Large language models trained to reason by mimicking human-like thought processes actually lose the cognitive efficiency that makes human reasoning so powerful, according to a new study from researchers at multiple institutions.
The Research
Yueqing Hu and colleagues (2026) tested the "Hán Dān Xué Bù" (Superficial Mimicry) hypothesis across 14 different AI models. They compared teacher models trained via reinforcement learning — which naturally match human difficulty scaling (correlation r=0.64) — with student models trained through supervised fine-tuning (SFT) to imitate the teacher's reasoning traces. The distilled students showed a dramatic "Functional Alignment Collapse": their alignment with human cognitive costs dropped to r=0.34. In many cases, the students performed worse than their own pre-distillation baselines — a phenomenon the authors call "Negative Transfer."
The team identified a "Cargo Cult" effect: student models ritualistically copied the linguistic form of reasoning (like being overly verbose) without internalizing the teacher's dynamic resource allocation policy. In other words, they learned to talk like someone thinking hard without actually thinking efficiently.
Why It Matters
This finding has profound implications for how we understand learning — both in AI and in humans. It suggests that human-like cognition isn't something you can copy by watching; it emerges from active reinforcement and trial-and-error. For anyone interested in improving their own thinking, it reinforces that passive observation (watching others solve problems) is far less effective than actively wrestling with challenges yourself.
What You Can Do
Don't just mimic experts — engage actively. When learning a new skill, try solving problems before seeing the answer. Use spaced repetition and self-testing instead of rereading notes. The brain builds efficient reasoning only through effortful practice, not passive imitation.
Source: arXiv q-bio.NC
Curious about your own brain? Take our free adaptive IQ test or try 306 brain training levels.
