EVENT DOSSIER
To address the poor performance of multimodal large language models (MLLMs) in complex structured scenarios, researchers proposed a method called First Things First Reinforcement Learning (FTF-rl). This method utilizes reinforcement learning to explicitly optimize the model’s ability to reason about multiple priority user requirements, focusing on three scenarios: unique solutions for required items, multiple solutions with different rankings, and abstaining from generating responses when no solution exists. The research team evaluated this method on 3,649 carefully constructed problems across real-world service domains such as e-commerce, booking, and ride-hailing, and found that existing state-of-the-art MLLMs suffered catastrophic failures in all scenarios. Experimental results show that FTF-rl significantly improves task success rates compared to strong baseline methods and demonstrates good general effectiveness in common logical and mathematical reasoning tasks such as LogicVista, MathVision, and InfoQA.
针对多模态大语言模型（MLLMs）在复杂结构化需求场景下表现不佳的问题，研究者提出了一种名为 First Things First Reinforcement Learning (FTF-rl) 的方法，旨在让基于 LLM 的代理优先处理…
针对多模态大语言模型（MLLMs）在复杂结构化需求场景下表现不佳的问题，研究者提出了一种名为 First Things First Reinforcement Learning (FTF-rl) 的方法，旨在优化模型对多重优先级用户需求的…
针对多模态大语言模型（MLLMs）在复杂结构化需求场景下表现不佳的问题，研究者提出了一种名为 First Things First Reinforcement Learning (FTF-rl) 的方法，旨在让基于 LLM 的代理优先处理必须项而非锦上添花项。该方法通过强化学习显式优化对多优先级用户需求的推理能力，并在涵盖电商、预订及叫车等真实服务场景的 3,649 个精心构建的问题上进行评估。实验结果显示，现有 MLLMs 在所有需求场景中均出现灾难性失败，而 FTF-rl 相比强基线方法显著提升了任务成功率，并在 LogicVista、MathVision 和 InfoQA 等逻辑与数学推理任务中展现出通用有效性。
针对多模态大语言模型（MLLMs）在复杂结构化需求场景下表现不佳的问题，研究者提出了一种名为 First Things First Reinforcement Learning (FTF-rl) 的方法，旨在优化模型对多重优先级用户需求的推理能力。该方法通过强化学习显式处理三类需求场景：必须满足项唯一确定解、必须满足项有多个解需结合次要项排序，以及必须满足项无解时 abstain（不生成响应）。研究团队在包含电商、预订及地图/网约车等真实服务场景的 3,649 个精心构建问题上对现有 SOTA MLLMs 进行了评估，发现其在所有场景中均存在灾难性失败。实验结果表明，FTF-rl 相比强基线方法显著提升了任务成功率，并在 LogicVista、MathVision 和 InfoQA 等通用逻辑与数学推理任务上…