←
Robert Bamler34 posts
Robert Bamler avatar

Robert Bamler

@robamler

Professor of Data Science and Machine Learning at @uni_tue, member of @ml4science and Tübingen AI Center.

◷ Joined October 2011▧ 4 Photos and videos
Robert Bamler avatar
Robert Bamler@robamler·13 Jun 2025

Great paper by my students @TimZXiao and @johanneszenn and collaborators that applies ideas from Monte Carlo sampling to (black-box) LLM execution to turn LLMs into better calibrated stochastic samplers.

✨ New paper: Flipping Against All Odds We found that large language models (LLMs) can describe probabilities—but fail to sample from them faithfully. Yes, even flipping a fair coin is hard. 🪙 🧵 Here’s what we learned—and how we fixed it. 🔗arxiv.org/abs/2506.09998 1/

Post attachment
Tim Xiao avatar
⟳ Robert Bamler retweeted
Tim Xiao@TimZXiao·13 Jun 2025

✨ New paper: Flipping Against All Odds We found that large language models (LLMs) can describe probabilities—but fail to sample from them faithfully. Yes, even flipping a fair coin is hard. 🪙 🧵 Here’s what we learned—and how we fixed it. 🔗arxiv.org/abs/2506.09998 1/

Post attachment
Tim Xiao avatar
⟳ Robert Bamler retweeted
Tim Xiao@TimZXiao·8 Jun 2024

🤔What about using an LLM as a function approximator for f(x; θ) where the parameters θ are natural language? 🤔Can we learn θ just like in machine learning (ML) where θ are numerical values? ✨Check out Verbalized ML, where data and models both operate in natural language! 🤩

Robert Bamler avatar
Robert Bamler@robamler·14 Dec 2023

Did you know that the training/test set split of the SVHN data set is biased, making SVHN unsuitable for evaluating generative models? Learn more from my students @TimZXiao and @johanneszenn at the DistShift workshop at #NeurIPS2023 tomorrow (10.30 am, room R06-R09).

🚨The training and test set of the Street View House Numbers (SVHN) dataset are NOT from the same distribution!🚨 Join us at the #NeurIPS2023 workshop on DistShift this Friday (10:30 am, room R06-R09) to find out more! arxiv.org/abs/2312.02168 w/ @johanneszenn @robamler

Post attachment
Johannes Zenn avatar
⟳ Robert Bamler retweeted
Johannes Zenn@johanneszenn·29 Apr 2023

There is no need for gradients due to resampling in Differentiable Sequential Monte Carlo Samplers! Check out our recent work (arxiv.org/abs/2304.14390) with @robamler and meet us at the poster on Friday!

My student @johanneszenn found a useful fact about differential sequential Monte Carlo samplers: you can ignore any gradients due to resampling because they vanish in expectation. Check out his accepted ICLR DEI paper and meet us at the poster on Friday. arxiv.org/abs/2304.14390

Robert Bamler avatar
Robert Bamler@robamler·22 Mar 2023

We just got the green light to hold another workshop on machine-learning-based data compression—this time at ICML. I'm very excited! Stay tuned for details and for the call for papers.

🎉Exciting news! Our "Neural Compression" workshop proposal has been accepted at #ICML 2023! Join us to explore the latest research developments, including perceptual losses and more compute-efficient models! @BerivanISIK, @YiboYang, @_dsevero, @karen_ullrich, @robamler

Robert Bamler avatar
Robert Bamler@robamler·10 Feb 2023

Looking for the one VAE to rule them all? The bad news: it doesn't exist. The good news: our recently accepted ICLR 2023 paper shows how to optimally allocate information to each latent layer depending on your application: arxiv.org/abs/2302.04855 @TimZXiao @ml4science #ICLR2023