FTF-BENCH
收藏资源简介:
FTF-BENCH(First Things First Benchmark)是由上海交通大学与字节跳动联合创建的多模态大语言模型评估基准,旨在测试模型在处理包含必须需求和可选需求的复杂用户指令时的推理能力。数据集包含3,649个图像-需求对,覆盖电商、预订服务和地图/打车三个真实场景,每个样本由界面截图、候选对象及结构化需求组成。其构建采用大模型驱动流水线,经图像收集、需求生成、口语化表达和人工审核确保质量。该基准聚焦于三类任务:唯一答案、多答案和不可回答,用于诊断MLLM在优先级感知推理中的缺陷,并推动服务型智能体泛化能力的提升。
FTF-BENCH (First Things First Benchmark) is a multimodal large language model evaluation benchmark jointly created by Shanghai Jiao Tong University and ByteDance, aiming to evaluate the reasoning capabilities of models when handling complex user instructions that involve both mandatory and optional requirements. The dataset consists of 3,649 image-demand pairs, covering three real-world scenarios: e-commerce, booking services, and map/taxi-hailing. Each sample comprises an interface screenshot, candidate objects, and structured requirements. Its construction follows a large model-driven pipeline, with quality ensured through four stages: image collection, requirement generation, colloquial paraphrasing, and manual review. This benchmark focuses on three task categories: single-answer, multi-answer, and unanswerable. It is designed to diagnose the flaws of multimodal large language models (MLLMs) in priority-aware reasoning, and to promote the enhancement of the generalization ability of service-oriented AI agents.
数据集概述:FTF-RL
FTF-RL 是一个面向服务场景的多模态强化学习项目,其核心贡献与数据资源如下:
主要贡献
-
基准与评估框架
构建了一个包含 3,649 个服务导向型多模态问题 的综合性评估框架,用于在结构化用户约束条件下评估模型性能。 -
算法创新:FTF-RL
提出了一种新颖的强化学习方法,显式优化模型对多优先级用户需求(“必须满足”与“最好满足”)的推理过程。 -
泛化能力提升
FTF-RL 在需求遵循型任务上显著提升了任务成功率,同时在标准逻辑与数学基准上也表现出更优的通用性能。
当前状态
- 该项目的 源代码、文档及预训练模型 正在准备中,预计在公告发布后的 两周内 正式开源。

- 1First Things First: Teaching LLM-Based Agents to Prioritize Must-Haves before Nice-to-Haves上海交通大学; 字节跳动 · 2026年



