Fabien Roger

Karma: 7,438

I am working on empirical AI safety.

Book a call with me if you want advice on a concrete empirical safety project.

Anonymous feedback form.

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

30 points

0 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

19 points

0 comments10 min readLW link

Refusals that could become catastrophic

Fabien Roger30 Jan 2026 4:12 UTC

71 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

33 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

40 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

80 points

7 comments12 min readLW link

Steering Language Models with Weight Arithmetic

Fabien Roger and constanzafierro

11 Nov 2025 16:30 UTC

82 points

2 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

173 points

37 comments2 min readLW link

Four places where you can put LLM monitoring

Fabien Roger and Buck

9 Aug 2025 23:10 UTC

49 points

0 comments7 min readLW link

Why Do Some Language Models Fake Alignment While Others Don’t?

abhayesian, John Hughes, Alex Mallen, Jozdien, janus and Fabien Roger

8 Jul 2025 21:49 UTC

158 points

14 comments5 min readLW link

(arxiv.org)

What can be learned from scary demos? A snitching case study

Fabien Roger24 Jun 2025 8:40 UTC

22 points

1 comment7 min readLW link

Modifying LLM Beliefs with Synthetic Document Finetuning

RowanWang, Johannes Treutlein, Avery, Ethan Perez, Fabien Roger and Sam Marks

24 Apr 2025 21:15 UTC

70 points

12 comments2 min readLW link

(alignment.anthropic.com)

Reasoning models don’t always say what they think

Joe Benton, Ethan Perez, Vlad Mikulik and Fabien Roger

9 Apr 2025 19:48 UTC

28 points

4 comments1 min readLW link

(www.anthropic.com)

Alignment Faking Revisited: Improved Classifiers and Open Source Extensions

John Hughes, abhayesian, Akbir Khan and Fabien Roger

8 Apr 2025 17:32 UTC

147 points

20 comments12 min readLW link