RSS

Fabien Roger

Karma: 8,041

I am working on empirical AI safety.

Anonymous feedback form.

Poi­son­ing Fine-tun­ing Datasets of Con­sti­tu­tional Classifiers

Chase Bowers and Fabien Roger
29 Apr 2026 17:04 UTC
14 points
1 comment11 min readLW link
(alignment.anthropic.com)

Con­trol pro­to­cols don’t always need to know which mod­els are scheming

Fabien Roger26 Apr 2026 19:16 UTC
38 points
1 comment6 min readLW link

Nar­row Se­cret Loy­alty Dodges Black-Box Audits

Alfie Lamerton and Fabien Roger
22 Apr 2026 9:41 UTC
48 points
1 comment13 min readLW link

How Un­mon­i­tored Ex­ter­nal Agents can Sab­o­tage AI labs

Elle Najt and Fabien Roger
9 Apr 2026 18:07 UTC
18 points
0 comments9 min readLW link

Mea­sur­ing and im­prov­ing cod­ing au­dit re­al­ism with de­ploy­ment resources

Connor Kissane, Monte M and Fabien Roger
23 Mar 2026 17:20 UTC
43 points
1 comment10 min readLW link
(alignment.anthropic.com)

Self-At­tri­bu­tion Bias: When AI Mon­i­tors Go Easy on Themselves

Dipika Khullar, Jack Hopkins and Fabien Roger
6 Mar 2026 21:54 UTC
43 points
5 comments6 min readLW link

Tools to gen­er­ate re­al­is­tic prompts help sur­pris­ingly lit­tle with Petri au­dit realism

Connor Kissane, Monte M and Fabien Roger
1 Mar 2026 8:18 UTC
44 points
2 comments7 min readLW link

3 Challenges and 2 Hopes for the Safety of Un­su­per­vised Elicitation

Callum Canavan, Aditya Shrivastava, Allison Qi, Jonathan Michala and Fabien Roger
27 Feb 2026 17:25 UTC
21 points
0 comments10 min readLW link

Re­fusals that could be­come catastrophic

Fabien Roger30 Jan 2026 4:12 UTC
84 points
12 comments7 min readLW link

Elic­it­ing base mod­els with sim­ple un­su­per­vised techniques

Callum Canavan, Aditya Shrivastava, Allison Qi, Tianyi (Alex) Qiu, Jonathan Michala and Fabien Roger
23 Jan 2026 18:06 UTC
34 points
2 comments8 min readLW link

Should con­trol down-weight nega­tive net-sab­o­tage-value threats?

Fabien Roger16 Jan 2026 4:18 UTC
35 points
0 comments10 min readLW link

Towards train­ing-time miti­ga­tions for al­ign­ment fak­ing in RL

Vlad Mikulik, gasteigerjo, Hoagy, Joe Benton, Benjamin Wright, Jonathan Uesato, Monte M, Fabien Roger and evhub
16 Dec 2025 21:01 UTC
39 points
1 comment5 min readLW link
(alignment.anthropic.com)

Eval­u­at­ing hon­esty and lie de­tec­tion tech­niques on a di­verse suite of dishon­est models

RowanWang, Sam Marks, Johannes Treutlein, evhub and Fabien Roger
25 Nov 2025 19:33 UTC
41 points
0 comments4 min readLW link
(alignment.anthropic.com)

Think­ing about rea­son­ing mod­els made me less wor­ried about scheming

Fabien Roger20 Nov 2025 18:20 UTC
88 points
7 comments12 min readLW link

Steer­ing Lan­guage Models with Weight Arithmetic

Fabien Roger and constanzafierro
11 Nov 2025 16:30 UTC
87 points
5 comments5 min readLW link

Rogue in­ter­nal de­ploy­ments via ex­ter­nal APIs

Fabien Roger and Buck
15 Oct 2025 19:34 UTC
34 points
4 comments6 min readLW link

Cur­rent Lan­guage Models Strug­gle to Rea­son in Ciphered Language

Fabien Roger and Shiyuan Guo
14 Oct 2025 9:08 UTC
78 points
7 comments5 min readLW link

Train­ing Qwen-1.5B with a CoT leg­i­bil­ity penalty

Fabien Roger9 Oct 2025 21:33 UTC
68 points
7 comments4 min readLW link

Train­ing fails to elicit sub­tle rea­son­ing in cur­rent lan­guage models

mishajw, Fabien Roger, Hoagy, gasteigerjo, Joe Benton and Vlad Mikulik
9 Oct 2025 19:04 UTC
49 points
3 comments4 min readLW link
(alignment.anthropic.com)

Inoc­u­la­tion prompt­ing: In­struct­ing mod­els to mis­be­have at train-time can im­prove run-time behavior

Sam Marks, Nevan Wichers, Daniel Tan, Aram Ebtekar, Jozdien, David Africa, Alex Mallen and Fabien Roger
8 Oct 2025 22:02 UTC
176 points
37 comments2 min readLW link