
🧵1/10 LLMs can answer in many languages. But do they think in them? Even when prompted in Swahili or Thai, models often switch to English for reasoning. This breaks interpretability and trust. So we ask: Can LLMs reason in the input language?

Fiete group @MIT, theoretical and computational neuroscience. Run by the group. Follows and retweets are not an endorsement.

🧵1/10 LLMs can answer in many languages. But do they think in them? Even when prompted in Swahili or Thai, models often switch to English for reasoning. This breaks interpretability and trust. So we ask: Can LLMs reason in the input language?

1/ Excited to share our recent work in #ICML2025, “A multi-region brain model to elucidate the role of hippocampus in spatially embedded decision-making”. 🎉 🔗 minzsiure.github.io/multireg… Joint w/ @FieteGroup @jaedong_hwang, @brody_lab, @PrincetonNeuro ⬇️ 🧵 for key takeaways that lead to an ongoing neurophysiological experimental verification of our prediction on grid cell code!

This paper got into @Nature!!! 🚀🚀🚀 Look at @SarthakChandra’s thread for a summary
🚨New Preprint! Wondered how grid cells form multiple discrete modules? Interested in continuous attractors and modularity? With @FieteGroup, we discover + generalize a physical mechanism for forming modules from smoothly varying parameters in a dynamical system!👇(1/15)

1/ Our paper appeared in @Nature today! nature.com/articles/s41586-0… @FieteGroup @KhonaMikail. Explains emergence of multiple grid cell modules, w/ excellent match to data! Novel mechanism for applying across vast systems from development to ecosystems. 🧵👇

1/ Super-excited to share our new work “Episodic and associative memory from spatial scaffolds in the hippocampus”, out in Nature today! nature.com/articles/s41586-0… @FieteGroup @sugsharma @rischaudhuri. Key insights and ideas 👇#tweeprint

Pleased to share that our work on "Episodic and associative memory from spatial scaffolds in the hippocampus" has been published in Nature today.
How does one brain circuit encode memories of both places and events? 🧠 The answer is out in Nature today! (proud to have lead this work as a Co-First author @FieteGroup ) 🔗 nature.com/articles/s41586-0… Talk: youtube.com/watch?v=P4kHqk0F… This work extends MESH: proceedings.mlr.press/v162/s…

How does one brain circuit encode memories of both places and events? 🧠 The answer is out in Nature today! (proud to have lead this work as a Co-First author @FieteGroup ) 🔗 nature.com/articles/s41586-0… Talk: youtube.com/watch?v=P4kHqk0F… This work extends MESH: proceedings.mlr.press/v162/s…
Finally a tweeprint on our recent preprint presenting Vector-HaSH ! Vector-HaSH extends MESH to unify two important and seemingly independent roles of hippocampus: Spatial Mapping and Episodic Memory ! Brief talk: youtube.com/watch?v=P4kHqk0F… Preprint: biorxiv.org/content/10.1101/… 1/n

Finally a tweeprint on our recent preprint presenting Vector-HaSH ! Vector-HaSH extends MESH to unify two important and seemingly independent roles of hippocampus: Spatial Mapping and Episodic Memory ! Brief talk: youtube.com/watch?v=P4kHqk0F… Preprint: biorxiv.org/content/10.1101/… 1/n

Thank you so much @SebastianSeung for appreciating our work! I am glad to hear that you consider it a breakthrough in theoretical neuroscience! For folks who would like to know more: Short talk on Vector HaSH youtu.be/P4kHqk0FltY?si=H2wn… Short talk on MESH youtu.be/2l7gF_cn-1Y?si=TMyh…
+1 to an amazing conceptual breakthrough in theoretical neuroscience that I think did not depend on big data or fancy technologies. What we can say is that BRAIN technologies will make the theory testable in a much more conclusive way than was ever possible.

Thanks @doristsao for your kind & thoughtful words about our work! I am glad that you find it beautiful, creative and conceptually insightful! For folks who would like to know more: Short talk on Vector-HaSH youtu.be/P4kHqk0FltY?si=H2wn… Short talk on MESH youtu.be/2l7gF_cn-1Y?si=TMyh…
Btw, I don't know if @FieteGroup is funded by BRAIN, but their recent work understanding grid cell-place cell dynamics as a general mechanism for episodic memory that factorizes problem of building attractors from problem of assigning content to them, is such a beautiful and creative synthesis of a large amount of observational data leading to an entirely new conceptual insight. The experimental work underpinning it goes back to John O' Keefe 1976.

Tweet thread coming up!
This work from @FieteGroup looks pretty cool: biorxiv.org/content/10.1101/… Showing how structure in the sensory cortex (hierarchy, topography, etc.) can emerge from self-organizing dynamics and spontaneous activity on a cortical sheet. Looking forward to reading this one.

Excited to announce: 🔥🧠 Associative Memory Under the Probabilistic Lens 🧠🔥 w @KhonaMikail @Andr3yGR @sanmikoyejo @neurostrow @FieteGroup & Nika Zahedi Appearing @ #NeurIPS2023 Associative Memory & Hopfield Networks Workshop ! 🧵👇 1/7

Announcing #2 in our #NeurIPS2023 workshop & conference papers series (2/10)!! 🔎🧠Testing Assumptions Underlying a Unified Theory for the Origin of Grid Cells🔎🧠 w @KhonaMikail @FieteGroup @sanmikoyejo & Adrian Bertagnoli Paper: openreview.net/forum?id=vYix… 🧵👇 1/8

Excited to begin announcing our #NeurIPS2023 workshop & conference papers (1/10)! 🔥🚀An Information-Theoretic Understanding of Maximum Manifold Capacity Representations🚀🔥 w/ amazing cast @vclecomte @BerivanISIK @sanmikoyejo @ziv_ravid @Andr3yGR @KhonaMikail @ylecun 1/7

Mechanistic interpretability is not only for ML or LLM, but is even more promising for Science! A few months ago, we proposed a brain-inspired method (BIMT) for NN interpretability; Now we're happy to see that it can give something back to neuroscience - Growing brains in RNNs!

In this work, with @ZimingLiu11, we solved a problem that I’ve been struggling with for years: How to make functional modules emerge in RNNs that are also anatomically/topographically segregated! Have a look —
Mechanistic interpretability is not only for ML or LLM, but is even more promising for Science! A few months ago, we proposed a brain-inspired method (BIMT) for NN interpretability; Now we're happy to see that it can give something back to neuroscience - Growing brains in RNNs!