Reproduce “The Value of Thinking: Learning When to Reason in Hybrid Decision Models” 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

Zhengxu Yu, Ke Xu, Meng Fang, Haitham Ammar, Weilin Luo, Jun Wang

HDM routes requests between a token-free fast answer and autoregressive reasoning, using a small block to estimate whether more reasoning is worth its cost. Its frozen backbone processes each request once; escalated requests reuse cached context. On JevBench’s 231 public items, HDM-4B scored 0.857 with median time 0.179 seconds, versus Jev’s 0.851 and 0.652 seconds. This supports competitiveness, not dominance; latencies span different backends.

Summary by alphaXiv.

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