[Paper Review] Learning to Rewrite Tool Descriptions for Reliable LLM-Agent Tool Use

Ruocheng Guo, Kaiwen Dong+2 more|arXiv (Cornell University)|Feb 23, 2026

Scientific Computing and Data Management0 citations

TL;DR

The paper introduces Trace-Free+, a curriculum-learning framework that improves tool interfaces for LLM-based agents by transferring knowledge from trace-rich training to trace-free deployment, enabling better tool selection and usage for unseen tools.

ABSTRACT

The performance of LLM-based agents depends not only on the agent itself but also on the quality of the tool interfaces it consumes. While prior work has focused heavily on agent fine-tuning, tool interfaces-including natural language descriptions and parameter schemas-remain largely human-oriented and often become a bottleneck, especially when agents must select from large candidate tool sets. Existing approaches to improving tool interfaces rely on execution traces, which are frequently unavailable in cold-start or privacy-constrained settings, and typically optimize each tool independently, limiting scalability and generalization to unseen tools. We propose Trace-Free+, a curriculum learning framework that progressively transfers supervision from trace-rich settings to trace-free deployment, encouraging the model to abstract reusable interface-usage patterns and tool usage outcomes. To support this approach, we construct a large-scale dataset of high-quality tool interfaces using a structured workflow over a diverse collection of tools. Experiments on StableToolBench and RestBench show consistent gains on unseen tools, strong cross-domain generalization, and robustness as the number of candidate tools scales to over 100, demonstrating that tool interface optimization is a practical and deployable complement to agent fine-tuning.

Motivation & Objective

• Improve the quality and generalizability of tool interfaces (descriptions and parameter schemas) for LLM-based tool-use agents.
• Enable robust tool selection and parameter generation in cold-start and privacy-constrained settings.
• Develop a scalable data synthesis workflow to produce high-quality tool interfaces across many tools.
• Demonstrate generalization across domains and scalability as candidate tool sets grow beyond 100.

Proposed method

• Construct a large, high-quality dataset of tool interfaces using a structured, agentic workflow over real-world tools (ToolBench seeds refined for health and completeness).
• Synthesize multi-step, dependency-aware user queries to reveal inter-tool usage patterns and failures.
• Train open-weight LLMs as tool description generators using a two-stage description refinement (D0 -> D1 general improvements; D1 -> D2 trace-based refinements via RIMRULE), enabling trace-based and trace-free inference.
• Apply curriculum learning to train models on both trace-rich and trace-free data, gradually increasing reliance on trace-free supervision (Trace-Free+).
• Evaluate under trace-free and trace-based settings with teacher-forcing protocols on RestBench and StableToolBench, measuring subtask-, query-, and tool-level metrics.

Figure 1 : An illustration of the proposed tool interface improvement pipeline. Compared to the original description ( $D0$ ), the learned description generator produces more effective tool descriptions that lead to better tool usage.

Experimental results

Research questions

• RQ1 Can trace-free training transfer the benefits of trace-based supervision to unseen tools during inference?
• RQ2 Do curriculum-learning strategies improve generalization and robustness as tool candidate sets scale?
• RQ3 How well can learned tool-description generators perform under trace-free conditions compared to trace-based baselines and prompt-based methods?
• RQ4 Are the improvements consistent across in-domain and cross-domain tool sets?
• RQ5 What is the impact of tool description quality on tool selection and API execution success for multi-hop tasks?

Key findings

• Trace-Free+ consistently improves subtask- and query-level success rates on unseen tools compared with trace-free and some baselines.
• Trace-Free+ outperforms D1 on harder, multi-hop subsets, indicating value of trace-informed curriculum in learning inter-tool dependencies.
• Trace-Free+ achieves strong cross-domain generalization, improving performance on RestBench (TMDB/Spotify) when trained on StableToolBench Split B.
• Trace-Free+ shows robustness as the number of candidate tools grows beyond 100, with smaller performance degradation than baselines.
• Trace-based models gain more from tool usage patterns in traces, but trace-free curriculum still yields competitive results under cold-start constraints.

Figure 2 : The SFT data synthesis pipeline.