Andy Arditi

Karma: 755

https://andyrdt.com

Finding “misaligned persona” features in open-weight models

Andy Arditi and RunjinChen

9 Sep 2025 14:15 UTC

45 points

5 comments15 min readLW link

Follow-up experiments on preventative steering

RunjinChen and Andy Arditi

6 Sep 2025 4:25 UTC

31 points

1 comment3 min readLW link

Persona vectors: monitoring and controlling character traits in language models

RunjinChen and Andy Arditi

1 Aug 2025 21:19 UTC

25 points

3 comments5 min readLW link

(arxiv.org)

Do models say what they learn?

Andy Arditi, marvinli, Joe Benton and Miles Turpin

22 Mar 2025 15:19 UTC

126 points

12 comments13 min readLW link

Finding Features Causally Upstream of Refusal

Daniel Lee, Eric Breck and Andy Arditi

14 Jan 2025 2:30 UTC

54 points

5 comments12 min readLW link

AI as systems, not just models

Andy Arditi21 Dec 2024 23:19 UTC

29 points

0 comments7 min readLW link

(andyrdt.com)

Unlearning via RMU is mostly shallow

Andy Arditi and bilalchughtai

23 Jul 2024 16:07 UTC

55 points

4 comments6 min readLW link

Refusal in LLMs is mediated by a single direction

Andy Arditi, Oscar Obeso, Aaquib111, wesg and Neel Nanda

27 Apr 2024 11:13 UTC

252 points

95 comments10 min readLW link

Refusal mechanisms: initial experiments with Llama-2-7b-chat

Andy Arditi and Oscar Obeso

8 Dec 2023 17:08 UTC

82 points

7 comments7 min readLW link