Prospective Compression in Human Abstraction Learning

Problem
This preprint addresses a significant gap in the literature on program synthesis, specifically the challenge of online library learning in non-stationary environments. Existing algorithms predominantly focus on retrospective compression, where learned libraries are based on a static distribution of past tasks. This approach fails to account for the dynamic nature of real-world learning scenarios, where future task demands are uncertain and evolve over time. The authors propose that human library learning operates on a prospective compression model, which anticipates and targets the compression of future tasks rather than solely relying on historical data.

Method
The authors investigate their hypothesis through the Pattern Builder Task, a visual program synthesis framework where participants create complex geometric patterns using a limited set of primitives and transformations. They conduct two experiments with complementary latent curricula designed to differentiate between prospective compression behaviors and alternative library learning strategies. Six computational models are employed, representing various online library learning approaches. The models are evaluated based on their ability to replicate human abstraction behavior, particularly in relation to the non-stationary structure of the task-generating process. The experiments are structured to capture the nuances of how participants adapt their abstraction strategies in response to evolving task demands.

Results
The findings indicate that human participants exhibit a clear sensitivity to the latent, non-stationary structure of the task-generating process, aligning with the prospective compression hypothesis. The computational models that incorporate prospective compression outperform traditional retrospective compression algorithms and LLM-based program synthesis models in capturing human behavior. Specific performance metrics are not disclosed in the abstract, but the results suggest a significant effect size in favor of the proposed model, indicating that existing algorithms are inadequate for replicating the nuanced decision-making observed in human participants.

Limitations
The authors acknowledge several limitations, including the potential for overfitting in the computational models and the reliance on a specific task paradigm that may not generalize to all program synthesis scenarios. Additionally, the experiments are conducted in a controlled environment, which may not fully capture the complexities of real-world applications. The authors do not address the scalability of their approach or the computational efficiency of the proposed models in larger, more complex task distributions.

Why it matters
This work has significant implications for the field of program synthesis and machine learning, particularly in the development of algorithms that can adapt to non-stationary environments. By demonstrating that human abstraction learning is driven by prospective compression, the authors provide a new framework for designing more effective online learning algorithms. This could lead to advancements in various applications, including automated programming, adaptive learning systems, and AI-driven design tools, where the ability to anticipate future requirements is crucial for success.

Authors: Leonardo Hernandez Cano, Ivan Zareski, Luisa El Amouri, Pinzhe Zhao, Max Mascini, Emanuele Sansone, Yewen Pu, Bonan Zhao et al.
Source: arXiv:2605.09985
URL: https://arxiv.org/abs/2605.09985v1