TL;DR
This paper introduces Behavior-Conditioned Symbolic Synthesis (BCSS), a framework that uses Large Language Models (LLMs) to automatically "reverse-engineer" the rules humans and AI systems follow when they make mistakes. By synthesizing interpretable symbolic programs that reproduce specific errors, the researchers discovered that both humans and machines are "satisficers"—they often settle for simplified, partial rules that only explain some of the available evidence.
What Problem Does This Paper Address?
For decades, cognitive scientists have used symbolic models (like logic-based programs) to understand the human mind because they are explicit and interpretable. However, these models are notoriously difficult to scale; they usually require experts to hand-write programs for every possible strategy a person might use.
The authors tackle this "engineering bottleneck" in the context of the Abstraction and Reasoning Corpus (ARC), a benchmark of visual logic puzzles that is notoriously difficult for AI. While we can see when a human or an AI fails an ARC task, we rarely understand the "buggy" logic behind the failure. This paper aims to bridge that gap by automatically generating symbolic descriptions of incorrect reasoning at scale.
Methods
The core innovation is Behavior-Conditioned Symbolic Synthesis (BCSS). Instead of trying to write every possible incorrect rule by hand, the authors use an LLM as a "proposal distribution."
- Reference Programs: They start with a correct program (written in a Python-like Domain-Specific Language or DSL) that solves a specific ARC task.
- LLM-Guided Search: They prompt an LLM with the task grids and the correct program, asking it to propose a small number of line-edits that would result in a specific incorrect output observed from a human or another AI.
- Verification: Each proposed "error program" is executed. If its output matches the actual behavior of the human or machine, it is retained as a valid model of that individual's reasoning.
- Classification: Finally, they use an LLM-based classifier to sort these error programs into a taxonomy of seven cognitive categories, such as "Attribute Misselection" or "Scope Restriction."
Datasets
- H-ARC: A large-scale dataset containing over 6,000 unique incorrect responses from 1,700+ human participants solving ARC tasks.
- ARC-AGI-1: The standard benchmark of 800 abstract reasoning tasks.
- Large Reasoning Models (LRMs): The authors evaluated two state-of-the-art models, Qwen3.5-27B and GPT-OSS-120B, across different "reasoning effort" levels (varying the amount of test-time compute/Chain-of-Thought).
Key Insights and Ideas
- The Simplicity Bias: Both humans and AI systems exhibit a strong bias toward simpler rules. Error programs were significantly shorter than the correct reference programs, suggesting that reasoners often prune complex logic in favor of "good enough" heuristics.
- Satisficing in Silicon: Machines, like humans, are "satisficers." They often adopt rules that perfectly explain the first few examples they see but fail to generalize to the entire set. Interestingly, as you give an AI model more "thinking time" (compute), its error patterns become more sophisticated and less prone to extreme oversimplification.
- Attribute Misselection is Universal: The most common error for both humans and AI is focusing on the wrong feature (e.g., coloring an object based on its position rather than its size).
- Shared Signatures, Different Biases: While humans and AI share broad error categories, they fail on different tasks. Humans are more likely to skip a drawing step entirely, while AI models are more likely to hallucinate incorrect arithmetic.
Key Results
- Coverage: BCSS successfully captured 52.5% of non-trivial human errors and up to 62.8% of machine errors, proving that most mistakes in this domain are systematic and rule-based rather than random.
- Complexity Reduction: Human error programs were 19.1% shorter on average than the correct solutions. For AI, this "simplicity bias" was even stronger in lower-performing models (up to 29.3% shorter).
- Discriminability: The authors built a classifier that could distinguish between a human error and a machine error with an AUC of 0.65–0.69, indicating that while their reasoning styles are similar, they possess distinct "inductive biases."
Limitations and Future Work
- The "Half-Empty" Glass: BCSS only captures about half of the errors. The authors note that errors involving adding new logic (rather than just editing existing lines) are harder for the current system to find.
- DSL Constraints: The symbolic language used was designed for solving ARC, not specifically for cognitive modeling. Some primitives might not perfectly map to how the human brain processes visual information.
- Process vs. Product: BCSS identifies the final rule an agent likely followed, but it doesn't yet model the dynamic process (the "Aha!" moment or the search through hypotheses) of how they arrived at that rule. Future work could integrate "think-aloud" protocols or eye-tracking data to refine these models.
Summary updated