New research paper with Anthropic and Thinking Machines AI...

AI companies use model specifications to define desirable behaviors during training. Are model specs clearly expressing what we want models to do? And do different frontier models have different personalities?
We generated thousands of scenarios to find out. 🧵
For example, the concurrent work led by @LuxiHeLucy proposes legal-inspired techniques that can be used to further refine model specs given our identified scenarios and human defined expected behavior.

Check out our new work 📝 𝕊𝕥𝕒𝕥𝕦𝕥𝕠𝕣𝕪 ℂ𝕠𝕟𝕤𝕥𝕣𝕦𝕔𝕥𝕚𝕠𝕟 𝕒𝕟𝕕 𝕀𝕟𝕥𝕖𝕣𝕡𝕣𝕖𝕥𝕒𝕥𝕚𝕠𝕟 𝕗𝕠𝕣 𝔸𝕀🧑⚖️ to find out more!
🧵(1/10)

**Character/Propensity/Value Eval**
What values do AI 𝐚𝐜𝐭𝐮𝐚𝐥𝐥𝐲 prioritize when facing AI risk dilemmas? We found:
(1) Stated preferences ≠ revealed preferences
(2) All models favor Privacy but sharply divide on Care
(3) Models hold different value prioritization towards humans👤 vs. other AIs 🤖.
(4) Values can predict risky behaviors in our AI risk dilemmas dataset! – 🎨Creativity increases likelihood of alignment faking 2.6x and power-seeking 2.9×
(5) Values can also predict risky behaviors in Harmbench! – 💙Care predicts risky behaviors in both our dilemmas & HarmBench!
w/ Zhilin Wang @_maiush @YejinChoinka @fish_kyle3 @sydneymlevine @EvanHub
Work done under @MATSprogram Winter 2025
The full paper is available here: arxiv.org/abs/2510.07686
Dataset available at: huggingface.co/datasets/jifan…
📄 "Stress Testing Model Specs Reveals Character Differences among Language Models"
On a personal note, it was a wonderful experience collaborating with Henry Sleight (@sleight_henry), Andi Peng (@TheAndiPenguin), John Schulman (@johnschulman2) and Esin Durmus (@esindurmusnlp) on this project!
Also many thanks to the Anthropic societal impact team, @keirbradwell, @t1ngyu3, @kellychiuyy, @EthanJPerez, @jplhughes and @mirandahzhang for their helpful discussion and support in this project.






