Researchers at Zhejiang University have introduced a new AI training method inspired by human cognition, aiming to improve how models categorize and generalize concepts. The approach incorporates human neural activity signals into training, helping AI systems align more closely with the way the brain organizes knowledge. The findings were published in Nature Communications.
Bigger models are not always better at abstraction
The study highlights a key limitation in the current scaling-driven AI paradigm. While increasing model parameters can improve recognition of specific objects, it does not necessarily strengthen a model’s understanding of abstract concepts. In some cases, the paper suggests, scaling up may even interfere with higher-level conceptual reasoning rather than improve it.
Aligning model structure with human cognition
To address this gap, the research team integrated neural activity data into the training process so that model representations could better reflect human cognitive structures. The idea is not just to make models memorize patterns more efficiently, but to help them organize concepts in a way that supports broader generalization and more robust categorization across unfamiliar inputs.
Strong results in low-example tasks
According to the reported experiments, the method delivered particularly strong gains in tasks involving abstract concept recognition. In settings with only a small number of examples, the system achieved a 20.5% improvement in distinguishing abstract concepts, outperforming larger models. The result challenges the prevailing assumption that model size alone is the most effective route toward more capable AI.
Overall, the research points to an alternative direction for AI development: instead of relying solely on scale, structured alignment with human cognition may offer a more effective path toward systems with stronger reasoning and more human-like conceptual understanding.