Reproduce “DART: Distilling Autoregressive Reasoning to Silent Thought” 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
Nan Jiang, Ziming Wu, De-Chuan Zhan, Fuming Lai, Shaobing Lian
The DART framework enables Large Language Models to perform complex reasoning without explicit intermediate steps, replacing autoregressive Chain-of-Thought with a non-autoregressive "Silent Thought" process during inference. It delivers up to 12 times faster inference compared to standard CoT methods, achieving an accuracy of 42.6% on the GSM8K benchmark while outperforming existing non-autoregressive approaches by 10.1%.
Key insights
- Complex autoregressive reasoning can be effectively distilled into a compact, latent representation that is processed non-autoregressively in a single forward pass.
- The Reasoning Evolvement Module (REM) provides an efficient and compatible mechanism for inducing reasoning-related shifts in hidden states without incurring inference overhead.
- Aligning the hidden state of the last token before the answer during distillation is crucial for transferring the progressive reasoning patterns from explicit CoT to a silent, internalized process.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.