iiinova — where your digital self keeps evolving

Antonia Wüst @toniwuest.bsky.social · 30/09/2026

In the human study, we found that participants could identify compositional rules that the models failed to recover. It would be interesting to see if you can reproduce it in your lego setup!

010

Antonia Wüst @toniwuest.bsky.social · 30/09/2026

Interesting! If you describe the observations in text, I think words vs. zendo should not make too much difference, since the overall task is similar. Compositionality definitely makes the task more challenging, as in your red block example

110

Antonia Wüst @toniwuest.bsky.social · 27/09/2026

We also evaluated Opus (4.8 at that time) btw! It's in the appendix 🤗

010

Antonia Wüst @toniwuest.bsky.social · 27/09/2026

Thanks! The visual component definitely makes it more challenging, presenting the scenes symbolically to the agent improves their performance. But they still struggle with more complex, compositional rules, don't know what kind of rules you tested :)

210

Antonia Wüst @toniwuest.bsky.social · 26/09/2026

It explicitly tests induction of complex rules over visual scenes, so inductive reasoning (the most interesting in my opinion)

010

Antonia Wüst @toniwuest.bsky.social · 25/09/2026

Paper resources: 📄 arxiv.org/abs/2607.08233 🌐 ml-research.github.io/ZendoWorld/ Huge thanks to my wonderful coauthors! 🙏

t.co

https://arxiv.org/abs/2607.08233

010

Antonia Wüst @toniwuest.bsky.social · 25/09/2026

Excited to share that our paper ZendoWorld has been accepted to #NeurIPS2026! 🎉 ZendoWorld tests interactive visual reasoning and demonstrates that current agents (VLM and NeSy) still struggle with both rule induction and informative experimentation 🤖💭

2156

Antonia Wüst @toniwuest.bsky.social · 21/07/2026

Thrilled to be in beautiful Rio this week for #CogSci2026! 🥥🌴🇧🇷 I'll be giving a talk tomorrow at 11:05 AM at the Foundations for Foundation Models workshop on improving visual reasoning of VLMs via program synthesis. Looking forward to great discussions!

031

Antonia Wüst @toniwuest.bsky.social · 14/07/2026

We find that ZendoWorld remains an open challenge: the best model reaches just 45.5% accuracy across the 22 games. The gap to humans is especially large on harder tasks involving compositional concepts. Precisely where systematic reasoning and experimentation matter most!

111

Antonia Wüst @toniwuest.bsky.social · 14/07/2026

Insight #2: VLM agents don’t know how to experiment When allowed to query the world, they often propose near-uninformative scenes that barely reduce uncertainty. They can reason from evidence, but struggle to ask the most useful next question.

100

Antonia Wüst @toniwuest.bsky.social · 14/07/2026

Insight #1: Accuracy ≠ understanding An agent can classify every observed scene correctly while inferring the wrong underlying rule. If you evaluate only prediction accuracy, you risk mistaking a good fit to the examples for genuine induction.

100

Antonia Wüst @toniwuest.bsky.social · 14/07/2026

ZendoWorld is an interactive environment inspired by Zendo, a game of induction. A hidden logical rule governs each scene. The agent observes labeled examples, infers the rule, and proposes new scenes to test it. Perception + induction + active experimentation, in one loop.

100

Antonia Wüst @toniwuest.bsky.social · 14/07/2026

Can AI agents act like scientists when reasoning about visual scenes? 🔬 Not just label what they see, but form explicit hypotheses, design experiments to test them, and revise their beliefs in light of the evidence. We built ZendoWorld to find out!

182

Antonia Wüst @toniwuest.bsky.social · 03/06/2026

Also, if you are interested in visual concept learning, program synthesis and neuro-symbolic AI, I'm happy to meet and have a chat! ☺️

000

Antonia Wüst @toniwuest.bsky.social · 03/06/2026

I'm in Denver at #CVPR2026 this week! 🏔️ Together with @wolfstammer.bsky.social, I will be presenting our recent work, Vision-Language Programs, at Poster Session 3 on Saturday ⏰ 11:45 am - 1:45 pm. Very excited to share our work and discuss the ideas behind it!

161

Reposted by Antonia Wüst

Charley Wu @thecharleywu.bsky.social · 23/03/2026

🚨Last week to apply for a 3 year postdoc with me and @dominikdeffner.bsky.social, embedded in theadaptivemind-excellencecluster.de Focus is on developing innovative experiments and computational models to understand social & cultural learning. Deadline is March 30th 👉 hmc-lab.com/SocialLearni...

hmc-lab.com

Postdoc position -- Social Learning and Cultural Evolution

Postdoc position -- Social Learning and Cultural Evolution posted on March 2, 2026 We are currently seeking a highly motivated individual...

02627

Antonia Wüst @toniwuest.bsky.social · 13/03/2026

We are extending the deadline to submit abstracts for the poster session to 21 March 2026 (AoE)! 🗓️ We invite researchers to present recent work on topics regarding human and artificial creativity. Previously published work is welcome 🤗 📥 Register and submit your abstract here: eveeno.com/342278190

eveeno.com

RMU AI and Creativity Symposium

097

Antonia Wüst @toniwuest.bsky.social · 25/02/2026

Excited to share that our paper "Synthesizing Visual Concepts as Vision-Language Programs" has been accepted to #CVPR2026! 🎉 We propose a novel method that combines VLMs with symbolic program synthesis to learn reliable programs of visual concepts. 🌐 ml-research.github.io/vision-langu...

132

Antonia Wüst @toniwuest.bsky.social · 10/02/2026

We are pleased to announce a one-day symposium on AI and Creativity in Darmstadt! Join us for an inspiring lineup of speakers and a full day dedicated to exploring creativity in modern machine learning models and the relationship between biological and artificial creation. 🎨🤖

0102

Antonia Wüst @toniwuest.bsky.social · 21/01/2026

Super excited that our recent work got featured in the Abstract Synthesis podcast! 🚀 I joined Brian to discuss inductive reasoning in vision and how we can combine Vision-Language Models with Program Synthesis to enable more reliable and interpretable reasoning 💡 Podcast: youtu.be/uefqvsButp8?...

youtu.be

Vision-Language Programs - Antonia Wüst

YouTube video by Ndea

071

Antonia Wüst @toniwuest.bsky.social · 01/12/2025

Thanks! The functions (like exist_object, get_objects) are predefined, however the symbols like "round" in this case are discovered by the VLM. By that we get an expressive DSL that still can adapt to the tasks at hand

010

Antonia Wüst @toniwuest.bsky.social · 01/12/2025

Thank you, very happy to hear that! I think there are definitely merits in combining the best of both worlds, that should not be overshadowed by the current focus on LLMs. Happy to discuss the topic!

130

Antonia Wüst @toniwuest.bsky.social · 01/12/2025

2 Dec 📌 Poster: Vision-Language Programs at WIML Workshop 🕕 6:00 pm to 9:00 pm 4 Dec 📌 Poster: Object-Centric Concept-Bottleneck with David Steinmann 🕟 4:30 pm to 7:30 pm 7 Dec 🎤 Oral: Vision-Language Programs at 01:30 pm 📌 Poster: 4:05 pm to 5:00 pm

010

Antonia Wüst @toniwuest.bsky.social · 01/12/2025

After an amazing time in LA and Joshua Tree Park, I’m excited to head to NeurIPS next week. My colleagues and I will be presenting some of our recent work (see below). Looking forward to connecting and starting new conversations. Feel free to reach out if you want to chat! 💬

120

Antonia Wüst @toniwuest.bsky.social · 30/11/2025

With VLP, we introduce VLM functions as a perceptual interface and combine them with symbolic operators. By that, VLP can discover concise concepts in the form of functional programs that faithfully follow the few-shot image examples.

250

Antonia Wüst @toniwuest.bsky.social · 30/11/2025

Instead of letting the VLM do all the reasoning in natural language, what if we only use it for perception, and then let a symbolic program do the reasoning on top of that? 💡

130

Antonia Wüst @toniwuest.bsky.social · 30/11/2025

Problem: Vision-language models are great at visual recognition, but often fail at faithful visual reasoning. They can output rules that sound plausible but violate the task constraints or contradict the images.

130

Antonia Wüst @toniwuest.bsky.social · 30/11/2025

🚨 New paper alert! We introduce Vision-Language Programs (VLP), a neuro-symbolic framework that combines the perceptual power of VLMs with program synthesis for robust visual reasoning.

1157

Reposted by Antonia Wüst

Martin Trapp @trappmartin.eurosky.social · 18/09/2025

Unfortunately, our submission to #NeurIPS didn’t go through with (5,4,4,3). But because I think it’s an excellent paper, I decided to share it anyway. We show how to efficiently apply Bayesian learning in VLMs, improve calibration, and do active learning. Cool stuff! 📝 arxiv.org/abs/2412.06014

arxiv.org

Post-hoc Probabilistic Vision-Language Models

Vision-language models (VLMs), such as CLIP and SigLIP, have found remarkable success in classification, retrieval, and generative tasks. For this, VLMs deterministically map images and text descripti...

25016

Antonia Wüst @toniwuest.bsky.social · 20/08/2025

And last but not least: the spirals are still spinning, each in their own direction 🌀

010

Antonia Wüst @toniwuest.bsky.social · 20/08/2025

📊 Updated results are also on our webpage! Link: ml-research.github.io/bongard-in-w... Curious to hear - should we evaluate other models too? 🤖

ml-research.github.io

Bongard in Wonderland

100

Antonia Wüst @toniwuest.bsky.social · 20/08/2025

🔎 Importantly, Task 2 continues to expose inconsistencies between the solved problems in Task 1 (64) and the problems where the model can correctly classify the individual images of the problem (only 34), given the gt options (Task 2).

100

Antonia Wüst @toniwuest.bsky.social · 20/08/2025

🤔 Surprisingly, even some easy problems like BP8 remain unsolved…

100

Antonia Wüst @toniwuest.bsky.social · 20/08/2025

Can the new GPT-5 model finally solve Bongard Problems? 👉Not quite yet! Using our ICML Bongard in Wonderland setup, it solved 64/100 problems - the best score so far! 📈 However, some issues still persist ⬇️

160

Reposted by Antonia Wüst

wolfstammer.bsky.social @wolfstammer.bsky.social · 07/07/2025

Can concept-based models handle complex, object-rich images? We think so! Meet Object-Centric Concept Bottlenecks (OCB) — adding object-awareness to interpretable AI. Led by David Steinmann w/ @toniwuest.bsky.social & @kerstingaiml.bsky.social . 📄 arxiv.org/abs/2505.244... #AI #XAI #NeSy #CBM #ML

0104

Antonia Wüst @toniwuest.bsky.social · 12/07/2025

I'll be at #ICML2025 next week presenting our recent work on VLMs and Bongard Problems! Feel free to reach out, happy to have a chat ☺️

030

Antonia Wüst @toniwuest.bsky.social · 02/05/2025

We also identified 10 particularly challenging Bongard Problems that none of the models could solve under any setting. The challenge remains wide open! 3 examples of the challenging BPs:

121

Antonia Wüst @toniwuest.bsky.social · 02/05/2025

Interestingly, success in solving the BPs (Open Question) doesn't translate to correctly categorizing individual images 👉 the sets of BPs solved in each task are not the same! This suggests that getting the right final answer doesn’t always mean genuine understanding 🤔

111

Antonia Wüst @toniwuest.bsky.social · 02/05/2025

Our evaluation shows the top-performing model (o1) solved 43 out of 100 problems, with the others trailing far behind. There’s still a long way to go for current AI models!

101

Antonia Wüst @toniwuest.bsky.social · 02/05/2025

Excited to share that our paper got accepted at #ICML2025!! 🎉 We challenge Vision-Language Models like OpenAI’s o1 with Bongard problems, classic visual reasoning challenges and uncover surprising shortcomings. Check out the paper: arxiv.org/abs/2410.19546 & read more below 👇

12410