BeNice! Closing Looping's Inference-Memory Gap via KV-Sharing

No. 206 of 245 · arXiv · Hayden Prairie, Jacob M. Ryan, Taylor Berg-Kirkpatrick, Daniel Y. Fu · published 23 Oct 2026, 04:00 UTC

AI & machine learning
Paper

2 judges. Scored by two of the three judges. The third reads only the day’s strongest candidates. How far stories are checked

The paper studies how to make looped language models share KV-caches across recurrence depths, which otherwise grow linearly in memory at inference. It proposes stochastic depth during pretraining to induce contractive residual dynamics, plus BeNice!, a post-training method that trains the first recurrence to produce KV-compatible residuals.

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

0.56total under rubric v2.0

Significance0.54
Reader interest0.60
Audience breadth0.41
Practical relevance0.60
Substantiated0.60
Technical depth0.75
New information0.90
Counts against: Promotional0.06
Counts against: Overclaim0.27

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	0.62	0.45	—	0.54	0.17	+0.25	+0.134
Reader interest	0.60	0.60	—	0.60	0.00	+0.20	+0.120
Audience breadth	0.47	0.35	—	0.41	0.12	+0.10	+0.041
Practical relevance	0.69	0.50	—	0.60	0.19	+0.15	+0.089
Substantiated	0.60	0.60	—	0.60	0.00	+0.10	+0.060
Technical depth	0.76	0.75	—	0.75	0.01	+0.10	+0.075
New information	0.94	0.85	—	0.90	0.09	+0.10	+0.090
Promotional	0.07	0.05	—	0.06	0.02	−0.15	−0.009
Overclaim	0.28	0.25	—	0.27	0.03	−0.15	−0.040
Sum of contributions							+0.560
Claim-check penalty							0.000
Total, as published							0.560

Spread is the highest minus the lowest judge answer. marks a spread above 0.40: the judges disagree, and the story is marked contested. Judge 3 did not score this story; its mean uses the other two.

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.