SebastianP

Karma: 348

The distillation double bind: Distilling misaligned models either transfers misalignment or it doesn’t

Alek Westover, SebastianP, Alexa Pan and Jozdien

18 Jun 2026 21:21 UTC

57 points

4 comments5 min readLW link

(blog.redwoodresearch.org)

How to reduce capability degradation from off-model SFT

Dylan Xu, SebastianP and Alek Westover

8 Jun 2026 16:24 UTC

21 points

0 comments3 min readLW link

Advice for making robust-to-training model organisms

SebastianP, Alek Westover, Vivek Hebbar, Julian Stastny and Dylan Xu

28 May 2026 17:26 UTC

38 points

8 comments12 min readLW link

(blog.redwoodresearch.org)

Why does off-model SFT degrade capabilities?

SebastianP, Dylan Xu, Alek Westover, Julian Stastny and Vivek Hebbar

21 May 2026 0:35 UTC

42 points

9 comments6 min readLW link

Incriminating misaligned AI models via distillation

Alek Westover, SebastianP, Alex Mallen, Jozdien, Alexa Pan, Julian Stastny and Vivek Hebbar

15 May 2026 21:43 UTC

115 points

12 comments5 min readLW link

Sleeper Agent Backdoor Results Are Messy

SebastianP, Alek Westover, Dylan Xu, Vivek Hebbar and Julian Stastny

28 Apr 2026 1:55 UTC

81 points

4 comments7 min readLW link

An Empirical Study of Methods for SFTing Opaque Reasoning Models

SebastianP, Alek Westover, Vivek Hebbar, Dylan Xu and Julian Stastny

24 Apr 2026 17:26 UTC

17 points

0 comments6 min readLW link

How do LLMs generalize when we do training that is intuitively compatible with two off-distribution behaviors?

Dylan Xu, Alek Westover, Vivek Hebbar, SebastianP, frisby and Julian Stastny

20 Apr 2026 16:58 UTC

62 points

5 comments20 min readLW link

Five approaches to evaluating training-based control measures

Alek Westover, SebastianP, Julian Stastny and Vivek Hebbar

18 Apr 2026 1:07 UTC

22 points

0 comments6 min readLW link

Model organisms researchers should check whether high LRs defeat their model organisms

Dylan Xu, SebastianP, Alek Westover, Vivek Hebbar and Julian Stastny

10 Apr 2026 0:07 UTC

40 points

0 comments5 min readLW link

Basic Legibility Protocols Improve Trusted Monitoring

SebastianP and theashwinner

12 Feb 2026 17:50 UTC

8 points

4 comments11 min readLW link