收录解读
This paper focuses on a real bottleneck in tool-use agents: tool documentation is usually written for humans, not for LLMs, so tool mastery often breaks down because models do not receive the right operational information in the right form. Instead of treating tool descriptions as fixed, the paper asks whether agents can refine them through interaction.
The core mechanism is DRAFT, a framework that dynamically refines tool documentation using feedback and trials gathered from the LLM’s own interactions with tools. This shifts tool mastery from passive reading of static docs to an active self-improvement loop grounded in execution experience.
For the repository, this is worth collecting because it adds an important dimension to the tool-use line: not only can agents retrieve and call tools, they can also improve the interface through which they understand them. That is a durable capability-extension pattern with clear spillover to enterprise tools, API ecosystems, and long-running agents.
It is not ranked higher because the contribution is narrower than infrastructure-scale works such as ToolLLM and more interface-specific than the broadest route changes in agent systems. But it is a strong and reusable tool-learning pattern.