Summary: New research shows that humans and artificial intelligence employ two complementary learning systems in similar ways: a fast, flexible in-context learning mechanism and a slower, incremental form of learning. Experiments reveal that AI can develop human-like in-context learning after extensive incremental training, and both systems show comparable trade-offs between adaptability and long-term retention.
The study indicates that difficult experiences tend to strengthen long-term memory, while easier, error-free practice promotes rapid flexibility. These insights may guide the design of AI systems that align more naturally with human cognition.
Key Facts
- Shared strategies: Humans and modern neural networks use both in-context and incremental learning to solve tasks.
- Meta-learning result: Flexible in-context learning in AI emerged only after thousands of incremental training episodes.
- Trade-offs: Both humans and AI balance quick rule learning and durable memory updates; harder tasks favor retention, easier tasks favor flexibility.
Source: Brown University
Overview
Researchers at Brown University report new evidence that links two long-recognized modes of learning—rapid, example-driven inference and slower, practice-based adaptation—across humans and artificial neural networks. Led by Jake Russin, a postdoctoral researcher in computer science, the team explored how the dynamics of in-context learning and incremental weight changes interact and how those interactions resemble the interplay between human working memory and long-term memory.
“These findings help explain why people sometimes behave like rule-based learners and at other times like gradual, incremental learners,” Russin said. “They also illuminate similarities between modern AI systems and human brain processes.”
Russin collaborated with Michael Frank, professor of cognitive and psychological sciences and director of the Center for Computational Brain Science, and Ellie Pavlick, associate professor of computer science and leader of the AI Research Institute on Interaction for AI Assistants, both at Brown.
The study appears in the Proceedings of the National Academy of Sciences (PNAS).
Human learning often operates in two distinct ways. In many tasks—such as deducing the rules of a simple game—people can quickly infer rules from a few examples, a capability the authors describe as in-context learning. In other domains—like mastering a musical piece—skill builds gradually through repeated practice and longer-term memory consolidation, an incremental learning process.
While previous work acknowledged that both humans and AI can use these learning modes, how they interact remained unclear. Drawing on ideas from computational neuroscience, Russin hypothesized that in artificial networks the relationship between in-context performance and incremental weight updates parallels the human distinction between working memory and durable long-term memory.
To test this, the team used meta-learning: a training regime that teaches AI systems how to learn. Through meta-learning, they probed when and how in-context learning capabilities arise in neural networks that otherwise update knowledge by changing internal weights.
One experiment adapted a human psychology task to AI: the network learned separate lists (for example, colors and animals) and was then tested on its ability to recombine items it had never seen paired before (such as “green giraffe”). After being trained across roughly 12,000 related tasks, the model reliably inferred new combinations, demonstrating emergent in-context compositionality.
This pattern—where fast, flexible inference appears after extensive incremental exposure—mirrors how humans often acquire rapid rule-learning skills only after many prior experiences with similar rules or structures. As Pavlick observed, “You may struggle with your first few board games, but by the hundredth, you can quickly pick up the rules of a new game you haven’t encountered before.”
The researchers also identified systematic trade-offs. Tasks that were more difficult and produced larger errors tended to be better retained by the network over time, analogous to error-driven updates that strengthen long-term memory in humans. Conversely, tasks learned without errors tended to enhance flexible in-context performance but produced weaker long-term retention.
Michael Frank noted that this finding aligns with long-standing observations in human learning: mistakes prompt memory updates, while smooth, error-free adaptation favors on-the-spot flexibility without engaging long-term consolidation processes as strongly.
Beyond theoretical implications, the work has practical relevance for building AI assistants that collaborate with people. Understanding where AI learning strategies align with human cognition—and where they diverge—can inform design choices to make AI tools more intuitive, reliable, and trustworthy, especially in sensitive areas such as mental health support.
“If AI assistants are to help people effectively, designers must consider both the similarities and differences in how humans and machines learn,” Pavlick said. “This study provides a valuable first step toward that goal.”
Funding: This research was supported by the Office of Naval Research and the National Institute of General Medical Sciences Centers of Biomedical Research Excellence.
About this AI and learning research news
Author: Kevin Stacey
Source: Brown University
Contact: Kevin Stacey – Brown University
Image: Image credit: Neuroscience News
Original research: Closed access. Title: “Parallel trade-offs in human cognition and neural networks: The dynamic interplay between in-context and in-weight learning” by Jake Russin et al., published in PNAS.
Abstract
Parallel trade-offs in human cognition and neural networks: The dynamic interplay between in-context and in-weight learning
Human learning displays a notable duality: in some situations we rapidly infer rules and compose logical structures from a few examples, while in other situations we rely on gradual, practice-driven improvement. Psychological theory often explains this by positing two distinct systems—one for fast, working-memory-like inference and another for slow, long-term adaptation.
This view faces a puzzle when compared to neural networks, which primarily learn through incremental weight changes. Recent work, however, shows that meta-learned networks and large language models can perform in-context learning: flexibly inferring task structure from a handful of examples without changing weights. In that sense, in-context learning resembles activation-based dynamics tied to working memory.
The present study demonstrates that the interaction between in-context learning and in-weight learning accounts for a wide range of human learning phenomena—such as curriculum effects, compositional generalization, and the trade-off between flexibility and retention. Emergent in-context capabilities can coexist with incremental learning in neural networks, offering an integrated perspective on dual-process theories of human cognition.