Latent learning: episodic memory complements parametric learning by enabling flexible reuse of experiences (2509.16189v1)

Abstract: When do machine learning systems fail to generalize, and what mechanisms could improve their generalization? Here, we draw inspiration from cognitive science to argue that one weakness of machine learning systems is their failure to exhibit latent learning -- learning information that is not relevant to the task at hand, but that might be useful in a future task. We show how this perspective links failures ranging from the reversal curse in language modeling to new findings on agent-based navigation. We then highlight how cognitive science points to episodic memory as a potential part of the solution to these issues. Correspondingly, we show that a system with an oracle retrieval mechanism can use learning experiences more flexibly to generalize better across many of these challenges. We also identify some of the essential components for effectively using retrieval, including the importance of within-example in-context learning for acquiring the ability to use information across retrieved examples. In summary, our results illustrate one possible contributor to the relative data inefficiency of current machine learning systems compared to natural intelligence, and help to understand how retrieval methods can complement parametric learning to improve generalization.

• The paper demonstrates that incorporating oracle episodic retrieval overcomes the reversal curse by enabling models to generalize latent information beyond explicit training.
• It introduces novel benchmarks—codebooks, simple reversals, and latent gridworld navigation—to rigorously evaluate latent learning capabilities.
• Empirical results reveal that retrieval-augmented architectures significantly improve generalization in both supervised and reinforcement learning environments.

Latent Learning and Episodic Memory: Complementary Mechanisms for Flexible Generalization in AI

Introduction

This paper investigates the limitations of parametric learning in current machine learning systems, particularly their inability to perform latent learning—acquiring information not immediately relevant to the current task but potentially useful for future tasks. Drawing on cognitive science, the authors argue that episodic memory, implemented as nonparametric retrieval, can complement parametric learning by enabling flexible reuse of experiences. The work formalizes latent learning, introduces novel benchmarks, and empirically demonstrates that oracle episodic retrieval mechanisms can overcome key failures of generalization in both supervised and reinforcement learning settings.

Figure 1: The reversal curse illustrates how parametric learners consolidate information in a task-dependent manner, failing to generalize to reversed relations even when the reversed relation is implied by the training data.

Formalization of Latent Learning

The authors provide a formal sketch distinguishing latent learning from standard generalization. Given input sequences $x \in X^*$ and task cues $t \in T^*$, models are trained to map $(x, t)$ to outputs $f(x, t)$. While parametric learners generalize within the training distribution, they fail to encode information about alternate tasks $t'$ not explicitly cued during training, even if the necessary information is present in $x$. This is exemplified by the reversal curse: models trained on "X taught Y" fail to answer "Who taught Y?" without explicit context, despite being able to do so when the relevant information is in context.

Benchmarks for Latent Learning

To empirically evaluate latent learning, the paper introduces four benchmarks:

• Codebooks: Tests the ability to use latent codebook indices for encoding sequences when only definitions (not usage) are seen during training.
• Simple Reversals: Assesses generalization to reversed relations not seen during training.
• Semantic Structure: Evaluates reasoning over naturalistic text, including reversals, syllogisms, and category-inclusion holdouts.
• Latent Gridworld Navigation: Inspired by rodent latent learning, agents must navigate to objects never used as goals in training but frequently encountered.

Figure 2: Codebooks benchmark tests the use of latent indices for encoding, highlighting failures of parametric generalization.

Figure 3: Simple reversals benchmark evaluates the ability to generalize to reversed relations.

Methods

Models are trained using standard transformer architectures for supervised tasks and IMPALA for RL tasks. Oracle retrieval is implemented by prepending relevant episodes or documents to the context, simulating perfect episodic memory. Importantly, retrieval is not used for direct supervision; gradients are propagated into the context, but the model is not trained to predict the retrieved information.

Empirical Results

Failures of Latent Learning in Parametric Models

Baseline transformer models exhibit high performance on standard generalization tasks but fail on latent learning tests. For example, in the codebooks benchmark, models can recall definitions and encode sequences with trained indices but cannot encode sequences using latent indices unless the definition is present in context.

Figure 4: Baseline models fail to generalize to latent encoding tasks, despite high performance on related validation tasks.

Benefits of Oracle Retrieval

Equipping models with oracle retrieval enables successful generalization to latent tasks. On both codebooks and simple reversals, retrieval models achieve above-chance performance on latent tests, demonstrating that episodic memory can bridge the latent learning gap.

Figure 5: Oracle retrieval enables generalization to latent encoding tasks in the codebooks benchmark.

Figure 6: Oracle retrieval enables generalization to latent reversal tasks.

Semantic Structure and Associative Cues

In the semantic structure benchmark, strong associative cues allow baseline models to generalize to some latent tasks. When associative cues are reduced, the advantage of retrieval becomes more pronounced, though benefits are muted without sufficient in-context learning (ICL) examples in training.

Figure 7: With strong associative cues, both baseline and retrieval models generalize well; retrieval advantage increases when cues are reduced.

Gridworld Navigation

In both RL and BC variants of the gridworld task, retrieval substantially improves performance on latent navigation goals, though absolute performance remains below ceiling, reflecting the complexity of multi-step reasoning and action planning.

Figure 8: Retrieval improves generalization to latent navigation goals in gridworld environments.

Importance of In-Context Learning

Ablation studies reveal that within-example ICL during training is necessary for models to learn to use retrieved memories effectively. Without ICL-supporting sequences, retrieval models fail to achieve strong performance on latent tests.

Theoretical and Practical Implications

The findings highlight a fundamental limitation of parametric learning: its inability to encode and flexibly reuse information not directly relevant to the training task. Episodic memory, implemented as retrieval, provides a mechanism for overcoming this limitation by reinstating relevant experiences into context, enabling in-context reasoning. This has direct implications for the design of AI systems, suggesting that retrieval-augmented architectures (e.g., RAG) can achieve forms of generalization inaccessible to pure parametric learners.

The results also clarify the role of associative learning and data diversity in enabling certain types of generalization, and suggest that training-time augmentation and episodic retrieval are complementary strategies. The boundary between latent learning and associative generalization remains an open area for future research.

Future Directions

Key avenues for future work include:

• Developing scalable and efficient episodic memory and retrieval mechanisms that approximate the flexibility of human memory.
• Investigating the interaction between data diversity, associative cues, and latent learning in large-scale models.
• Exploring hybrid approaches that combine training-time augmentation with test-time retrieval for robust generalization.
• Extending latent learning benchmarks to more complex, real-world tasks and multimodal settings.

Conclusion

This paper provides a rigorous analysis of latent learning as a critical gap in current AI systems and demonstrates that episodic memory, via retrieval, can complement parametric learning to enable flexible reuse of experiences. The empirical results across diverse benchmarks substantiate the claim that retrieval mechanisms are essential for achieving human-like generalization in artificial agents. The work motivates further research into episodic memory architectures and hybrid learning strategies to bridge the gap between natural and artificial intelligence.