Sid Black

Personascope: Measuring how deeply LLMs adopt personas

by Benji Berczi, Kyuhee Kim, Sid Black, and Cozmin Ududec

Benji Berczi, Kyuhee Kim, James Requeima, Sid Black, Cozmin Ududec This is work done by Benji and Kyuhee during MATS Winter 2026, mentored by Cozmin Ududec, and advised by James and Sid. Figure 1. A model can take on a persona fully in voice while not changing its behaviour at...

Jul 735

When Role-playing, Do Models Believe What They Say?

by Sturb, David Africa, and Sid Black

TL;DR * When a model role-plays a persona, does it only change what it says, or also what it internally represents as true? * To study this, we induce personas in five ways: prompting, in-context learning (ICL), supervised fine-tuning (SFT), Open Character Training (OCT), and Emergent Misalignment (EM). We measure...

Jul 254

Preliminary investigation: KL penalties in RL can increase CoT unfaithfulness

by 7vik, Sid Black, and Joseph Bloom

Authors: Satvik Golechha, Sid Black, Joseph Bloom Work done as part of the Model Transparency team at UK AISI. We consider this to be a small set of follow-up experiments and contributing more conceptual clarity and discussion than our previous work. Executive Summary In our recent work replicating MacDiarmid et...

Jun 3081

Machinic Psychopharmacology: Do LLMs Self-Medicate?

UK AISI, Model Transparency Team Epistemic status: Most experiments were run over a period of ~2-3 days during a hackathon at UK AISI, and were fairly heavily vibe coded. Expect some of this to be rough around the edges. Tl;dr We give two language models (Qwen3-8B and Qwen3-32B) access to...

Jun 10135

vLLM-Lens: Fast Interpretability Tooling That Scales to Trillion-Parameter Models

by Alan Cooney and Sid Black

TL;DR: vLLM-Lens is a vLLM plugin for top-down interpretability techniques[1] such as probes, steering, and activation oracles. We benchmarked it as 8–44× faster than existing alternatives for single-GPU use, though we note a planned version of nnsight closes this gap. To our knowledge it’s also the only tool that supports...

Apr 2376

(Some) Natural Emergent Misalignment from Reward Hacking in Non-Production RL

by 7vik, Sid Black, and Joseph Bloom

Authors: Satvik Golechha*, Sid Black*, Joseph Bloom * Equal Contribution. This work was done as part of the Model Transparency team at the UK AI Security Institute (AISI). Our code is available on GitHub and some model checkpoints and rollouts are available on HuggingFace. Executive Summary In Natural Emergent Misalignment...

Mar 30133

Do LLMs know what they're capable of? Why this matters for AI safety, and initial findings

by Casey Barkan, Sid Black, and Oliver Sourbut

This post is a companion piece to a forthcoming paper. This work was done as part of MATS 7.0 & 7.1. Abstract We explore how LLMs’ awareness of their own capabilities affects their ability to acquire resources, sandbag an evaluation, and escape AI control. We quantify LLMs' self-awareness of capability...

Jul 13, 202553

Sid Black

Sid Black

(Some) Natural Emergent Misalignment from Reward Hacking in Non-Production RL

The Singular Value Decompositions of Transformer Weight Matrices are Highly Interpretable

Current themes in mechanistic interpretability research

White Box Control at UK AISI - Update on Sandbagging Investigations

Sid Black

(Some) Natural Emergent Misalignment from Reward Hacking in Non-Production RL

The Singular Value Decompositions of Transformer Weight Matrices are Highly Interpretable

Current themes in mechanistic interpretability research

White Box Control at UK AISI - Update on Sandbagging Investigations

Personascope: Measuring how deeply LLMs adopt personas

When Role-playing, Do Models Believe What They Say?

Preliminary investigation: KL penalties in RL can increase CoT unfaithfulness

Machinic Psychopharmacology: Do LLMs Self-Medicate?

vLLM-Lens: Fast Interpretability Tooling That Scales to Trillion-Parameter Models

(Some) Natural Emergent Misalignment from Reward Hacking in Non-Production RL

Do LLMs know what they're capable of? Why this matters for AI safety, and initial findings