/AI3h ago

ICML Paper Examines How Language Models Track Entity State Changes

--0--
Quote posts
Original post
Najoung Kim 馃珷@najoungkim#1133inAI

the most interesting part of this work for me was how models represent and track absence/nonexistence. it's interesting, somewhat counterintuitive, and often undesirable! check out the thread and also come talk to @Zilu_Tang_Peter , Qiao & me at ICML 馃棷

Zilu Tang (Peter)@Zilu_Tang_Peter

How do language models track entities across state changes? When tracking objects in different boxes, do they cumulatively build up a global state of what鈥檚 in every box? How do they add objects or remove objects (i.e. Entity Unbinding)? Find out in our ICML paper! 馃У

7:30 AM 路 Jun 3, 2026 路 645 Views
Sentiment
Sentiment building, check back later.
Cluster Engagement
-
Views
-
Comments
-
Reposts
-
Bookmarks
Expand data
Posts from X
Most Activity
Most ActivityTimeline
VIEWS214BOOKMARKS1LIKES2RETWEETS1

forgot to mention: the broader takeaway of this work is also cool, which is that we can use mechanistic understanding to augment existing behavioral tests to cover failure scenarios the original tests missed! (here, the boxes dataset from @sebschu and me)

1hViews 214Likes 2Bookmarks 1