Playing ZendoWorld: Challenging AI Agents on Active Visual Concept Induction

Jul 16, 2026

ZendoWorld addresses the core challenge of enabling AI agents to jointly perceive complex visual inputs, induce hypotheses about hidden logical rules, and design informative experiments via active scene generation. The work proposes a controlled multi-turn environment inspired by the inductive logic game Zendo, combining raw visual observations with a DSL and game-feedback. Experiments across VLM reasoning, Bayesian particle filtering, dynamic concept discovery, and neuro-symbolic approaches show that high label prediction accuracy does not guarantee rule recovery, revealing distinct perception and induction bottlenecks and weak experimental informativeness.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages33

Reading Time1 h 23 min

Introduction

Related Work and Motivation

ZendoWorld Benchmark

Agent Evaluation Approach

Agent Variants and Methods

Key Findings and Human Comparison

arXiv :2607 .08233v 1 [ cs .AI] 9 Jul 2026 Playing ZendoWorld: Challenging AI Agents on Active Visual Concept Induction Sophia Koehler 1 ,2 Antonia Wüst 1 Inga Ibs3 ,4 Wasu Top Piriyakulkij5 Wolfgang Stammer6 Constantin Rothkopf2 ,3 ,4 Kevin Ellis5 Kristian Kersting 1 ,2 ,4 ,7 1AIML Lab, TU Darmstadt; 2Hessian Center for AI (hessian.AI); 3 Psychology of Information Processing, TU Darmstadt; 4 Centre for Cognitive Science, TU Darmstadt; 5 Cornell University;6Max Planck Institute for Informatics, SIC; 7 German Center for AI (DFKI); Abstract A central challenge in building intelligent systems is enabling agents to jointly perceive complex inputs, form hypotheses about hidden patterns, and design infor mative experiments to test them. To study this problem, we propose ZendoWorld, a controlled interactive environment in which agents must infer a logical rule about visual game observations, acquire information by proposing new scenes, and refine their hypotheses based on feedback from the game environment. We evaluate several agents spanning pure VLM reasoning, Bayesian particle filtering, dynamic concept discovery, and neuro-symbolic methods. Our main findings are: (1) high accuracy in predicting labels for observed examples does not imply recovery of the underlying rule; (2) perception and induction are distinct bottlenecks for different agent classes; and (3) VLM-based agents propose near-uninformative experiments, failing to actively reduce hypothesis uncertainty. To compare these results, we col lect human data on the task, which reveals a gap in inductive reasoning, particularly for more complex rules. Overall, ZENDOWORLD takes an important step toward evaluating intelligent agents and identifies concrete avenues for improvement, par ticularly in domains like scientific discovery. Code: [https://github.com/ml-research/ZendoWorld](https://github.com/ml-research/ZendoWorld) Data: [https://huggingface.co/datasets/ss567uhg/zendo-synthetic-data](https://huggingface.co/datasets/ss567uhg/zendo-synthetic-data) 1 Introduction The ability to form, test, and revise hypotheses in light of new evidence is a hallmark of human intelligence, reflecting the interplay between perception and inductive reasoning. Cognitive science has long emphasized this process as central to learning, showing how humans actively explore hypothesis spaces through informative experiments [10, 7, 24] . From a causal perspective, this process can be understood as learning not only from observations but from interventions that probe underlying mechanisms [26, 27] . With the growing interest in AI agents and AI scientists [16, 20, 38], structured benchmarks and controlled analyses are needed to assess these capabilities systematically. Recent progress in AI has focused on inductive reasoning from few-shot examples [31, 19, 12, 37], typically evalu- Figure 1: ZENDOWORLD combines perception, induction, and experimentation in a controlled visual environment. ated on grid-based puzzles such as ARC-AGI [6] or visual reasoning tasks over synthetic and real images [3, 39, 14, 34, 30] (cf. Figure 1 blue ∩ yellow regions) . While these benchmarks highlight Preprint. Table 1: Taxonomy of benchmarks at the intersection of perception, reasoning and experimentation. Our environment (ZENDOWORLD) is the first to require grounded visual induction through active experimentation. | Benchmark / Env | Primary Modality | Perception | Induction | Experimentation | | --- | --- | --- | --- | --- | | ARC-AGI-1/2 [6] | Symbolic (Grid) | × | ✓ | × | | Bongard, RAVEN [3, 39, 23] | Graphic | ✓ | ✓ | × | | Bongard-style [14, 34, 21, 25] | Visual (3D) | ✓ | ✓ | × | | Atari, NetHack [17, 33] | Visual (2D) | ✓ | ×* | ✓ | | MineDojo [9] | Visual (3D) | ✓ | ×* | ✓ | | CLEVR-AVR [40] | Visual (3D) | ✓ | ×** | ✓ | | Zendo, ActiveACRE [4][28] | Symbolic (Text) | × | ✓ | ✓ | | ARC-AGI-3, AutumnBench [1, 32] | Symbolic (Gr

FAQs

How do different agent types struggle in ZendoWorld?

The results suggest distinct bottlenecks: perception and induction limit different agent classes, and VLM-based agents tend to propose near-uninformative experiments that fail to actively reduce hypothesis uncertainty.

Jul 16, 2026

Upload by Theodora