Martin Engelcke

Take a look at Andrew's thread or the paper for more details! 6/6

We suggest that retrieving relevant experiences from an episodic memory during training and testing could be a potential part of the solution for overcoming these limitations, and we show that this leads to better generalisation on the benchmarks above. 5/

That is, even when the model has all the necessary "pieces" of knowledge for solving a task. 4/

In contrast, we show on several benchmarks (codebook translation, simple reversals, semantic structures, and gridworld navigation) that parametric models often fail to apply information they have encountered before during training to new tasks. 3/

Tolman conducted experiments showing that rats are able to navigate to food and water in a maze more quickly if they have seen the maze before, even if they were previously not motivated to seek out food and water. 2/

Our work on "Latent learning: episodic memory complements parametric learning by enabling flexible reuse of experiences" led by @lampinen.bsky.social and with Effie Li, @arslanchaudhry.bsky.social, and James McClelland is now available on arXiv!

Link: arxiv.org/abs/2509.16189

Thread: 1/

Hello, World!