Research map / Large language models / Agents, code and math
Automated Creation of Reusable and Diverse Toolsets for Enhancing LLM Reasoning
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes KTCE, a two-stage framework that builds a three-layer knowledge tree from problems and induces reusable atomic Python tools from it, then evolves the toolset with LLM-driven selection, mutation and crossover to improve diversity.
Problem
Tools automatically created by LLMs are overly problem-specific (low reusability) and the resulting toolsets are narrow (limited diversity) for LLM reasoning.
Approach
Treat tools as executable domain knowledge: extract knowledge (topic/concept/key points) from training problems into a clustered knowledge tree and induce atomic tools per key point, then apply genetic-programming-style evolutionary search (selection, mutation, crossover via LLM) guided by a tool-effectiveness and coverage objective.
Key results
- KTCE achieves 69.00% on MATH Algebra vs 57.03% for TROVE (best baseline), and improves accuracy by 6.23% to 18.49% on average across datasets.
- KTCE reaches 64.50%, 82.90%, and 62.00% task coverage on MATH, TabMWP, and SCIBENCH, versus at most 30.50% for prior tool-creation methods.
- Toolset from GPT-3.5-Turbo transfers to GPT-4o-Mini on MATH (+4.06%) and DeepSeek-Coder on SCIBENCH (+5.00%).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use KTCE instead of Library | GPT-3.5-Turbo · MATH Algebra | accuracy: 58.80 -> 69.00 | Helped |
| Use KTCE instead of Library | GPT-3.5-Turbo · MATH Geometry | accuracy: 33.40 -> 40.29 | Helped |
| Use KTCE instead of Creator (SR) | GPT-3.5-Turbo · TabMWP | accuracy: 87.80 -> 90.00 | Helped |
| Use KTCE instead of Creator (SR) | GPT-3.5-Turbo · SCIBENCH | accuracy: 32.00 -> 37.00 | Helped |
| Add Knowledge-based Tool Creation (Stage 1) (ablation) | KTCE on MATH · MATH | accuracy: 50.16% -> 53.14% | Helped |
| Add Tool Evolutionary Search (Stage 2) (ablation) | KTCE on MATH · MATH | accuracy: 50.50% -> 53.14% | Helped |
| Add selection to Tool Evolutionary Search (ablation) | KTCE on MATH · MATH | accuracy: 51.72% -> 53.14% | Helped |
| Add crossover to Tool Evolutionary Search (ablation) | KTCE on MATH · MATH | accuracy: 50.62% -> 53.14% | Helped |
| Add mutation to Tool Evolutionary Search (ablation) | KTCE on MATH · MATH | accuracy: 51.56% -> 53.14% | Helped |
Surprising findings
- Removing Stage 1 (knowledge-based creation) collapses toolset coverage to 9.16% on MATH, showing the knowledge tree is critical for coverage.
Limitations
- All experiments use GPT-3.5-Turbo for toolset creation and the agent; generalization to other creators is limited to a few evaluations.
- Evaluation covers only three reasoning domains (MATH, TabMWP, SCIBENCH) with a 100-problem SCIBENCH test subset.
- Tools are restricted to Python functions and the approach does not address multimodal or embodied tools.
Experimental setup
Datasets: MATH TabMWP SCIBENCH GSM8K
Benchmarks: MATH TabMWP SCIBENCH GSM8K
Metrics: accuracy task coverage tool usage frequency toolset size cyclomatic complexity Halstead volume Halstead effort
Baselines: CoT PoT Library Wolfram Creator Creator (SR) CRAFT TROVE
Builds on: Program-of-Thought CRAFT TROVE CREATOR Genetic Programming
Closest papers by meaning
- Large Language Models as Tool Makers (ICLR 2024 · similarity 0.91)
- CRAFT: Customizing LLMs by Creating and Retrieving from Specialized Toolsets (ICLR 2024 · similarity 0.90)
- Adapting While Learning: Grounding LLMs for Scientific Problems with Tool Usage Adaptation (ICML 2025 · similarity 0.90)
- ToolLLM: Facilitating Large Language Models to Master 16000+ Real-world APIs (ICLR 2024 · similarity 0.90)
- Advancing Tool-Augmented Large Language Models: Integrating Insights from Errors in Inference Trees (NeurIPS 2024 · similarity 0.89)
- ToolACE: Winning the Points of LLM Function Calling (ICLR 2025 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.