ルールなしで物理を発見するAI:「Baba in Wonderland」のブレークスルー

arXiv cs.AI May 2026
Source: arXiv cs.AIArchive: May 2026
「Baba in Wonderland」と呼ばれる新しい研究フレームワークは、実行可能な世界モデルの核心的な課題を解決します。ルール記述や報酬信号なしに、AIが環境の状態依存ダイナミクスをゼロから自律的に発見できるようにします。このブレークスルーにより、エージェントはパターンマッチングから脱却します。
The article body is currently shown in English by default. You can generate the full version in this language on demand.

The fundamental limitation of current AI world models is their tendency to learn superficial semantic correlations—mapping inputs to outputs—rather than the underlying causal laws that govern an environment. This 'semantic trap' has prevented agents from truly understanding physics. A new framework, 'Baba in Wonderland,' directly addresses this by allowing an agent to induce a readable, editable, and reusable world model purely from online interaction evidence. The agent plays a modified version of the puzzle game 'Baba Is You,' where it must discover how objects interact (e.g., 'rock is push,' 'flag is win') without being told the rules. Through systematic trial and error, it constructs a state-transition graph that encodes the true dynamics. The significance is twofold: first, it represents a paradigm shift from supervised world model learning to self-supervised discovery, eliminating reliance on human-crafted rules or reward functions. Second, it provides a transparent, interpretable model that can be edited and reused across tasks. This directly enables robots to plan in novel environments, video models to understand scene physics, and large language models to perform genuine causal reasoning. The work signals a critical step toward agents that don't just predict the next token, but actively discover and rewrite the laws of their world.

Technical Deep Dive

The 'Baba in Wonderland' framework tackles a deceptively hard problem: how can an agent learn the true causal dynamics of an environment when it has no prior knowledge of the rules, no language descriptions, and no reward signal? The core insight is to treat world model learning as a problem of *inductive inference from interaction traces*.

Architecture & Algorithm:
The agent operates in a grid-world environment based on the game 'Baba Is You,' where objects have hidden properties (e.g., 'rock is push,' 'wall is stop,' 'flag is win'). The agent can move, push objects, and observe the results. The key algorithmic innovation is a two-stage process:
1. Hypothesis Generation: The agent maintains a set of candidate 'rules' in a symbolic form (e.g., `object_A + action_push -> object_A moves`). After each interaction, it generates new hypotheses that are consistent with the observed state transition.
2. Hypothesis Elimination: The agent actively designs experiments—specific sequences of actions—to disambiguate between competing hypotheses. For example, if it suspects 'rock is push' but also 'rock is stop,' it might try to push a rock into a wall to see which rule holds. This is essentially a Bayesian inference loop over state-transition functions.

Why It Works:
The critical technical achievement is avoiding the 'semantic trap.' Traditional neural network world models (e.g., Dreamer, MuZero) learn a compressed latent representation that predicts next frames. They can easily learn that 'pushing a rock into a wall' usually results in the rock stopping, but they cannot distinguish *why*—is it because 'rock is stop' or because 'wall is stop'? The Baba framework forces the agent to represent dynamics as explicit, compositional rules. This makes the model *executable*: you can edit a rule (e.g., change 'wall is stop' to 'wall is push') and the agent will immediately adapt its planning, without retraining.

Relevant Open-Source Repositories:
- Baba Is You (game engine): The original puzzle game by Hempuli. While not a research repo, it provides the testbed. The game's source code is available on GitHub, and the community has built custom level editors.
- Gymnasium (formerly OpenAI Gym): The standard interface for reinforcement learning environments. The Baba framework can be integrated as a custom environment. The repo has over 40k stars and is the de facto standard for RL benchmarking.
- DreamerV3 (by Danijar Hafner): A state-of-the-art model-based RL algorithm that learns a world model from pixels. While it doesn't solve the semantic trap, it represents the baseline approach. The repo has over 2k stars.

Performance Data:
| Metric | Baba in Wonderland | DreamerV3 (baseline) | Random Agent |
|---|---|---|---|
| Levels Solved (out of 50) | 47 | 12 | 2 |
| Rule Discovery Accuracy | 94% | 18% | N/A |
| Planning Horizon (steps) | Unlimited (symbolic) | 15 (latent) | 1 |
| Model Editability | Full (symbolic rules) | None (neural weights) | N/A |
| Training Steps to Convergence | 5,000 | 200,000 | N/A |

Data Takeaway: The Baba framework achieves a 4x improvement in levels solved and 5x improvement in rule discovery accuracy compared to the leading model-based RL baseline. More importantly, it does so with 40x fewer training steps, because it learns the *structure* of the dynamics rather than memorizing transitions. The symbolic model is also fully editable, which is impossible with neural networks.

Key Players & Case Studies

This research is not tied to a single corporate lab but emerges from a collaboration between academic institutions focused on causal representation learning. The lead authors are from the Max Planck Institute for Intelligent Systems and the University of Tübingen, groups known for work on causal discovery and interactive learning.

Comparison with Competing Approaches:
| Approach | Key Entity | Method | Reliance on Rules | Interpretability | Sample Efficiency |
|---|---|---|---|---|---|
| Baba in Wonderland | MPI-IS / Tübingen | Symbolic induction | None | High (symbolic rules) | Very High |
| DreamerV3 | Google DeepMind | Latent dynamics model | None | Low (neural) | Low |
| MuZero | Google DeepMind | Learned model + MCTS | None | Low (neural) | Medium |
| NEAR (Neuro-Symbolic) | IBM Research | Hybrid neural + logic | Partial (logic templates) | Medium | Medium |
| Grounding Language Models | Microsoft / OpenAI | LLM prompting | High (language priors) | Low (black-box) | N/A |

Data Takeaway: The Baba framework occupies a unique niche: it requires no human priors, yet produces fully interpretable models. This is a significant departure from both pure neural approaches (Dreamer, MuZero) and language-grounded approaches (LLMs), which either sacrifice interpretability or rely on semantic priors that may not transfer to novel environments.

Case Study: Robotics Planning
A concrete application is in robotic manipulation. Consider a robot tasked with stacking blocks. A traditional neural world model might learn that 'pushing a block causes it to slide,' but if the surface changes from ice to carpet, the model fails. The Baba framework would discover the rule 'block is slide' on ice and 'block is stop' on carpet, and can adapt instantly when the surface changes. This is being tested in simulated environments like MetaWorld and RLBench.

Industry Impact & Market Dynamics

This breakthrough has the potential to reshape several industries:

1. Robotics: The ability to discover physics from interaction is a holy grail for general-purpose robots. Companies like Boston Dynamics, Tesla (Optimus), and Figure AI are all struggling with the 'sim-to-real' gap—robots trained in simulation fail in the real world because they don't understand the true physics. A self-discovering world model could allow robots to adapt to new environments on the fly.

2. Video Generation & Simulation: Models like Sora (OpenAI) and Genie (Google DeepMind) generate video by predicting pixels, not physics. They often produce physically implausible results (e.g., objects passing through each other). Integrating a discovered physics engine could ground these models in real causal dynamics.

3. Gaming & Virtual Worlds: Game AI that can discover and edit its own physics would enable unprecedented player agency. Imagine a game where you can change the rule 'fire burns' to 'fire heals' and the AI immediately adapts its behavior.

Market Data:
| Sector | Current Market Size (2025) | Projected Growth (CAGR) | Key Players |
|---|---|---|---|
| Autonomous Robotics | $45B | 25% | Boston Dynamics, Tesla, Figure AI, NVIDIA |
| AI Video Generation | $8B | 40% | OpenAI, Google, Runway, Pika |
| Game AI & Simulation | $12B | 18% | Unity, Epic Games, NVIDIA Omniverse |
| Causal AI Platforms | $2B | 35% | CausaLens, Microsoft, IBM |

Data Takeaway: The total addressable market for technologies that enable causal world models is over $60B and growing at 25-40% annually. The Baba framework is still in the research phase, but its core idea—self-supervised discovery of symbolic dynamics—could become a foundational component for all these sectors.

Risks, Limitations & Open Questions

While the results are impressive, several limitations remain:

1. Scalability to Complex Environments: The current framework works in grid-worlds with discrete actions and objects. Scaling to continuous control (e.g., robotic arm manipulation) or high-dimensional visual inputs (e.g., raw pixels) is non-trivial. The symbolic hypothesis space explodes combinatorially.

2. Partial Observability: The agent assumes full observability of the state. In real-world scenarios, the agent cannot see all relevant variables (e.g., friction coefficients, object mass). Discovering dynamics under partial observability remains an open problem.

3. Stochasticity: The framework assumes deterministic dynamics. Real-world physics has noise and randomness (e.g., wind, sensor noise). Extending to stochastic transitions is a major challenge.

4. Computational Cost: While sample-efficient, the hypothesis elimination process requires the agent to actively design experiments. In a large state space, the number of experiments needed could be prohibitive.

5. Ethical Concerns: A robot that can discover and edit its own physics could be dangerous if it learns to exploit loopholes. For example, a robot might discover that 'human is pushable' and act accordingly. Ensuring safety constraints are built into the discovery process is critical.

AINews Verdict & Predictions

The 'Baba in Wonderland' framework is not just another incremental improvement—it is a genuine paradigm shift in how we think about world models. By moving from *learning* to *discovering* physical laws, it addresses the fundamental limitation of current AI: the inability to reason causally.

Our Predictions:
1. Within 12 months, we will see a version of this framework scaled to continuous control tasks, likely in simulation (e.g., MuJoCo, Isaac Gym). The first commercial application will be in game AI, where developers will use it to create adaptive physics puzzles.

2. Within 24 months, a major robotics company (likely Tesla or Figure AI) will announce a research collaboration to integrate this approach into their robot control stack. The key metric will be 'time to adapt to a new environment'—currently hours, targeted to minutes.

3. Within 36 months, the concept of 'self-discovered physics' will become a standard module in video generation models. We predict that OpenAI's Sora or a competitor will release a version that uses a symbolic physics engine to ground its video predictions, significantly reducing physical implausibilities.

What to Watch:
- The release of a public benchmark for 'world model discovery' (similar to the Atari benchmark for RL).
- The first GitHub repo that implements Baba-style discovery for continuous control.
- Any paper that combines this approach with large language models for natural language grounding of discovered rules.

The era of AI as a passive pattern-matcher is ending. The era of AI as an active discoverer of physical laws has begun.

More from arXiv cs.AI

PopuLoRA:集団進化がRLHFを超える自己改善型AI推論を実現する方法PopuLoRA represents a paradigm shift in how large language models (LLMs) can autonomously improve their reasoning capabiGRIDフレームワーク:LLMが脅威インテリジェンスからセキュリティ知識グラフを自動構築GRID represents a paradigm shift in how security knowledge graphs are built. For years, the cybersecurity industry has sAIメモリ過負荷:パーソナライズシステムが約束を守れない理由The AI industry has been locked in a race to expand context windows, with models like GPT-4 Turbo boasting 128K tokens aOpen source hub352 indexed articles from arXiv cs.AI

Archive

May 20262076 published articles

Further Reading

バイナリ・スパイキングニューラルネットワークの解明:SATソルバーがニューロモルフィック・ブラックボックスに論理をもたらす研究者らは初めて、バイナリ・スパイキングニューラルネットワーク(BSNN)をバイナリ因果モデルとして形式化し、SATおよびSMTソルバーを活用して各ニューロンの発火に対する最小限で正確な因果説明を生成しました。このニューロモルフィックコンピマルチフィデリティデジタルツインとLLM:航空機故障診断に因果関係の魂を吹き込む革新的な診断フレームワークは、マルチフィデリティデジタルツインを用いて希少な故障データを生成し、FMEAベースの因果知識を注入し、LLMを活用して自然言語レポートを作成します。これにより、航空保守におけるブラックボックス時代に終止符を打つこPopuLoRA:集団進化がRLHFを超える自己改善型AI推論を実現する方法PopuLoRAは、共有された凍結ベースモデル上の特殊なLoRAアダプターが教師集団と生徒集団として共進化する、集団ベースの非同期自己対戦フレームワークを導入します。自己キャリブレーションを相互評価に置き換えることで、ますます困難な問題を生GRIDフレームワーク:LLMが脅威インテリジェンスからセキュリティ知識グラフを自動構築GRIDは、大規模言語モデルが非構造化のサイバー脅威インテリジェンスからセキュリティ知識グラフを自動構築できる、革新的なエンドツーエンドフレームワークを導入します。計算可能な報酬メカニズムにより、ドメイン知識や教師信号の不足を克服します。

常见问题

这篇关于“AI Discovers Physics Without Rules: The 'Baba in Wonderland' Breakthrough”的文章讲了什么?

The fundamental limitation of current AI world models is their tendency to learn superficial semantic correlations—mapping inputs to outputs—rather than the underlying causal laws…

从“Baba in Wonderland world model GitHub implementation”看,这件事为什么值得关注?

The 'Baba in Wonderland' framework tackles a deceptively hard problem: how can an agent learn the true causal dynamics of an environment when it has no prior knowledge of the rules, no language descriptions, and no rewar…

如果想继续追踪“executable world model vs neural world model comparison”,应该重点看什么?

可以继续查看本文整理的原文链接、相关文章和 AI 分析部分,快速了解事件背景、影响与后续进展。