arXiv proposes TAC-CoT for tool-empowered agent reasoning
Original title: TAC-CoT: Tool-Empowered Implicit Chain-of-Thought via Interleaving Explicit Tokens
TAC-CoT replaces explicit CoT with compact latent segments while retaining tool call tokens improving accuracy and reducing forward passes.
LLM-based agentic systems combine CoT reasoning with tool use but explicit CoT incurs overhead. Implicit CoT does not support tool interfaces. TAC-CoT replaces explicit traces with latent segments retaining textual tokens only for tool calls observations and answers. Experiments on agentic math and QA benchmarks show up to 8.8 point accuracy gain and up to 6.8x fewer forward passes. Analyses confirm improved tool-use quality in single- and multi-tool settings.
- replaces explicit CoT with latent segments
- retains tokens only for tool calls
- up to 8.8 point accuracy gain
- up to 6.8x fewer forward passes
- improved tool-use quality
From the original article
1 min read at arxiv.orgLLM-based agentic systems are increasingly combined with CoT reasoning to solve tasks that require accurate reasoning, external knowledge, and multi-step deliberation. Yet, explicit CoT incurs substantial generation overhead, while implicit CoT, which replaces textual rationales with compact continuous tokens, does not naturally support the explicit communication interface required for agentic tool-use.
By Juseon Do · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.