Research Scientist @GoogleDeepMind // Machine Learning (formerly Physics) // Previously @bethgelab @MPI_IS @uni_tue @GoogleAI
San Francisco, CA
Joined August 2009
- Tweets124
- Following500
- Followers499
- Likes4.9K
Roland S. Zimmermann retweeted
for alignment folks, worth taking a look at the hollistic loss of control assessment (3.4.3), the alignment assessment (3.4.4), and chain of thought transparency/monitorability assessment (3.4.5) -- lots of new empirical work there that I'd love your eyes and feedback on
Roland S. Zimmermann retweeted
We did a lot to understand Gemini 4's capabilities and propensities, and to improve our mitigations for this launch -- you can get a tl;dr here blog.google/innovation-and-a… and dive into the report storage.googleapis.com/deepm…
Roland S. Zimmermann retweeted
Excited to launch Gemini 4 Argon, designed to help with hardest coding and knowledge work tasks with its frontier capabilities..
For developers and enterprises, it’s available in the paid tier of the Gemini API and the Gemini platform, for Ultra subscribers in Google Antigravity, and Ultra and Pro subscribers in the Gemini app and Search.
Read more: blog.google/innovation-and-a…
Roland S. Zimmermann retweeted
Frontier models can now detect AI control interventions.
In particular, GPT-6 Astra is incredibly good at this. When we published the initial version of our CIAware benchmark in May most models were not much above chance level, now a step change happened.
I think this is a big problem and challenges the effectiveness of control protocols. High control intervention awareness means that interventions leaks information about the monitor which helps circumventing the monitor.
Details in the thread 🧵
Roland S. Zimmermann retweeted
Introducing Gemini 4 Argon – our new frontier model.
It’s built for complex workflows across coding, enterprise knowledge work, and cybersecurity defense – rolling out today to a set of trusted testers through our Fairwind Program.
Roland S. Zimmermann retweeted
Introducing Gemini 3.8 Flash, another jump in Gemini's agentic + coding capabilities, and our 3rd updated Flash model in only 6 weeks...
This model has been a ton of fun to work with, excited to see what you all think!
Control interventions are a common approach to increase the safety of agents - but how well can they detect such interventions? My MATS fellows investigated this and built a new benchmark for investigating this. Check out @JSchaeff3r's thread on the paper!
Roland S. Zimmermann retweeted
Our MATS 9.0 project with @zimmerrol is out!
We studied whether models can detect interventions on their trajectories from control protocols. So far, they're bad at it.
However, they can achieve non-trivial accuracy by inventing their own watermarking schemes and embedding signals in messages for future detection.
Now since this this paper (and the watermarking strategy) is public, I expect models to get better at it. We hope our benchmark helps track that progress and inform safety teams!
Roland S. Zimmermann retweeted
RL assumes that LLMs explore well during training. What if they choose not to?
In our new ICML paper with @GoogleDeepMind, we train LLMs that strategically resist RL capability elicitation by under-exploring.
We study this threat model, called exploration hacking.
Roland S. Zimmermann retweeted
Is training against the CoT always bad?
RL training can lead to obfuscated CoT making it difficult to 'read an LLMs thoughts'. How can we predict when obfuscation occurs?🤔
Our new @GoogleDeepMind paper introduces a framework to predict this before training starts!
Roland S. Zimmermann retweeted
New DeepMind x UK AISI paper: what would it take to prevent harm from misaligned AI agents via monitoring in real deployments? We wrote a safety case sketch for control monitoring
No flashy results but lots of important details for deploying future AI agents safely!
Roland S. Zimmermann retweeted
Extrajudicial killings of suspected drug dealers, bombing another country and kidnapping its leader w/o congressional approval. Yes, Maduro is a dictator, but without due process and international law being upheld, if it's Venezuela today who's to say it's not Greenland tomorrow?
Roland S. Zimmermann retweeted
I used to be able to claim that tech billionaires didn't actually do this — that they just wanted to refine their gadgets. But unfortunately in the current administration we've seen all three.
Roland S. Zimmermann retweeted
We built and validated an autorater that can provide a leading indicator of CoT illegibility
teapawt.net/emmons_scott/status/19…
Roland S. Zimmermann retweeted
The latest version of our framework expands the scope of safety case reviews and includes a new harmful manipulation CCL.
As we build increasingly powerful AI models, we’re committed to responsible development.
We’re implementing our latest Frontier Safety Framework – our most comprehensive approach yet for identifying and staying ahead of emerging risks.
Find out more → goo.gle/3W1ueFb
Roland S. Zimmermann retweeted
MATS is a great opportunity to start your career in AI safety! For MATS 9.0 I'll be running a research stream together with @emmons_scott @jenner_erik and @zimmerrol If you want to do research on AI oversight and control, apply now!
Roland S. Zimmermann retweeted
the model card builds on foundational research on stealth and situational awareness from our team
led by @MaryPhuong10 @zimmerrol @vkrakovna @davlindner, Ziyue Wang, @sarah_cogan, @AllanDafoe @_lewisho and @rohinmshah
arxiv.org/abs/2505.01420
Roland S. Zimmermann retweeted
Can reasoning models strategically sabotage their own reinforcement learning training by deliberately under-exploring?
I’m currently exploring this question in MATS 8.0, alongside @yoenoo_ and @DamonFalck, supervised by @emmons_scott, @davlindner and @zimmerrol.
