Tag
Apple's new paper introduces SynthAPI, showing that for tool-use agents, diversity of apps and database states matters more than raw task count, deriving a data-mix scaling law that predicts held-out generalization far better than task-count-only laws.
Trajectory2Gym 提出一种决策落地的组合式交互合成方法用于工具调用智能体的训练监督，在五个 Qwen3 规模上于 Tau1、Tau2 和 BFCL 基准上提升 16.9–29.8 个百分点。
The paper proposes CITA, a method that trains a Comparative Inference Model (CIM) to estimate the long-horizon value of possible next tool invocations before executing them, improving Tool F1 and task success across three tool-use benchmarks and multiple LLM backbones.
SEAD 提出将工具使用智能体的攻击与防御统一为部分可观测状态控制问题，其中 DART 通过分解有害目标并利用工具执行反馈指导轨迹搜索来提升攻击成功率，SAGE 则通过只读查询在执行前检查状态来拦截有害操作，在在线攻防评估中将 DART 的可执行攻击成功率从 48.0% 降至 4.0%。