Why Constrained Updates Still Forget: A Functional-Path Analysis of Continual SFT

No. 85 of 245 · arXiv · Qingsen Ma, Dianyun Wang, Jiahong Liu, Irwin King · published 23 Oct 2026, 04:00 UTC · first published 21 Oct 2026 · 88 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

This paper analyzes why constrained updates still cause forgetting during supervised fine-tuning of language models, decomposing old-behavior loss changes into constraint, probe mismatch, gradient staleness, and finite-step terms.

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

0.66total under rubric v2.0Contestedjudges differ by 0.43 on significance, 0.42 on audience breadth

Significance0.59
Reader interest0.66
Audience breadth0.53
Practical relevance0.65
Substantiated0.87
Technical depth0.96
New information0.89
Counts against: Promotional0.03
Counts against: Overclaim0.22

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	0.65	0.35	0.78	0.59	0.43 , above the contested threshold	+0.25	+0.148
Reader interest	0.53	0.55	0.90	0.66	0.37	+0.20	+0.132
Audience breadth	0.58	0.30	0.72	0.53	0.42 , above the contested threshold	+0.10	+0.053
Practical relevance	0.64	0.50	0.80	0.65	0.30	+0.15	+0.097
Substantiated	0.97	0.80	0.84	0.87	0.17	+0.10	+0.087
Technical depth	0.99	0.90	0.98	0.96	0.09	+0.10	+0.096
New information	0.88	0.80	0.98	0.89	0.18	+0.10	+0.089
Promotional	0.04	0.03	0.01	0.03	0.03	−0.15	−0.004
Overclaim	0.25	0.30	0.12	0.22	0.18	−0.15	−0.033
Sum of contributions							+0.665
Claim-check penalty							0.000
Total, as published							0.665

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.