Reproduce “Reasoning Models Can Be Effective Without 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
Wenjie Ma, Jingxuan He, Charlie Snell, Tyler Griggs, Sewon Min, Matei Zaharia
NoThinking tests whether bypassing an explicit generated thinking trace improves answers under constrained token budgets. It prefills the model’s thinking block with a short “finished thinking” message, then generates a solution and answer. On AMC 2023 at about 700 tokens, pass@1 was 51.3% for NoThinking versus 28.9% for Thinking. The findings concern DeepSeek-R1-Distill-Qwen-32B and do not establish that models stop reasoning internally.
Key insights
- Models specifically trained to produce explicit thinking steps can perform comparably or better without them, suggesting that reasoning capabilities are implicitly embedded within the model's parameters.
- The explicit Chain-of-Thought process, while seemingly mimicking human-like reasoning, is not strictly necessary for achieving strong performance in advanced LLMs.
- NoThinking, when combined with parallel sampling, offers a more efficient and effective approach to achieving high reasoning accuracy at significantly lower latency and token cost.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.