encoder through distillation to reproduce the effect of written reasoning steps while keeping the model's behavior close to that of the frozen backbone, and then reinforces it on answer correctness.1
• Beyond the mathematics setting, the same recipe trained on a multi-domain corpus improves the frozen backbone on twelve evaluation benchmarks spanning mathematics, science, and commonsense, without task-specific adaptation.1

Sources (1)

• [1]INTACT: Inducing Latent Reasoning on a Frozen Backbone
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 14, 04:35 AM

  We use a small external encoder to induce latent reasoning in a frozen backbone model.

  Latent reasoning replaces natural-language chains of thought with compact continuous "soft-token" representations, reducing the number of intermediate tokens that must be decoded autoregressively.

Extractive summary: sentences quoted from the sources.