# Lessons from Neuroscience for AI: How integrating Actions, Compositional Structure and Episodic Memory could enable Safe, Interpretable and Human-Like AI | Mahault Albarracin Ph.D. **By:** Mahault Albarracin Ph.D. **Published:** 2026-01-02T16:20:54.769Z **Source:** [LinkedIn](https://www.linkedin.com/posts/mahault-albarracin-ph-d-1742bb153_lessons-from-neuroscience-for-ai-how-integrating-activity-7412890645733949440-KfJQ?rcm=ACoAAAfsct4BVBNkOVI5cyYn-UOIIe2RSRhT-cI) --- # Mahault Albarracin Ph.D.’s Post [Mahault Albarracin Ph.D.](https://ca.linkedin.com/in/mahault-albarracin-ph-d-1742bb153?trk=public_post_feed-actor-name) Interesting new paper. In essence, LLMs sortof minimize VFE, but only very locally, and do not use any sort of compositionality or causal structure, which ultimately limits their ability to reason at all. [https://lnkd.in/ePj8sMy9](https://www.linkedin.com/redir/redirect?url=https%3A%2F%2Flnkd%2Ein%2FePj8sMy9&urlhash=yUhq&trk=public_post-text) [ Lessons from Neuroscience for AI: How integrating Actions, Compositional Structure and Episodic Memory could enable Safe, Interpretable and Human-Like AI arxiv.org ](https://www.linkedin.com/redir/redirect?url=https%3A%2F%2Farxiv%2Eorg%2Fabs%2F2512%2E22568v1&urlhash=Gv7z&trk=public_post_feed-article-content) There are some very good ideas here. For folks who don’t love deep learning (I’m looking at you, [Denis O.](https://www.linkedin.com/in/denis-o-b61a379a?trk=public_post_comment-text) - whatchoo doing liking this post!? lol), it’s still not getting us out of the dog house. the emphasis on compositionality and causality is very much what i have been arguing for a very long time. the tide is changing! Most of the effort to identify flaws in LLM could be more productive if it considered semantic based logic models. [http://aicyc.org/2025/09/16/extending-sam-to-handle-causality-with-do-calculus](http://aicyc.org/2025/09/16/extending-sam-to-handle-causality-with-do-calculus?trk=public_post_comment-text)/ I find it alarming that we are scaling a gradient optimized statistical model that iterates on weighted guesses optimized by "teaching". At its core there is no belief system or logic engine. The system is great for digesting large amounts of data but not intelligent or adaptive decision making.