STALE: Can LLM Agents Know When Their Memories Are No Longer Valid?
Hanxiang Chao, Yihan Bai, Rui Sheng, Tianle Li, Yushi Sun, [0.6em] Wuhan University, The Chinese University of Hong Kong, The Hong Kong University of Science, Technology, Wuhan, China; Hong Kong, China[0.6em] {chx_whu, yihanbai}@whu.edu.cn , [email protected] , {rshengac, ysunbp}@connect.ust.hk · 2026
Read STALE: Can LLM Agents Know When Their Memories Are No Longer Valid? by Hanxiang Chao, Yihan Bai, Rui Sheng, Tianle Li, Yushi Sun, [0.6em] Wuhan University, The Chinese University of Hong Kong, The Hong Kong University of Science, Technology, Wuhan, China; Hong Kong, China[0.6em] {chx_whu, yihanbai}@whu.edu.cn , [email protected] , {rshengac, ysunbp}@connect.ust.hk (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2605-06527.txt.
Description
Hanxiang Chao, Yihan Bai, Rui Sheng, Tianle Li, and Yushi Sun ask whether language model agents notice when a stored memory has quietly gone stale. Personal agents must revise beliefs when later evidence changes an earlier fact, often without any explicit contradiction. The authors call this implicit conflict and build the STALE benchmark to test it.
STALE holds 400 expert-checked scenarios and 1,200 queries probing whether models resolve outdated states, resist false premises, and apply updates in later behavior, with contexts up to 150 thousand tokens. Even the best model scores only about fifty-five percent overall. A prototype system using structured state consolidation improves on this, pointing to explicit state checks.
Catalog record
| CREATOR | Hanxiang Chao, Yihan Bai, Rui Sheng, Tianle Li, Yushi Sun, [0.6em] Wuhan University, The Chinese University of Hong Kong, The Hong Kong University of Science, Technology, Wuhan, China; Hong Kong, China[0.6em] {chx_whu, yihanbai}@whu.edu.cn , [email protected] , {rshengac, ysunbp}@connect.ust.hk |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 109 KB |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2605.06527