Reproduce “Concise-9B: Removing the Slack from a Reasoning Model's Thinking” 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
Alexander Boesgaard Lorup
The authors measure answer-stability slack in Qwen3.5-9B and train a shorter-thinking version through concise self-training on its own short, correct traces. Forced readouts every 2,048 thinking tokens estimate when answers stop changing, though stability does not imply correctness. At a 32K budget, the trained model uses 0.66× the tokens with equivalent accuracy, but trails on hard problems at 80K.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.