Reproduce “Trajectory-Triggered Sparse Collaborative Decoding” 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
Amit Ranjan Trivedi, Sruthikesh Surineni, Kyong Ho Lee, Mohit Mittal, Steven Li, Rex Hill
Sparse collaborative decoding lets a draft model handle most reasoning and uses a learned trigger to call a stronger target when trajectory features predict failure. The target adds bounded text, then the draft resumes, without requiring target logits or activations. On GSM8K, failure prediction reaches AUC 0.864, while MATH-500 exposes limits.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.