đź”— Source: arXiv

Training Large Language Models to Reason in a Continuous Latent Space

🚀 Technical Novelty

  • Mechanism: Directly feeds the LLM’s last hidden state (a “continuous thought”) back as the next input embedding instead of decoding to discrete tokens, trained via a multi-stage curriculum.
  • Nuance: Differs from prior SOTA by eliminating language-space constraints entirely during reasoning, allowing superposition of multiple reasoning paths and implicit breadth-first search without explicit tree-search algorithms or pause tokens.

đź’ˇ Yield

  • Outperforms standard CoT on logical reasoning tasks (ProntoQA, ProsQA) with significantly fewer generated tokens.
  • Achieves a superior accuracy-efficiency trade-off on math reasoning (GSM8k) by effectively chaining continuous thoughts.
  • Emerges with implicit BFS-like planning capabilities without explicit training objectives for search.

⚠️ Limitations

  • Requires a multi-stage curriculum guided by language CoT data; struggles to learn latent reasoning purely from end-to-end gradient descent on Q&A pairs.
  • Performance gap remains compared to carefully tuned iCoT baselines, indicating room for improved training strategies.
  • Generalization to pretraining and broader reasoning challenges needs further exploration.