Reproduce “TAC-CoT: Tool-Empowered Implicit Chain-of-Thought via Interleaving Explicit Tokens” 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
Juseon Do, Sangmin Bae, Jeesu Jung, Hwanjun Song
TAC-CoT enables tool-using language models to alternate short latent reasoning segments with explicit tool calls and observations before answering. Each segment has a fixed length, while the number of segments adapts by query. On tool-enabled GSM8K, Qwen2.5-0.5B-Instruct achieved 61.2% accuracy versus 52.4% for Explicit-CoT, using 43.9 versus 109.2 forward passes per query, a result from one in-domain task.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.