Cao, Schooler and Zafarani compare LLM memory with human memory

Cao, Schooler and Zafarani compare LLM memory with human memory

Authors Zhaoyang Cao, Lael Schooler and Reza Zafarani use human memory research to examine whether large language models show memory effects found in people. They report that both people and models remember less when overloaded with information (list length effect), remember better with repeated exposure (list strength effect), and can confuse overlapping information (fan effect). Both can also show false memories—“remembering” related words that were never presented. The authors report that LLMs can apply prior learning to new, related situations (cross-domain generalization). The authors also report differences: large language models are less influenced by the order in which information is presented (positional bias) and more robust when processing random or meaningless material (nonsense effect). Their argument is not that model memory and human memory are identical: the results show both similarities and divergences in how they reconstruct memory. The comparison may help readers understand particular memory-like behaviors and errors in models, while the authors point to architectural differences as a possible source of distinct patterns. The abstract does not establish how these findings translate to use outside the evaluated paradigms.

Publication: fe1e7299-9d4b-45fa-80e9-288527cac18d

attribution — pass: The reading correctly credits the paper to Zhaoyang Cao, Lael Schooler and Reza Zafarani. It keeps the findings attributed with phrasing such as 'they report' and 'the authors interpret', rather than presenting them as established fact.

Authors:Zhaoyang Cao, Lael Schooler, Reza Zafarani

summary — pass: The reading faithfully covers the seven phenomena, the similarities and the two differences, and each technical term gets a plain-language gloss drawn from the abstract. The question follows from the evidence. Minor nuance: the abstract says architectural differences 'lead to' distinct patterns, while the reading softens this to 'a possible source'. The smallest repair is to restore 'the authors attribute distinct patterns to architectural differences'. This does not prevent a general reader from understanding the development.

while also highlighting the architectural differences that lead to distinct patterns of error and success

date — pass: The displayed original publication date of 2025-09-21 matches the v1 submission date. It is not the date of the later v3 revision (2 Nov 2026) or the date of the feed entry.

[Submitted on 21 Sep 2025 (v1), last revised 2 Nov 2026 (this version, v3)]

support — pass: Every claim appears in the abstract. The third claim's excerpt covers only the differences, not the shared false memories and cross-domain generalization, but those points are stated elsewhere in the abstract. The significance sentence is hedged and makes no unsupported claims about benefits or costs. The limitations correctly note that only the abstract is available.

Like humans, LLMs are susceptible to falsely "remembering" words that were never shown but are related to others (false memories), and they can apply prior learning to new, related situations (cross-domain generalization).

media — pass: The selected media is the verified link to the arXiv abstract page itself. This is the direct original source.