🤖 AI

Hypothesis Generation and Inductive Inference in Children and Language Models

This paper compares human children and LLM-based agents in an inductive inference task, revealing that while both adapt similarly to environmental uncertainty and evidence reliability through distinct computational mechanisms, LLMs exhibit unique tendencies toward over-observation and instruction compliance compared to children.

Original authors: Jeffrey Qin, Wasu Top Piriyakulki, Zhuangfei Gao, Mia Radovanovic, Jessica Sommerville, Kevin Ellis, Marta Kryven

Published 2026-05-26

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Jeffrey Qin, Wasu Top Piriyakulki, Zhuangfei Gao, Mia Radovanovic, Jessica Sommerville, Kevin Ellis, Marta Kryven

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are in a room with five locked boxes and a pile of 13 keys. You don't know which key opens which box. To make things tricky, a teacher has just told you a lie: "The red key opens the red box, the blue key opens the blue box." But in reality, the boxes open based on a secret code involving numbers and shapes that you can't see until you pick the boxes up.

This is the setup of the "Box Task," the central experiment in this paper. The researchers wanted to understand how humans (specifically children) and Artificial Intelligence (specifically Large Language Models, or LLMs) figure out the truth when the world is confusing, the information is incomplete, and the instructions might be wrong.

Here is a breakdown of what they found, using simple analogies.

1. The Two Ways of Thinking: The "Constraint" vs. The "Program"

The researchers built two different computer models to simulate how these agents think.

The "Constraint" Model (For Children): Imagine a child's mind as a detective holding a notepad. Every time they try a key and it fails, they write a new rule on the pad: "Okay, red doesn't work." They don't necessarily build a perfect, logical system all at once. Instead, they juggle a few possible ideas (hypotheses) at once. Sometimes they guess, sometimes they follow the teacher's lie, and sometimes they invent a new rule on the fly. The researchers call this "Sets of Constraints."
The "Program" Model (For AI): Imagine the AI as a computer programmer. Instead of just writing notes, the AI writes a tiny computer program (a set of instructions) that says, "If the key has a number, check if it matches the box." If the program fails, the AI rewrites the code. This is called "Program Synthesis."

2. The Big Discovery: They Act Alike, But for Different Reasons

The researchers put both the children and the AI through the same puzzle, sometimes making the keys "sticky" (unreliable) so they wouldn't always open the box even if they were the right key.

The Similarity:
Both children and AI got confused in the same way.

They doubted the evidence: When a key failed to open a box, both groups often thought, "Maybe the lock is just stuck," rather than immediately realizing, "My rule is wrong." They gave the "sticky lock" the benefit of the doubt.
They solved the puzzle without understanding it: Many children (and AI) managed to open all five boxes by luck or by using a messy mix of rules, but when asked to explain the real rule to a new box, they failed. They got the job done without learning the lesson.

The Difference:

The AI is a "Good Student": The AI strictly followed the teacher's instructions. If the teacher said "Match the colors," the AI tried to match colors first, even when it was clearly wrong. It was very obedient.
The Children are "Skeptics": The children were less obedient. Many of them ignored the teacher's lie immediately and started testing random keys.
The "Over-Observer" vs. The "Under-Observer":
- The AI was like a robot that felt it had to check every single box before trying a key. It systematically picked up every box to count the shapes, even when it could have guessed.
- The Children were like lazy detectives. Many of them didn't pick up any boxes at all. They guessed based on what they could see from a distance. They seemed to treat "picking up a box" as a cost they wanted to avoid.

3. Why This Matters

The paper argues that while children and AI can both solve puzzles in uncertain environments, they do it with different "internal costs."

For the AI: The cost of checking information is zero. It doesn't get tired, it doesn't get bored, and it doesn't mind following instructions blindly. It will check every box if the math says it's the most efficient way to be sure.
For the Child: There is an invisible "cost" to gathering information. Maybe it's social (they don't want to look silly picking up boxes), or maybe it's just mental energy. Because of this, they often guess, take shortcuts, and sometimes get the answer right without actually knowing why it's right.

The Bottom Line

The researchers used the AI as a "test subject" (a model organism) to see what happens when you change the rules of the game. They found that if you give an AI the same uncertainty and unreliable evidence that children face, the AI starts acting more like a child: it gets confused, it hesitates to change its mind, and it sometimes solves the puzzle by accident.

However, the AI is still fundamentally different: it is too obedient and too eager to gather data. Children, on the other hand, are willing to take risks, ignore bad advice, and skip steps to save energy. The paper concludes that to truly understand human intelligence, we need to account for these hidden "costs" of thinking and observing that machines don't have.

Technical Summary: Hypothesis Generation and Inductive Inference in Children and Language Models

Problem Statement

Real-world decision-making requires constructing mental models under uncertainty regarding evidence, underlying causal rules, and the state of the world itself. A central challenge in cognitive science and robotics is identifying the computational principles that enable agents to discover coherent theories under these conditions. Specifically, there is a scarcity of experimental frameworks that systematically vary interacting sources of uncertainty across different systems (human vs. artificial) to compare inference mechanisms.

This paper addresses this by modeling inductive inference in the Box Task, a physical puzzle-solving domain originally designed to study how 7–10 year old children respond to misleading instructions under uncertainty. The task involves opening five locked boxes using thirteen keys within a time limit. The true rule (matching the number of shapes on a box to a key's number) is latent, while children are misled by a teacher's video suggesting a color-matching rule. The environment is partially observable (shape counts are hidden until boxes are picked up) and features unreliable evidence (mechanical failures can prevent correct key-box combinations from opening).

Methodology

The authors formalize the Box Task as a Partially Observable Markov Decision Process (POMDP) where an agent must discover a latent rule $h$ from stochastic state transitions and incomplete observations. They implement two complementary frameworks to approximate the solution using Sequential Monte Carlo Samplers (SMC-S) for online Bayesian updating:

1. Sets of Constraints (SoC) – Cognitive Modeling

This implementation treats hypotheses as structured collections of evidence-derived constraints. It serves as a cognitive model fitted to human children's behavior ( $N=100$ ).

Hypothesis Representation: Hypotheses are mappings of boxes to keys.
Proposal Distribution: A mixture of:
- Pre-specified rules: Salient candidate rules (color, shape, order) with assigned priors.
- Online generator: A mechanism to propose novel constraints based on accumulated evidence, upweighting consistent assignments and downweighting falsified ones.
Subjective Reliability: Introduces a parameter $\rho \in (0, 1]$ representing the agent's subjective estimate of evidence reliability. This allows the model to attribute failures to external causes (e.g., a stiff lock) rather than immediately falsifying the hypothesis.
Variants: The authors tested four variants, including lesioned models (removing the generator or subjective reliability) and a full model (SoC-Full).

2. LLM-based Program Synthesis (LLM-PS) – Model Organism

This implementation treats Large Language Models (LLMs) as controllable "model organisms" to probe how behavior changes under systematic variation of environmental conditions.

Hypothesis Representation: Hypotheses are executable Python functions (predict(key, box) -> bool) generated and revised by the LLM.
Inference Engine: The system maintains a population of hypothesis particles. When a particle is selected for "rejuvenation," the LLM is prompted to synthesize a revised program consistent with the history of actions and outcomes.
Action Selection: Agents select actions to maximize expected information gain (EIG) over the particle distribution.
Variants: The authors implemented conditions matching the human task:
- LLM-PS-S: Full observability, unreliable keys.
- LLM-PS-P: Partial observability (requires OBSERVE actions to reveal shape counts), unreliable keys.
- ReAct Baseline: A standard ReAct agent that generates actions directly without maintaining explicit hypothesis representations.

Key Contributions

Formalization of Inductive Inference: The paper extends the Sampling Hypothesis to settings with unreliable evidence and unbounded hypothesis spaces, providing two complementary implementations: constraint-based (SoC) and program synthesis (LLM-PS).
Computational Analysis of Children: The study demonstrates that children's behavior is best explained by the joint operation of subjective evidence reliability (discounting failures as mechanical errors) and online hypothesis generation (creating approximate solutions). This accounts for the dissociation between task completion and true-rule generalization.
Controlled Comparison: By treating LLMs as model organisms, the authors systematically vary evidence reliability and observability. They show that LLM-based agents replicate children's performance patterns in uncertain conditions (e.g., discounting unreliable evidence, seeking to resolve partial information) but exhibit distinct inductive biases, specifically a tendency to over-observe and over-comply with instructions compared to children.

Results

SoC Results

Model Fit: The SoC-Full model (incorporating both subjective reliability and an online generator) provided a significantly better fit to children's trajectories and generalization rates than any lesioned variant ( $p < .001$ ).
Generalization: While 66% of children opened all boxes, only 22% generalized the true rule. The SoC-Full model replicated this, converging on the correct rule in approximately 20% of simulations.
Mechanism: The results indicate that neither subjective reliability nor online generation alone is sufficient; their combination is required to reproduce the distribution of attempts and the low rate of correct generalization.

LLM-PS Results

Replication of Patterns: LLM-based agents replicated children's responses to changes in evidence reliability and observability. They successfully discounted unreliable evidence and sought to resolve partial information.
Task Completion vs. Generalization: Like children, LLM agents often completed the task without discovering the true rule. When this occurred, they relied on "approximate" hypotheses that combined disparate causes (e.g., using color for some boxes and numbers for others) rather than a unified rule.
Divergence in Behavior:
- Instruction Compliance: LLMs strictly adhered to the misleading color-matching instruction on the first trial, whereas 21% of children immediately tested non-color-matching combinations.
- Observation Costs: LLM-PS-P agents systematically observed all boxes before resuming attempts, whereas children varied significantly in observation frequency, with many observing few or no boxes. This suggests children implicitly trade off information value against cognitive or social costs absent in the LLM setting.
Backend Capability: Generalization to the true rule required sufficient backend capability (e.g., GPT-5.2 succeeded with a single particle, while weaker models like DEEPSEEK V3.2 failed to discover the true rule even under full observability).

Significance and Claims

The paper claims to provide a controlled framework for evaluating inductive inference under uncertainty across human and artificial agents. Its primary significance lies in:

Bridging Cognitive Science and AI: It demonstrates that LLMs can serve as valid "model organisms" for cognitive science, allowing for systematic manipulation of environmental variables that is difficult to achieve with human subjects.
Understanding Approximate Inference: The study highlights that successful task completion does not necessarily imply the discovery of the true underlying rule. Both children and LLMs can utilize "approximate" hypotheses that integrate disparate causes to solve immediate problems, which may serve as a stepping stone for later unification.
Inductive Biases: The results suggest that while children and LLMs adapt similarly to environmental structure, their information-seeking behaviors are shaped by different underlying costs and priors. Children appear to operate with a broader prior and implicit costs for observation, whereas LLMs exhibit a bias toward structured, small hypothesis spaces and strict instruction following.

The authors modestly note that their interpretation of LLM behavior remains indirect, as they do not definitively prove that LLMs maintain explicit hypothesis representations or perform likelihood-based updates versus retrieval. They suggest future work could use mechanistic interpretability to distinguish these possibilities and further investigate the "subjective reliability" and rule priors of LLMs.