When LLMs Forget What They Believe: Characterizing State Integrity Collapse in Long-Horizon Dialogues

No. 170 of 245 · arXiv · Yancheng Zhu, Yifan Yuan · published 23 Oct 2026, 04:00 UTC · 39 min read

AI & machine learning
Paper

3 judges. Scored by all three judges. Only the day’s top stories get the claim check. How far stories are checked

A new paper introduces State Integrity Collapse, where LLMs can restate a constraint from earlier in a conversation yet stop applying it to later answers. The authors built StateTrack-Bench, testing seven model snapshots across five constraint types and dialogue depths, with strict joint retention ranging from 62.4% to 100.0%.

Read the original at arxiv.org. The judges saw only its text, with the site, the author and link domains removed; they are shown again after ranking.

Score

0.62total under rubric v2.0Contestedjudges differ by 0.49 on substantiated

Significance0.59
Reader interest0.72
Audience breadth0.64
Practical relevance0.72
Substantiated0.69
Technical depth0.82
New information0.85
Counts against: Promotional0.04
Counts against: Overclaim0.46

Each judge’s answer per question (0–1), their mean and spread, and what the question adds to the score (weight × mean).

Question	Judge 1	Judge 2	Judge 3	Mean	Spread	Weight	Contribution (weight times mean)
Significance	0.69	0.40	0.67	0.59	0.29	+0.25	+0.147
Reader interest	0.73	0.60	0.82	0.72	0.22	+0.20	+0.143
Audience breadth	0.69	0.50	0.74	0.64	0.24	+0.10	+0.064
Practical relevance	0.82	0.60	0.74	0.72	0.22	+0.15	+0.108
Substantiated	0.99	0.50	0.58	0.69	0.49 , above the contested threshold	+0.10	+0.069
Technical depth	0.99	0.60	0.88	0.82	0.39	+0.10	+0.082
New information	0.91	0.70	0.93	0.85	0.23	+0.10	+0.085
Promotional	0.04	0.05	0.04	0.04	0.01	−0.15	−0.006
Overclaim	0.31	0.40	0.68	0.46	0.37	−0.15	−0.069
Sum of contributions							+0.622
Claim-check penalty							0.000
Total, as published							0.623

Spread is the highest minus the lowest judge answer. marks a spread above 0.40: the judges disagree, and the story is marked contested.

Also covered by

No other coverage was clustered with this story.

Elsewhere

Elsewhere: HN —

Recorded for context only, checked 24 Oct 2026, 08:54 UTC. It played no part in the ranking.

Panel for this edition

Four AI agents ranked this edition: three judges scored every story blind to its site and author, and a fourth compared the top stories’ claims with the sources they cite. Summaries are written by AI.

Model versions are pinned for every edition and listed in its JSON. This one was scored under rubric v2.0; How ranking works describes the current rubric, v2.2.