adrilmanurung/momento
收藏资源简介:
Momento是一个基准数据集,用于评估基于大型语言模型(LLM)的代理在跨多个会话的对话中,通过工具介导完成任务的持久记忆和推理能力。任务基于餐厅服务领域,要求代理回忆过去的用户偏好、解决跨多个交互的目标、以正确顺序调用正确工具,并遵守领域策略约束。与单轮或单会话基准不同,Momento明确测试代理是否能利用长期记忆,基于先前的对话历史在新会话中正确服务用户。数据集包含场景数据(如scenarios.jsonl),每个任务包括任务ID、用户ID、指令、会话历史、工具调用DAGs、图像引用等,支持多模态和代理评估。
Momento is a benchmark dataset for evaluating LLM-based agents on persistent, tool-mediated task completion across multiple conversational sessions. Tasks are grounded in a restaurant service domain and require agents to recall past user preferences, resolve goals that span multiple interactions, and invoke the correct tools in the correct order, all while respecting domain policy constraints. Unlike single-turn or single-session benchmarks, MOMENTO explicitly tests whether agents can leverage long-term memory to serve users correctly in new sessions based on prior conversation history. The dataset includes scenario data (e.g., scenarios.jsonl), with each task comprising task ID, user ID, instruction, session history, tool call DAGs, image references, etc., supporting multimodal and agentic evaluation.





