Timo Kaufmann

Student Researcher at @GoogleDeepMind. PhD Student at LMU Munich (@AIML_LMU). Focus on alignment / human preferences. Opinions my own.

[Timo Kaufmann @timokauf]

Presenting ResponseRank at #NeurIPS2025! Come by poster #405 at 4:30pm today if you're in San Diego 👋

[Timo Kaufmann @timokauf]

Very excited to start my Student Researcher position at @GoogleDeepMind next week! Flying out tomorrow, ready to wield my #GoogleInterns propeller head proudly

[Artificial Intelligence and Machine Learning @ LMU @AIML_LMU]

🎉We are proud of @timokauf, who is off to @GoogleDeepMind, London as a Student Researcher! On Monday he'll begin 6 months on the Amplified Oversight team, which tackles a central challenge in AI safety: how humans can reliably oversee models that are becoming ever more capable🚀

[Timo Kaufmann @timokauf]

@ZacKenton1 @DavidDAfrica @jacob_pfau Thanks for making me aware! Inverse constitutional learning has a lot of overlap with the ICAI + FF combination and understanding motivations would be great for alignment. Right now the method is more focused on expressed traits, but might be extensible (@arduinfindeis)

[Timo Kaufmann @timokauf]

Somebody recorded me at our poster. Cool to have a video! Don't expect a full explanation though, it's just a random excerpt and I had no idea I was being filmed 🙃

[Starc @Starc_Institute]

#NeurIPS2025 has passed, but we hope the celebration will last forever.

Here is a poster presentation we recorded at the event and we hope it can last forever in cyberspace.

The paper: ResponseRank: Data-Efficient Reward Modeling through Preference Strength Learning

Thanks for the authors: @timokauf, @metz_yannick, Daniel Keim, Eyke Hüllermeier

[Timo Kaufmann @timokauf]

@Starc_Institute Thanks for posting this, nice to have a video of the session!

I had no idea I was being filmed. The video starts sort of in the middle of the explanation, so take a look at the paper if you're interested.

Paper: https://t.co/suPIUZpewO
Poster etc.: https://t.co/5gYGIxNaem

[Timo Kaufmann @timokauf]

Had a good time presenting this! (Though who decided poster sessions should go until 7:30pm?) Thanks to Max Muschalik for helping me present even without being an author. Poster presentations really need two people!

[Timo Kaufmann @timokauf]

Presenting ResponseRank at #NeurIPS2025! Come by poster #405 at 4:30pm today if you're in San Diego 👋

[Timo Kaufmann @timokauf]

Joint work with @metz_yannick, Daniel Keim, and @eyke_hu.

[Timo Kaufmann @timokauf]

Presenting ResponseRank at #NeurIPS2025! Come by poster #405 at 4:30pm today if you're in San Diego 👋

[Timo Kaufmann @timokauf]

Benefits: improved reward model generalization, better data efficiency, and improved policies. Looking forward to seeing you at the poster!

Paper and more: https://t.co/5gYGIxNaem

[Timo Kaufmann @timokauf]

The key insight is that these signals only need to be locally valid and relative (e.g., within one annotator's comparisons). No need to model the exact relationship to strength. Just rank which comparisons are stronger.

[Timo Kaufmann @timokauf]

The core idea: Not all preferences are equal. ResponseRank learns preference strength from implicit signals in your data, like inter-annotator agreement, stated confidence, or response times.

[Arduin Findeis @arduinfindeis]

I think Gemini 3 Pro’s personality and style notably improved over 2.5.

It uses fewer common shortcuts to bias human annotators, e.g. long responses or overly polite tone. That makes the strong LMArena performance quite a bit more impressive!

Some examples of the differences 🧵

[Timo Kaufmann @timokauf]

@janleike Thanks for pointing this out! Do you think Anthropic will be able to offer visa sponsorship in future iterations? I'm a German PhD student and would love to apply, but cannot self-sponsor as far as I can tell.

[Timo Kaufmann @timokauf]

Nice demo of feedback forensics, check out our tool and Arduin's analysis!

[Arduin Findeis @arduinfindeis]

How is GPT-5's personality different to GPT-4o?

A quantitative analysis using Feedback Forensics 🧵

[Timo Kaufmann @timokauf]

Just noticed the key deadlines for #ICLR2026 out! PSA for everyone else who's been waiting.

Full paper: Sept 24 AoE.

[Timo Kaufmann @timokauf]

This was a lot of fun!

[Artificial Intelligence and Machine Learning @ LMU @AIML_LMU]

@arduinfindeis and our lab member @timokauf presented Inverse Constitutional AI. Had a blast, the pictures speak for themselves! Joint work with @eyke_hu, @SamuelAlbanie, @RobDMullins.
Paper📰https://t.co/49eGzYchQs
GitHub: https://t.co/dMlHYlCl5r
🧵2/4

[Timo Kaufmann @timokauf]

Looking forward to present tomorrow. Come by our poster #520 if you're at ICLR!

[Arduin Findeis @arduinfindeis]

Excited to be in Singapore for ICLR! Keen to chat about interpreting feedback data and detecting model characteristics ⚖️

Reach out or come by our poster on Inverse Constitutional AI tomorrow, Friday 25 April from 10am-12.30pm (#520 in Hall 2B) - @timokauf and I will be there!

[Timo Kaufmann @timokauf]

Currently visiting @arduinfindeis in Cambridge. I didn't realize it's this beautiful!

Do I know anyone here that I haven't met up with yet?

[Timo Kaufmann @timokauf]

Ever wondered how models on chatbot arena differ? Feedback forensics gives some answers, check it out!

[Arduin Findeis @arduinfindeis]

🤖 3. Discovering model strengths

How is GPT-4o different to other models? → Uses more numbered lists, but Gemini is more friendly and polite

https://t.co/bDCQXGtRFW

[Arduin Findeis @arduinfindeis]

🕵🏻💬 Introducing Feedback Forensics: a new tool to investigate pairwise preference data.

Feedback data is notoriously difficult to interpret and has many known issues – our app aims to help!

Try it at https://t.co/4HubCg52Pi

Three example use-cases 👇🧵

[Timo Kaufmann @timokauf]

Our paper on query-efficient reward learning will be at AAAI! Unfortunately I won’t be attending, but Xuening will be at the poster - stop by or reach out online!

[Artificial Intelligence and Machine Learning @ LMU @AIML_LMU]

🚀 Excited to share our latest work at #AAAI2025!

How can we make Reinforcement Learning from Human Feedback (RLHF) more query-efficient? Introducing DUO: Diverse, Uncertain, On-policy query generation and selection.
🧵👇 1/6