RSS

cloud

Karma: 1,382

[Paper] Out­put Su­per­vi­sion Can Obfus­cate the CoT

jacob_drori, lukemarks, cloud and TurnTrout
20 Nov 2025 22:41 UTC
75 points
2 comments5 min readLW link
(arxiv.org)

Om­nis­cal­ing to MNIST

cloud8 Nov 2025 19:42 UTC
87 points
3 comments10 min readLW link

Re­con­tex­tu­al­iza­tion Miti­gates Speci­fi­ca­tion Gam­ing Without Mod­ify­ing the Specification

ariana_azarbal, Victor Gillioz, TurnTrout and cloud
14 Oct 2025 0:53 UTC
124 points
14 comments9 min readLW link

cloud’s Shortform

cloud17 Sep 2025 7:41 UTC
6 points
1 comment1 min readLW link

Nar­row fine­tun­ing is different

cloud and Stewy Slocum
5 Aug 2025 14:29 UTC
67 points
3 comments4 min readLW link

[Re­search Note] Op­ti­miz­ing The Fi­nal Out­put Can Obfus­cate CoT

lukemarks, jacob_drori, cloud and TurnTrout
30 Jul 2025 21:26 UTC
198 points
23 comments6 min readLW link

Sublimi­nal Learn­ing: LLMs Trans­mit Be­hav­ioral Traits via Hid­den Sig­nals in Data

cloud, mle and Owain_Evans
22 Jul 2025 16:37 UTC
342 points
39 comments4 min readLW link

Selec­tive Gen­er­al­iza­tion: Im­prov­ing Ca­pa­bil­ities While Main­tain­ing Alignment

ariana_azarbal, Matthew A. Clarke, Jorio Cocola, Cailley Factor and cloud
16 Jul 2025 21:25 UTC
67 points
4 comments7 min readLW link

Distil­la­tion Ro­bus­tifies Unlearning

Bruce W. Lee, Addie Foote, alexinf, leni, Jacob G-W, Harish Kamath, Bryce Woodworth, cloud and TurnTrout
13 Jun 2025 13:45 UTC
234 points
43 comments8 min readLW link
(arxiv.org)

Selec­tive mod­u­lar­ity: a re­search agenda

cloud and Jacob G-W
24 Mar 2025 4:12 UTC
66 points
2 comments24 min readLW link

[Question] Is weak-to-strong gen­er­al­iza­tion an al­ign­ment tech­nique?

cloud31 Jan 2025 7:13 UTC
22 points
1 comment2 min readLW link

Gra­di­ent Rout­ing: Mask­ing Gra­di­ents to Lo­cal­ize Com­pu­ta­tion in Neu­ral Networks

cloud, Jacob G-W, Evzen, Joseph Miller and TurnTrout
6 Dec 2024 22:19 UTC
169 points
14 comments11 min readLW link
(arxiv.org)