Axel Ahlqvist@_axelahlqvist
Joined October 2021
5 Photos and videos
Alignment audits only work if the model can't tell it's being audited. In our new paper, we make Petri audits far more realistic, tripling realism win rate and decreasing verbalized eval awareness. 🧵
4
7
58
8,134
For more information check out the the full paper and our LessWrong post: Full paper: arxiv.org/abs/2609.02302 Less wrong post: lesswrong.com/posts/9DyiNexL… Critique refinement: github.com/AxelAhlqvist1995/… Petri Dish : github.com/meridianlabs-ai/p…

Improving Evaluation Realism with Inference-Time Compute and...

A core obstacle to alignment evaluation is evaluation awareness: capable models can tell when they are being tested rather than deployed, weakening the conclusions a safety evaluation can support....

arxiv.org
1
32
Fixed Less wrong link: lesswrong.com/posts/9DyiNexL…

Improving Audit Realism With Inference Time Compute and Deployment Scaffolds — LessWrong

tl;dr:  • We introduce critique refinement, a method to iteratively improve the realism of auditor outputs in Petri, and DISH (Deployment Imitating S…

lesswrong.com
5
This work was carried out as a collaboration between Meridian Cambridge, @cbai_ai, @AISecurityInst and @anthropic, jointly authored by @_axelahlqvist, @richardguanxyz, @riverapablojuan, Adeline kassler, @mitroitskii, @alexandrasouly, @kaifronsdal, @_robertkirkk and @jplhughes
1
4
294
For coding settings we go further and run the target inside a real agent scaffold (Claude Code, Codex CLI, or Gemini CLI), so the system prompt, tools, and system reminders match deployment exactly. The auditor only simulates the user and tool results. We call this DISH.
1
2
267
Our methods compose. DISH improves realism substantially, and when combined with critique refinement gives the highest realism on every target we tested.
2
283
Load more