A new study reveals that the internal memory updates of certain AI models—linear transformers—are mathematically identical to century-old theories of animal learning. This surprising link predicts how these models behave and opens new ways to understand both machine and human cognition.
The Research
Mu Qiao, a researcher at an undisclosed institution, published a paper on arXiv (August 2025, v3 August 2026) showing that major linear-attention families implement named models from classical learning theory. Linear attention mirrors Hebbian contiguity, DeltaNet implements Rescorla–Wagner error correction, and decay variants like RetNet mimic contiguity with a stimulus trace. The study found an exact closed form for Kamin blocking—a phenomenon where prior learning blocks new associations—verified in simulation to less than 10^-7 across five learning rates. Qiao also predicted a dissociation: error-correcting attention exhibits cue competition, while contiguity-based attention does not, and validated this empirically. Additionally, single-state capacity scales with exponents of 1.22 for retrieval and 1.89 for identification, and retrieval error depends primarily on total state size, not head partition. The study also proved no spontaneous recovery for analyzed single-state recurrences under cue-orthogonal retention trials.
Why It Matters
This research bridges AI and psychology, showing that principles from animal learning—like cue competition—apply to machine memory. For your brain, it suggests that how we learn associations may follow similar rules, and understanding these can help you optimize your own learning. The findings also imply that AI models, like humans, can show biases in how they form memories, which has implications for trust and fairness in AI.
What You Can Do
To leverage these insights, vary your learning contexts to avoid cue competition, and space out practice to strengthen associations. For AI users, be aware that models may exhibit biases similar to human conditioning.
Source: arXiv q-bio.NC
Curious about your own brain? Take our free adaptive IQ test or try 306 brain training levels.