CogEval-Bench
收藏资源简介:
CogEval-Bench是一个结构诊断基准,旨在评估智能体在连续事件流下是否形成认知结构,而不仅仅是检索事实的能力。该数据集源自论文《CogniFold: Always-On Proactive Memory via Cognitive Folding》的结构评估部分,包含6个场景(学术研究、客户支持、健康历程、新闻流、软件工程师、团队项目),每个场景提供小规模(约50个事件)和中规模(约90个事件)数据。数据生成采用自上而下方法:先人工定义黄金概念图(包括概念、关系、层次结构、预期意图和多跳推理链),然后生成第一人称事件流,注入约12%的干扰事件并进行时间打乱。数据集文件包括gold_graphs/目录(存储参考认知结构JSON文件,详细描述概念、关系、植入链、意图和状态跟踪)和generated/目录(存储生成事件流与问答捆绑包JSON文件,包含事件流和从黄金图衍生的问答对)。评估方向涵盖概念质量(语义对齐)、拓扑结构(图与层次恢复、多跳链遍历)以及压缩与主动性(冗余事件压缩和意图结晶)。该数据集适用于评估和开发展现结构化记忆和认知推理能力的智能体系统。
CogEval-Bench is a structural diagnostic benchmark designed to evaluate whether agents form cognitive structures under continuous event streams, rather than merely measuring their fact retrieval capabilities. This dataset is the structural evaluation component from the paper CogniFold: Always-On Proactive Memory via Cognitive Folding. It includes 6 scenarios (academic research, customer support, health journey, news stream, software engineer, team project), each providing data at two scales (small scale with about 50 events, medium scale with about 90 events). Data generation follows a top-down approach: first, a gold concept graph is manually specified to define concepts, relationships, hierarchical structures, expected emergent intents, and embedded multi-hop reasoning chains; then, first-person event streams are generated from this gold graph, injecting approximately 12% of noise events and temporally shuffling the event stream. The dataset files are divided into two categories: the `gold_graphs/` directory contains JSON files for reference cognitive structures, detailing concepts, relationships, embedded chains, expected intents, state tracking, etc.; the `generated/` directory contains JSON files for generated event streams and question-answer bundles for each scenario and scale, which embed the corresponding gold graph and include event streams (each with fields such as ID, timestamp, source, type, title, description) and various QA pairs derived from the gold graph. The benchmark scores memory systems across three evaluation dimensions: Dimension A assesses concept quality (semantic alignment of concepts); Dimension B evaluates topological structure (recovery of inter-concept graphs and hierarchical structures, as well as multi-hop chain traversal capabilities); Dimension C assesses compression and proactivity (ability to compress redundant events into stable concepts and crystallize intents from converging evidence). The dataset is suitable for evaluating and developing agent systems with structured memory and cognitive reasoning capabilities.




