Patrick Knab

We structure the design space of CBMs, disentangle key components, discuss recurring challenges, and outline a roadmap for how the field could evolve. If we missed any relevant work published up to the end of 2025, please let us know—we will incorporate it in the next version.

Together with David Steinmann , Udo Schlegel, and @WolfStammer , we decided it was time to take a step back: we review and categorize more than 100 papers into a unified architectural taxonomy. In doing so, we try to answer the question: What’s in the bottle?

What’s in the Bottle? A Survey and Roadmap of Concept Bottleneck Models CBMs are a rapidly growing direction in interpretable-by-design machine learning. However, the field has become increasingly fragmented. Preprint Link: researchgate.net/publication/40…

@icmlconf As soon as ICLR decisions are out, amount will roughly decrease by number of accepted papers

The #ICML2026 abstract deadline has passed! We're at 33540 active abstracts (and dropping). How many will make it over the finish line? 🏁

We wish everyone the best for tomorrow’s decisions—and try not to take them too seriously. 🙂 If you’re curious what tabular ML predicts for your paper, you can check it out here: Website: lnkd.in/dxNDPryB Code: lnkd.in/dzXvmaBS

Framing the problem as primarily tabular—using numerical review scores such as soundness and contribution—we evaluated standard tabular models (TabPFN, CatBoost, logistic regression, and decision trees) without any hyperparameter tuning.

We took a quick look at the new PaperDecision results for ICLR 2026 and asked a simple question: do we really need LLMs for this task? We did the same but with tabular data: s-marton.github.io/TabICLR/ #ICLR2026

@dyhTHU Super cool. We have built a similar tool, but based on tabular data. Let's see which of the two works best ;) s-marton.github.io/TabICLR/

📢ICLR2026 Acceptance Prediction is out! 🚀Find the acceptance of your paper in advance (predicted): paperdecision.netlify.app 🛠️Code of Multi-Agent Framework and Benchmark is available: github.com/PaperDecision/… 🎯Our goal is to understand the how and why behind paper decisions.

@HeNordlinder @dyhTHU We build sth similar with more benchmarks and completely based on tabular data ;) s-marton.github.io/TabICLR/

@dyhTHU Do you have a benchmark vs logistic regression on the features you included?

@bfl_ml Congrats! Also a huge milestone for AI in south Germany :) Just out of curiosity ( I haven’t been able to find sth on the website), are you also offering PhD internships?

We've raised $300M in Series B funding from Salesforce Ventures and Anjney Midha (AMP) FLUX is used by millions every month and powers production workflows across the world's leading platforms. This funding will allow us to invest deeply in research and build the foundations for visual intelligence.

@rdesh26 @iclr_conf Sure, totally agree. We'll prepare a brief summary for the AC to summarize the work of our rebuttal, since we had some useful discussions.

@p_knab @iclr_conf Yeah I feel for them, but it seems to me that there could have been better ways than wiping off thousands of person-hours of work and creating more work for ACs (over the holiday season) with this blanket decision.

@iclr_conf is such a shit-show! We spent a lot of time (as I imagine many others) to respond to our reviewers' feedback, who themselves engaged with us very constructively. All of this is going to be wiped off? I'll stick to good old ICASSP/Interspeech next time!

For everyone who doesn’t know it yet — SAM 3 is out! Excited to see what the next generation of Segment Anything brings: openreview.net/forum?id=r35cl… #SAM3 #ICLR2026

We present MoTIF (Moving Temporal Interpretable Framework), a transformer-inspired framework that adapts CBMs for video classification with three perspectives: • Global concepts across the entire video • Local concepts in specific windows • Temporal dependencies of a concept

Concept Bottleneck Models (CBMs) have shown strong potential for interpretable image classification. But bringing them from static images to video data is not straightforward — temporal dependencies are key for capturing actions and events.

🚀 Preprint out: Concepts in Motion: Temporal Bottlenecks for Interpretable Video Classification 📄 Preprint: arxiv.org/pdf/2509.20899 💻 Code: github.com/patrick-knab/M…

We systematically review LIME extensions, introduce a new taxonomy, and offer: 🔹 A structured overview for researchers 🔹 Practical guidance for practitioners 🔹 Open questions for future work 🌐 Website: patrick-knab.github.io/which-lime-to-…

Which LIME should I trust? Our new paper, accepted at XAI 2025 in Istanbul, answers this question! LIME is a go-to for post-hoc explanations—but with so many variants, which one should you use? 🧵👇 📝 Paper: arxiv.org/abs/2503.24365… #XAI2025 #XAI #ML

TL;DR: Improved image classification using a modified LIME framework, enabling user-driven hierarchical feature analysis through the integration of semantic segmentation foundation models.

Excited for insightful discussions and new research—check out our results here: 👇 Project: patrick-knab.github.io/DSEG-LIME/ See you in Singapore! 🇸🇬

Our paper “Beyond Pixels: Enhancing LIME with Hierarchical Features and Segmentation Foundation Models” has been accepted to the Foundation Models in the Wild workshop at ICLR 2025! arxiv.org/pdf/2403.07733 #ICLR2025 #XAI