Shruti Joshi

@_shruti_joshi_

phd student in identifiable repl. prev. research programmer @MPI_IS Tübingen, undergrad @IITKanpur '19.

Montreal, Canada shrutij01.github.io Joined August 2018

940 Following 686 Followers 231 Posts

Latest posts 5

Pinned post

Shruti Joshi @_shruti_joshi_ Aug 18

How can we reliably measure whether a representation identifies latent factors? Can current identifiability metrics evaluate interpretability or unsupervised mechanism discovery?

Our UAI 2026 paper finds systematic false positives and false negatives in these evaluations.

0:11

Shruti Joshi @_shruti_joshi_ Nov 3

Do check out our EMNLP tutorial on why identifiability and causality matter so much for interpretability!

It was brilliant fun putting this together with @rpatrik96, @amuuueller, and @dhanya_sridhar ✨

Huge thanks to everyone who came along and participated!!

Aaron Mueller @amuuueller Nov 2

Interested in putting interpretability on more rigorous foundations? Check out our #EMNLP2026 tutorial slides!

I learned a ton from my fantastic collaborators @rpatrik96, @_shruti_joshi_, @dhanya_sridhar. Recommended for newcomers and experts alike!

📚 tinyurl.com/causality4mech…

View quoted post on X

Shruti Joshi @_shruti_joshi_ Aug 20

Drop by our poster today @UncertaintyInAI to chat about compositional generalisation, sparse coding, and how existing SAE architectures can be made better at generalising OOD

Shruti Joshi @_shruti_joshi_ Apr 1

SAEs fail at OOD tasks. Why?

Features in superposition are linearly representable but not linearly accessible. Instead of discarding sparse coding, we embrace the geometry of superposition and use methods equipped to handle the nonlinearity it induces.

View quoted post on X

Shruti Joshi @_shruti_joshi_ Aug 3

Thrilled to co-organise the first InterpScience workshop at NeurIPS! As interpretability and other fields across the sciences bring complementary systems-level views, we're bringing everyone together to sharpen how we understand LLMs and test that understanding. Papers due Aug 28

Dhanya Sridhar @dhanya_sridhar Aug 3

📣 We are organizing the first InterpScience Workshop @ NeurIPS 2026 in Sydney! The goal of the workshop is to build a more rigorous scientific foundation of LLM interpretability.

📝 Papers due: Aug 28
🌐 interpscience.github.io
✉️ interpscience@gmail.com

[1/7]
View quoted post on X

Shruti Joshi @_shruti_joshi_ Jul 7

Go catch @vpacela presenting our work on why SAEs fall short for compositional OOD generalisation under superposition and how sparse coding can help, at the CompLearn workshop at #ICML2026.
arxiv.org/abs/2603.28744

arxiv.org Stop Probing, Start Coding: Why Linear Probes and Sparse... The linear representation hypothesis states that neural network activations encode high-level concepts as linear mixtures. However, under superposition, this encoding is a projection from a...

Vitória Pacela @vpacela Jul 6

I'm in Seoul for ICML! Excited to meet people and present this as a poster at the Compositional Learning Workshop! Saturday, 4 pm, Auditorium.
View quoted post on X