When Agency Beats Intelligence: What the Conjunctive Handicap Tells Us About Scientific Reasoning in Humans and LLMs

Causal reasoning sits at the heart of what we usually call intelligence. We predict, explain, and intervene in the world precisely because we can build mental models of what causes what. A long line of work in cognitive science has used a deceptively simple toy, the "blicket detector," to probe how people learn causal structure. And one of the more uncomfortable findings from that literature is that adults are surprisingly bad at one specific kind of causal rule. The recent paper "Human Adults and LLMs as Scientists: Who Benefits from Active Exploration?" by Samiei, Yiu, and colleagues revisits this finding and arrives at a conclusion that should reshape how we think about evaluating both human and machine reasoning.

The conjunctive handicap, and why it is strange

In the classic blicket detector paradigm, a machine lights up when certain objects, or combinations of objects, are placed on it. Learners observe a few trials and then must infer which objects are causally relevant and under what rule the machine activates. Two rule types matter here:

• Disjunctive (OR) rules: the machine turns on when any causal object is present.
• Conjunctive (AND) rules: the machine turns on only when all causal objects are present simultaneously.

The well-documented result is that adults default to disjunctive interpretations and struggle to infer conjunctive rules, even when the evidence supports them. More striking still, preschool-aged children often outperform adults on conjunctive inference (Lucas et al., 2014). This "conjunctive handicap" has held up across cultures and socioeconomic contexts (Wente et al., 2019), which made it look like a robust feature of adult cognition, perhaps a consequence of strong learned priors that favor simple OR explanations.

But there is a methodological catch that the authors zero in on. Nearly all of these demonstrations relied on passive observation. Participants watched a fixed sequence of trials chosen by the experimenter. They never got to ask their own questions of the world. And causal learning, as decades of developmental work suggests, is fundamentally an active process. Children learn better when they design their own tests; adults learn worse when interventions are imposed on them rather than self-directed (Schulz et al., 2007; Sobel & Kushnir, 2006).

This sets up the central question of the paper: is the conjunctive handicap a genuine limit of adult causal competence, or an artifact of denying people the ability to generate the evidence they actually need?

What active exploration changes

To test this, the authors built a custom web platform, the Nexiom Text Adventure, implemented in Streamlit. The task is functionally identical to the blicket detector but uses an invented term, "nexiom," to strip away any priors people might bring from prior blicket-style tasks or intuitions about familiar objects. Participants could freely add or remove four objects from the machine, explicitly test any combination, observe whether it switched on or off, and decide for themselves when they had gathered enough evidence to commit to an answer about both the causal objects and the rule.

The design includes three conditions that, taken together, are what make the study interesting:

1. Active exploration: participants choose their own interventions and observe the outcomes.
2. Passive observation (yoked control): each participant is paired with an active explorer and simply watches that person's actions and outcomes. Crucially, the evidence is identical to what the active learner saw.
3. A condition in which participants propose interventions but observe outcomes from someone else's tests, decoupling choice from contingent feedback.

The headline result: active exploration substantially improves adults' conjunctive reasoning. Active adults showed strong performance on conjunctive rules and, notably, did so while generating relatively small but highly informative sets of tests. The conjunctive handicap did not vanish entirely, since conjunctive rules still required more tests to infer than disjunctive ones, but the large competence gap reported in the passive literature mostly collapsed.

The more subtle and, to me, more important finding comes from the third condition. Generating informative tests was not enough. Participants who proposed interventions but watched outcomes from someone else's tests performed as poorly as fully passive learners. The benefit of agency was not in the act of choosing, nor merely in seeing good evidence. It came from the tight coupling between a self-generated intervention and its own contingent outcome. You have to ask the question and receive the answer to your own question for the learning advantage to materialize.

This is a clean dissociation. It tells us that the active learning advantage is not reducible to better information content in the evidence stream, because the yoked condition controlled for exactly that. Something about the closed loop between hypothesis, action, and feedback is doing the work.

The LLM comparison

The authors then placed a range of large language models in the same Nexiom environment, letting models select interventions and observe outcomes just as human explorers did. This is a natural extension given the growing interest in LLMs as general-purpose reasoning agents that generate hypotheses and design experiments (Piriyakulkij et al., 2024; GX-Chen et al., 2025).

The results are nuanced rather than dismissive. Some state-of-the-art models approached human-level performance on hypothesis inference accuracy. But two consistent weaknesses appeared:

• Models explored less efficiently, generating less informative test sequences than the best human explorers.
• Models showed a similar conjunctive-disjunctive performance gap, and unlike humans, did not consistently benefit from choosing their own interventions.

That last point is the crux. For humans, the loop between self-directed action and contingent outcome unlocked conjunctive reasoning. For LLMs, having control over interventions did not reliably help. The models could often produce the right final answer when the evidence was favorable, but they were weaker at the upstream skill of adaptively searching the hypothesis space and pruning hypotheses as evidence accumulated.

My take: we are benchmarking the wrong half of science

This study crystallizes a distinction I think the field has underweighted. There are two separable competencies in causal discovery:

1. Inference: given a body of evidence, recover the underlying structure.
2. Experiment design: decide which intervention to perform next to maximize information gain.

Almost all of our LLM evaluation infrastructure targets the first. We hand models a fixed context and grade the answer. But real scientific reasoning, and arguably real intelligence, lives substantially in the second. Choosing what to test next, under uncertainty, with the cost of each test in mind, is the harder and more consequential skill. The Nexiom results suggest that LLMs are closing the gap on inference while still lagging on adaptive search.

There is a deeper inference to draw here about why humans benefit from the closed loop and LLMs do not. When a human runs their own test, the intervention is generated by the same internal model that will interpret the outcome. The prediction and the surprise are registered against an active hypothesis. An LLM processing a transcript of its own prior actions does not obviously maintain that same predictive state across turns. Its "agency" is, in a sense, simulated through text rather than grounded in a persistent internal model being updated by prediction error. If that interpretation is right, then the fix is not larger models but architectures or training regimes that preserve and update an explicit belief state across an interaction.

Some honest limitations temper all this. The Nexiom task is small: four objects and two rule structures. Whether the active learning advantage scales to high-dimensional hypothesis spaces, where exhaustive testing is impossible and search strategy matters far more, is an open empirical question. The LLM evaluation also depends heavily on prompting and scaffolding choices, and "less efficient exploration" could partly reflect interface mismatch rather than a pure reasoning deficit. The human sample is adults in a constrained online task, which limits ecological claims about scientific practice writ large.

Closing thoughts

The slogan version of this paper is that agency, not raw intelligence, is the bottleneck for inferring AND rules, at least for humans. The conjunctive handicap was never a fixed cognitive limit; it was an artifact of taking the steering wheel away. The more durable lesson for AI is that matching human accuracy on answers is not the same as matching human competence at science. The capability we should care about, and increasingly benchmark, is the active strategy of choosing what to test next and updating beliefs against the results.

The open question I would most like to see answered: can we build agents whose interventions and interpretations are bound by a shared, persistently updated model, so that they too benefit from closing the loop? If active exploration is what unlocks conjunctive reasoning in people, then engineering genuine, not merely textual, agency may be the path toward machines that can actually do science rather than describe it.