Playful Problem Solving | ARC Prize @ MIT

2025-10-27 · Source: ARC Prize · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems, Data Science & Analytics · Depth: Advanced, long

Summary

Jun Y Chu, a post-doctoral researcher at Stanford, presented her work on "playful problem solving," focusing on how humans flexibly create, select, and solve new problems, and how cognitive processes change through social interaction. Her research, relevant to games and agentic play, explores human intelligence beyond mere problem-solving, emphasizing curiosity, reasoning about others' goals, attention allocation, and flexible abstract concept reasoning. Chu analyzed the ARC 3 benchmark, noting that the learning process and invention of new problems were more interesting than individual level completion. She uses the classic puzzle game Sokoban in her post-doctoral study, collecting over 2,000 game records from human participants to understand what makes puzzles enjoyable or difficult. Initial findings suggest success correlates with enjoyment, and efficiency (fewer steps) also contributes, though significant noise remains regarding why some struggles are perceived as more interesting or conducive to learning.

Key takeaway

For AI Scientists developing general intelligence benchmarks, you should consider expanding evaluation criteria beyond task completion. Focus on measuring an agent's ability to generate novel problems, assess problem validity, and demonstrate curiosity, rather than solely optimizing for efficiency in solving predefined tasks. This shift could lead to more robust and human-like AI systems capable of flexible reasoning and learning.

Key insights

Human intelligence encompasses problem creation, selection, and assessment, not just problem-solving.

Principles

• Intelligence involves curiosity and reasoning about others' mental states.
• Intelligent agents should discern important from distracting information.
• Flexible reasoning about abstract concepts is a hallmark of intelligence.

Method

Studying human puzzle-solving in games like Sokoban, analyzing enjoyment, difficulty, and strategic learning through game records and state space visualizations.

In practice

• Design AI benchmarks that reward problem generation, not just solution.
• Analyze player engagement beyond success metrics (e.g., struggle types).
• Incorporate human-like curiosity into agent design.

Topics

• Human Intelligence
• Playful Problem Solving
• ARC 3 Benchmark
• Sokoban Puzzle Game
• Cognitive Processes

Best for: AI Scientist, Research Scientist, AI Student

Related on AIssential

• On Games -- and Play | ARC Prize @ MIT — Human intelligence is uniquely defined by the capacity to invent arbitrary problems and objective functions, not just efficient skill acquisition.
• Will human minds still be special in an age of AI? — Human and artificial intelligence are fundamentally different, shaped by distinct constraints and leading to varied problem-solving approaches.
• Measuring progress toward AGI: A cognitive framework — A cognitive taxonomy and evaluation protocol are proposed to measure AI progress toward AGI against human benchmarks.
• Interaction-Centered Intelligence: Toward Interaction as the Primary Unit of Analysis in Co-Creative AI and Human-AI Systems — Intelligence emerges from evolving interaction dynamics among agents and environments, not just internal computation.
• How a 27M Model Beat Trillion-Parameter AI — Iterative internal processing, not scale or brain-like hierarchy, drives efficient problem-solving in AI.
• We destroyed AI. For Good? — Current AI research is overly fragmented, lacking a unified theoretical framework for intelligence.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by ARC Prize.