Reproduce “Token-Budget-Aware LLM Reasoning” 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

Tingxu Han, Zhenting Wang, Chunrong Fang, Shiyu Zhao, Shiqing Ma, Zhenyu Chen

TALE estimates question-specific reasoning budgets and uses them to guide shorter chains of thought. On six GPT-4o-mini task sets, TALE-EP averages 81.03% accuracy and 148.72 output tokens, versus Vanilla CoT at 83.75% and 461.25 tokens. This is 67.8% fewer tokens with a 2.72-percentage-point accuracy drop. On MathBench-College, accuracy falls to 70% versus 78%, showing the average masks task-level losses.

Summary by alphaXiv.

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