💬 NLP

Human Adults and LLMs as Scientists: Who Benefits from Active Exploration?

This paper demonstrates that granting human adults agency through active exploration significantly improves their ability to identify conjunctive causal rules compared to passive observation, while revealing that although large language models can match human inference accuracy, they often employ less efficient exploration strategies and exhibit similar performance gaps between conjunctive and disjunctive reasoning.

Original authors: Mandana Samiei, Eunice Yiu, Anthony GX-Chen, Dongyan Lin, Jocelyn Shen, Blake A. Richards, Alison Gopnik, Doina Precup

Published 2026-06-05

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Mandana Samiei, Eunice Yiu, Anthony GX-Chen, Dongyan Lin, Jocelyn Shen, Blake A. Richards, Alison Gopnik, Doina Precup

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

The Big Question: Are Adults Bad at "Teamwork" Logic?

Imagine you are trying to figure out how a mysterious machine works. You have four different keys (let's call them "Nexioms"). You need to find out which keys turn the machine on and how they work together.

There are two ways the machine could work:

The "Or" Rule (Disjunctive): The machine turns on if you put in any one of the correct keys. (Like a light switch: flip one, and the light goes on).
The "And" Rule (Conjunctive): The machine only turns on if you put in all the correct keys at the same time. (Like a safe: you need the code, the fingerprint, and the key all together).

The Old Finding:
Scientists used to think adults were bad at figuring out the "And" rule. When people just watched someone else test the keys, adults usually guessed the "Or" rule, even when the evidence suggested an "And" rule. It was like adults had a mental blind spot for teamwork logic.

The New Question:
The authors of this paper asked: Is this because adults are actually bad at logic, or is it because they were just sitting passively watching?

They wondered if adults would do better if they could act like scientists—picking up the keys themselves, testing them, and seeing what happened immediately.

The Experiment: The "Nexiom Detector"

The researchers built a digital game called the "Nexiom Detector."

The Active Group: These participants could click to add or remove keys and press "Test" to see if the machine turned on. They controlled the experiment.
The Passive Group: These participants just watched the Active Group's screen. They saw the same results but couldn't touch anything.
The "Planner" Group: A third group could plan which keys to test, but they didn't see the results of their own plans. Instead, they watched the results of someone else's tests.

They also tested several Large Language Models (LLMs) (AI chatbots) to see if they could solve the puzzle as well as humans.

What They Found

1. Giving Adults a "Remote Control" Changed Everything

When adults were allowed to actively test the keys themselves, their performance skyrocketed.

The Result: They became very good at figuring out the tricky "And" rules. They successfully identified which keys were needed and that all of them had to be present.
The Takeaway: Adults aren't bad at "And" logic. They just struggle when they can't design their own tests. When they act like scientists, they are just as smart as children at figuring out complex rules.

2. The "Planner" vs. The "Doer"

Here is a crucial finding: Just thinking about what to test isn't enough.

The "Planner" group (who thought about tests but watched someone else's results) did poorly. They performed almost as badly as the Passive group.
The Analogy: Imagine you are a chef. If you write down a recipe (planning) but someone else cooks the dish and you only watch them eat it, you won't learn how to cook. You need to taste the food you cooked to know if the salt was right.
The Lesson: Active learning works best when your action is tightly linked to the result. You need to see the immediate consequence of your own choices.

3. The "And" Rule is Still Harder (But Doable)

Even with active control, the "And" rule took more time and more tests to figure out than the "Or" rule.

The Analogy: Finding a single key that opens a door is easy (one try might work). Finding the exact combination of three keys that opens a safe takes more trial and error.
However, the adults didn't give up. They just ran more tests to be sure. They didn't get stuck; they just worked a little harder.

4. How Did the AI (LLMs) Do?

The researchers pitted the humans against advanced AI models.

The "Or" Rule: The AI was great. Some models performed just as well as the top human scientists.
The "And" Rule: The AI struggled more than the humans. While some smart models got close, many of them were less efficient. They often took more tests than necessary or got confused by the "And" logic.
The Takeaway: AI can be a good scientist for simple rules, but humans are still better at the strategic, step-by-step detective work required for complex "teamwork" rules.

Summary: What Does This Mean?

This paper tells us that adults are not "broken" when it comes to complex logic. The problem wasn't their brains; it was the situation.

Passive Observation: Like watching a magic show from the audience. You see the trick, but you don't understand how it's done because you can't touch the props.
Active Exploration: Like being the magician's assistant. You get to hold the props, try the tricks, and see what happens when you pull the wrong lever.

The Main Conclusion:
When we let adults act like scientists—giving them the freedom to test their own ideas and see the immediate results—they are excellent at figuring out complex "And" rules. The "handicap" we thought they had was actually just a lack of agency.

However, even with this freedom, figuring out complex rules takes more effort than simple ones, and while AI is getting good at this, humans still have a slight edge in being efficient, strategic explorers.

Technical Summary: Human Adults and LLMs as Scientists: Who Benefits from Active Exploration?

Problem Statement

A persistent finding in causal learning literature is that adult humans struggle to identify conjunctive causal rules (where an effect requires the simultaneous presence of multiple causes, $A \land B$ ), often defaulting to disjunctive rules (where any single cause is sufficient, $A \lor B$ ). This "conjunctive handicap" has traditionally been demonstrated using passive observation paradigms, where learners observe a fixed sequence of evidence without control over how that evidence is generated.

The central question addressed by this paper is whether this bias persists when adults are granted agency through active exploration. Specifically, the authors investigate if the inability to infer conjunctive rules stems from a fixed limitation in causal competence or from the constraints of passive evidence presentation, which prevents learners from generating and evaluating the specific evidence needed to update their priors. Furthermore, the study situates human performance within a broader computational context by comparing adult learners to Large Language Models (LLMs) acting as general-purpose reasoning agents in the same active causal discovery setting.

Methodology

Experimental Task: The "Nexiom Detector"

The authors developed a custom web-based platform, Nexiom Text Adventure, functionally equivalent to the classic "blicket detector" paradigm but utilizing the novel term "nexiom" to minimize prior knowledge from established literature.

Setup: Participants interact with a machine that activates based on the presence of specific objects ("nexioms") and an underlying causal rule.
Conditions:
1. Active Exploration: Participants freely select objects, test arbitrary combinations, and decide when to stop exploring to infer both the causal objects and the rule (Conjunctive vs. Disjunctive).
2. Passive Observation (Yoked Control): Participants observe the actions and test outcomes of a paired Active participant without performing any interventions themselves.
3. Passive Proposer: A supplementary condition where participants generate hypotheses and propose interventions but receive outcomes generated by a matched Active participant, rather than their own. This isolates whether benefits arise from planning interventions or from receiving contingent feedback.

Participants and Agents

Humans: 306 adult participants (ages 22–35) recruited via Prolific. They were assigned to Active, Passive Observation, or Passive Proposer conditions.
LLMs: Six models were tested (gpt-5, gpt-5-mini, gemini-2.5-flash, deepseek-reasoner, o4-mini, deepseek-chat). Models were given the same action space as humans: sequentially adding/removing objects, testing combinations, and observing outcomes.
Design: Experiments utilized both Conjunctive and Disjunctive rule structures. Human data included 51 participants per rule type per condition; LLMs underwent 24 trials per rule type (48 total trials per model).

Evaluation Metrics

Accuracy: Binary measures for Object Identification (exact match of causal objects), Rule Inference (Conjunctive vs. Disjunctive), and Full Hypothesis Accuracy (both correct).
Exploration Process:
- Cumulative Information Gain: Quantifying uncertainty reduction (in bits) across tests.
- Hypothesis Pruning: Tracking the number of consistent hypotheses remaining after each test.
- Efficiency: Number of tests, actions, and time per test.

Key Contributions and Results

1. Active Exploration Overcomes the Conjunctive Handicap

Contrary to prior passive studies, active exploration substantially improved adult performance on conjunctive causal rules.

Accuracy: Active adults achieved high accuracy in inferring conjunctive rules, significantly outperforming passive observers.
Efficiency: While conjunctive rules still required more tests to infer than disjunctive rules (Mean tests: 9.6 vs. 6.4 for average humans), active adults generated relatively small but highly informative sets of tests.
Implication: The adult "conjunctive handicap" reflects constraints of passive evidence presentation, not a fixed limitation in causal competence. Active adults demonstrated flexibility similar to that observed in young children.

2. The Necessity of Contingent Feedback

The Passive Proposer experiment revealed that active intervention planning alone is insufficient.

Participants who proposed interventions but observed outcomes from someone else's tests performed as poorly as passive observers (e.g., 0.12 object identification accuracy in conjunctive settings vs. 0.69 for Active Explorers).
Conclusion: The benefits of active exploration depend critically on the tight coupling between self-generated interventions and their contingent outcomes.

3. LLM Performance vs. Human Performance

Disjunctive Rules: Several state-of-the-art models (e.g., gpt-5, deepseek-reasoner) approached or matched top-performing human accuracy. Disjunctive environments appear easier to resolve once agents can actively generate informative interventions.
Conjunctive Rules: Significant performance gaps emerged. While top human explorers achieved perfect accuracy, many LLMs (including smaller or less structured models) showed pronounced declines. Even strong reasoning-oriented models generally lagged behind top human performance in conjunctive settings.
Exploration Strategies:
- Humans: Top-performing humans rapidly converged on targeted multi-object interventions to prune the hypothesis space.
- LLMs: While some models adjusted intervention complexity, they often exhibited less efficient exploration strategies and similar conjunctive-disjunctive performance gaps as humans, though less effectively mitigated.
- Error Patterns: LLMs frequently made "near-miss" errors (misclassifying a single causal object), whereas human errors were more heterogeneous.

4. Search Efficiency and Information Gain

Disjunctive Settings: Both humans and models rapidly accumulated information, with marginal utility approaching zero after a few tests.
Conjunctive Settings: Required longer exploration trajectories. Informative interventions continued to emerge later in the sequence. Successful discovery was associated with progressive hypothesis pruning and adaptive sequential search strategies.

Significance and Claims

The paper claims that adult causal learning is more adaptive and flexible than prior passive paradigms suggested. The apparent failure in conjunctive reasoning is not due to an inability to represent complex causal structures but arises from the interaction between strong default heuristics (favoring disjunctive rules) and constrained learning environments.

By allowing adults to act as "scientists"—exploring and generating their own evidence—the study reveals forms of causal competence that passive observation obscures. The findings align with interventionist accounts of causal learning, suggesting that causal knowledge is fundamentally about understanding how actions change the world.

Furthermore, the comparison with LLMs suggests that active intervention alone is insufficient for human-like causal discovery. Successful performance depends on the quality, efficiency, and informativeness of the tests, particularly in conjunctive environments. While some LLMs can match humans on hypothesis inference accuracy, they often lack the strategic testing and evidence selection capabilities that characterize top human explorers.

The study concludes that developmental differences between children and adults may reflect differences in the strength of default heuristics and the willingness to revise beliefs, rather than fundamental differences in causal reasoning mechanisms. When provided with an environment supporting hypothesis-driven exploration, adult learning resembles the flexibility of younger children.