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

Adults often struggle to learn rules that require multiple things to happen at once (an "AND" rule). They frequently default to simpler "OR" rules instead. This difficulty is a well-documented phenomenon in cognitive science known as the "conjunctive handicap." A new study from researchers at Mila, McGill, and UC Berkeley suggests this isn't necessarily a permanent brain defect. Instead, it may be a byproduct of how we are taught. When adults move from passive observers to active experimenters, their ability to solve these complex puzzles improves significantly.

The limitations of passive observation

In the study of causal learning—how agents figure out that "Action A causes Effect B"—researchers have long noted a discrepancy between children and adults. Historically, studies using the "blicket detector" paradigm (a task where learners identify which objects trigger a machine) found that while preschoolers could easily grasp conjunctive rules, adults often failed. As shown in, prior studies indicated that adults struggle to infer these "AND" rules when they are simply watching a fixed sequence of events unfold.

This reliance on passive observation creates a significant bottleneck. In a passive setting, the learner is a spectator. They are forced to interpret a pre-determined stream of evidence. This limits their ability to resolve ambiguity. If a machine turns on, a passive observer might not know if it turned on because of Object A, Object B, or the combination of both. Without the ability to intervene, adults tend to fall back on "disjunctive" (OR) heuristics. These are mental shortcuts that assume any single cause is sufficient. These shortcuts are easier to verify with limited, unguided data.

Testing via the Nexiom Detector

To determine if agency could break this handicap, the authors developed the "Nexiom Detector." This is a text-based interactive platform designed to minimize existing biases. Unlike previous studies, the researchers gave participants control over the environment. As illustrated in, the experiment was split into two primary modes: Active Exploration and Passive Observation.

In the Active Exploration condition, participants acted as scientists. They could add or remove specific objects from the machine. They could then explicitly trigger a "test" to see the outcome. This allowed them to design targeted interventions (experiments meant to rule out specific possibilities). In the Passive Observation (yoked control) condition, participants merely watched a paired active explorer perform these actions.

The researchers also added a critical third group: Passive Proposers. These participants were allowed to plan their own experiments and propose specific interventions. However, they did not see the results of their own actions. Instead, they observed the outcomes of a different person's tests. This distinction was vital for isolating the mechanism of learning. The researchers wanted to know if the benefit comes from the mental act of planning a test, or from the "contingent feedback" (seeing exactly how your specific action changed the world).

Evidence for agency-driven learning

The results indicate that active participation fundamentally changes the learning trajectory. The authors report that active exploration substantially improves adult performance on conjunctive causal rules. While conjunctive rules still require more effort, the "handicap" is mitigated when the learner is in the driver's seat.

The study highlights a clear trade-off in complexity and effort. As seen in [Table 1], adults conducting active exploration performed more tests in conjunctive settings ($M=9.6$) than in disjunctive settings ($M=6.4$). This reflects the inherent informational demand of "AND" rules. A single successful test of two objects does not prove they are both required. It only proves that the combination works. To truly prove a conjunctive rule, one must also test the objects individually. This ensures neither is sufficient on its own.

Perhaps the most striking finding involves the Passive Proposers. The authors report that this group performed substantially worse on object identification in conjunctive environments (0.12 accuracy) compared to both Passive Observers (0.45) and Active Explorers (0.69). This demonstrates that successful causal discovery depends on the tight coupling between a self-generated intervention and its immediate, observable consequence. Simply thinking about an experiment is not enough. You must see the fallout of your own choices.

Divergent paths for LLMs and humans

The researchers extended this framework to several large language models (LLMs), including GPT-5 and DeepSeek-Reasoner. They wanted to see if artificial agents benefit from agency in the same way. The results reveal a widening gap between human intelligence and current AI when complexity increases.

In disjunctive (OR) environments, the gap is narrow. The authors report that some state-of-the-art models, such as GPT-5, achieve near-human performance in these simpler settings. However, the conjunctive condition exposes the limits of current architectures. While top-performing humans maintain high accuracy, many LLMs exhibit pronounced declines. Even the strongest reasoning-oriented models generally remain below the performance of top-tier human explorers in these complex settings.

Furthermore, the error profiles differ. As shown in, when LLMs fail, they often commit "near-miss" errors.

These involve misclassifying only a single causal object. Human errors are more heterogeneous (varied in type). The study also notes that while models can mimic human-like accuracy in simple tasks, they often lack the efficient, adaptive search strategies seen in the most successful human participants.

The verdict on active discovery

If you are building an autonomous agent for scientific discovery or complex troubleshooting, the takeaway is clear. Providing an agent with an "action space" (a set of possible moves) is a necessary but insufficient condition for robust causal reasoning. An agent must not only be able to propose an experiment. It must also be able to observe the direct, contingent consequences of its specific interventions.

The study suggests that the "conjunctive handicap" in humans is likely a product of environmental constraints. It is not necessarily a structural inability to represent complex logic. For AI development, the goal should not just be increasing parameter counts. Developers should also focus on improving the efficiency and adaptivity of the exploration process. Until models can match the strategic, hypothesis-pruning capabilities of a human scientist, they will continue to struggle with the "AND" rules of the physical world. Code for the Nexiom platform is reportedly available via the project's Streamlit implementation.

Figures from the paper