tenurehq/precisionmembench
收藏资源简介:
PrecisionMemBench是一个用于大型语言模型(LLM)内存系统的多维度检索基准。它测量了四个正交属性,这些属性是单轮答案质量基准无法检测的:1. 检索精度:在固定的35个信念种子语料库中,正确的信念是否浮现,且仅浮现该信念,涵盖两个领域范围、一个超链替换链和一个次要用户固定装置;2. 噪声隔离:在10轮会话中,离题漂移轮次引入的信念是否污染后续无关轮次的检索;3. 会话轮次延迟:在会话负载下,检索延迟是否相对于单轮基线下降;4. 信念可变性:在会话中更新的信念是否通过别名丰富飞轮在同一会话中立即浮现。这些属性是独立的,系统可能在某个属性上通过而在另一个上失败。基准包含89个案例,覆盖别名解析、范围消歧、超链替换链排除、模糊匹配、跨用户隔离、预算驱逐、排名稳定性、多轮主题漂移下的会话级噪声隔离等方面。数据集旨在评估LLM内存系统的检索性能,特别关注精度和噪声控制,适用于文本检索、文本分类、事实核查等任务。
PrecisionMemBench is a multi-dimensional retrieval benchmark for LLM memory systems. It measures four orthogonal properties that single-turn answer-quality benchmarks cannot detect: retrieval precision, noise isolation, session-turn latency, and belief mutability. These properties are independent. A system can pass on precision and fail on drift. A system can have clean single-turn latency and degrade 4x under session load. A system with no write-time mutation primitive cannot be scored on the fourth property at all, it is an architectural absence, not a performance difference. Every case specifies not just what the memory system must return, but what it must not. Noise is a hard failure, not an invisible inference cost. 89 cases covering: alias resolution, scope disambiguation, supersession chain exclusion, fuzzy matching, cross-user isolation, budget eviction, ranking stability, session-level noise isolation under multi-turn topic drift.




