State-Path Tool Menu boosts agent success on ToolBench to 0.898
Researchers have introduced the State-Path Tool Menu, a framework designed to improve how AI agents select tools from large libraries. Traditional methods often rank tools by surface-level relevance, which can fail to include necessary prerequisite tools or order them correctly for multi-step tasks. The new approach treats the tool menu as an execution prior, mapping a pre-execution route from…
Key points
- State-Path Tool Menu improves ToolBench online success from 0.737 to 0.898.
- Framework orders tools by execution dependencies rather than simple relevance ranking.
- Method outperforms baselines without altering the underlying language model agent.
The system uses an encoder to represent tool dependencies and a retriever to identify missing input producers and final actions. A reranker then ensures producers are placed before consumers. In tests on the ToolBench benchmark, this method increased online success rates from 0.737 to 0.898. It outperformed standard retrieval, reranking, and routing baselines without modifying the underlying agent model.
Notably, the State-Path menu achieved higher coverage of complete action chains using only 32 tools compared to 128 in the official list. The performance gains were consistent across different executor families, suggesting the method is robust to variations in model capacity. The code for this framework is available on GitHub for further research and implementation.
The headline, key points and digest above were generated by Digest AI's editorial model from the linked sources. Automated summaries can contain errors: the sources are the record. Spotted a mistake? Tell us.
More in Research
All →- Guide: How AI Embeddings Encode Meaning into Numerical Vectors · 1 src
- LLM-Anchored Paralinguistic Boost for Alzheimer's Detection · 1 src
- New probability-wave framework links trader behavior to AGI architecture design · 1 src
- MAAGL framework improves graph reasoning with multi-agent collaboration · 1 src
- XAI-Arena uses LLMs to benchmark explainable AI quality · 1 src
Comments
via GitHub Discussions