Is One Score Enough? Rethinking the Evaluation of Sequentially Evolving LLM Memory
The paper introduces SeqMem-Eval, a new evaluation framework for assessing the memory of large language models (LLMs) during sequential tasks. It emphasizes the importance of understanding memory evolution and retention rather than relying solely on final performance metrics. The authors demonstrate that higher accuracy does not always correlate with better memory quality, highlighting the need for more nuanced evaluation methods.
- ▪SeqMem-Eval targets the evaluation of LLM memory in a test-time setting with external, prompt-mediated memory.
- ▪The framework measures online utility, hold-out generalization, backward transfer, and forgetting.
- ▪The study reveals that many high-performing methods suffer from significant forgetting or negative transfer.
Opening excerpt (first ~120 words) tap to expand
Computer Science > Machine Learning arXiv:2605.15384 (cs) [Submitted on 14 May 2026] Title:Is One Score Enough? Rethinking the Evaluation of Sequentially Evolving LLM Memory Authors:Songwei Dong, Zihan Chen, Chengshuai Shi, Peng Wang, Jundong Li, Cong Shen View a PDF of the paper titled Is One Score Enough? Rethinking the Evaluation of Sequentially Evolving LLM Memory, by Songwei Dong and 5 other authors View PDF HTML (experimental) Abstract:Memory plays a central role in enabling large language models (LLMs) to operate over sequential tasks by accumulating and reusing experience over time.
…
Excerpt limited to ~120 words for fair-use compliance. The full article is at arXiv cs.AI.
Discussion
0 commentsMore from arXiv cs.AI
