CogInterp Workshop @ NeurIPS 2025

At Upper Level Room 5AB of the conference venue!

[CogInterp Workshop @ NeurIPS 2025 @CogInterp]

We’re excited to announce the first workshop on CogInterp: Interpreting Cognition in Deep Learning Models @ NeurIPS 2025! 📣

How can we interpret the algorithms and representations underlying complex behavior in deep learning models?

🌐 https://t.co/sKn7LYWtR7

1/

[NYU Center for Data Science @NYUDataScience]

Can LLMs evolve human-like semantic categories?

CDS-affiliated @NogaZaslavsky and PhD student Nathaniel Imel show that, via simulated cultural transmission, LLMs reorganize color categories toward efficient compression.

🔗https://t.co/Dhiy0r1ScL

[Ari Holtzman @universeinanegg]

this slide is solid gold

[Goodfire @GoodfireAI]

Our last Stanford guest lecture - @EkdeepL on what counts as an explanation & a neuro-inspired "model systems approach" to interp

Plus, how in-context learning and many-shot jailbreaking are explained by LLM representations changing in-context (as a case study for that approach)

00:33 - What counts as an explanation?
04:47 - Levels of analysis & standard interpretability approaches
18:19 - The "model systems" approach to interp
[Case study on in-context learning]
23:36 - How LLM representations change in-context
44:10 - Modeling ICL with rational analysis
1:10:54 - Conclusion & questions

Thanks again to @SuryaGanguli for having us in his class!

[Christopher Potts @ChrisGPotts]

Safety-oriented interpretability researchers should be focused on AI systems, not individual model artifacts. A snippet from the NeurIPS CogInterp workshop panel on Sunday:

[Noga Zaslavsky @NogaZaslavsky]

Honored and thrilled that our work received the @CogInterp best paper award! 💫

📄 Extended paper: https://t.co/TrQcZgrygv

🧵 Highlights: https://t.co/5xRT08StX3

@NeurIPSConf #NeurIPS2025

[Ari Holtzman @universeinanegg]

this was so awesome. Jay still killin' it five decades later

[CogInterp Workshop @ NeurIPS 2025 @CogInterp]

Our Best Paper Award goes to Nathaniel Imel and Noga Zaslavsky @NogaZaslavsky for their excellent paper “Culturally transmitted color categories in LLMs reflect a learning bias toward efficient compression”!

[Justin Angel @JustinAngel]

At the @CogInterp workshop at NeurIPS.

https://t.co/owG9CXhXj3

This slide explains MechIntrep vs CongIntrep:

[CogInterp Workshop @ NeurIPS 2025 @CogInterp]

We are about to start our panel discussion, join us for some hot takes about what cognitive interpretability should be about.

[CogInterp Workshop @ NeurIPS 2025 @CogInterp]

Our final speaker @sydneymlevine makes a radical proposal: building computational models of human moral judgements to use as an AI system for making moral judgements.

[CogInterp Workshop @ NeurIPS 2025 @CogInterp]

Jay proposes shifting from representing context as a sequence of tokens to a sequence of thoughts. The model learns a latent 'thought gestalt' from previous sentences to guide downstream prediction.

[CogInterp Workshop @ NeurIPS 2025 @CogInterp]

Jay McClelland, opens with a question, "Do LMs have thoughts?" Are LMs stochastic parrots or is there some understanding?

[CogInterp Workshop @ NeurIPS 2025 @CogInterp]

Visualizing how LLMs handle object-property binding, he argues that even with scale, transformers might not be forming the kind of 'integrated representations' that human cognition relies on.

[CogInterp Workshop @ NeurIPS 2025 @CogInterp]

A big crowd for Jay McClelland’s talk!

[CogInterp Workshop @ NeurIPS 2025 @CogInterp]

Swing by a super happening poster session where ML and CogSci meet!

[CogInterp Workshop @ NeurIPS 2025 @CogInterp]

In our fourth spotlight talk, neural network legend Paul Smolensky uses symbolic programs such as production systems to understand how neural networks process symbols

[CogInterp Workshop @ NeurIPS 2025 @CogInterp]

For our third spotlight talk, Sonia Murthy @soniakmurthy uses probabilistic cognitive models to understand value trade-offs in LLMs that enable pragmatic reasoning about politeness in speech acts

[CogInterp Workshop @ NeurIPS 2025 @CogInterp]

Erin Grant @ermgrant discusses dissociations between function and representation, and asks whether representational alignment is enough for understanding deep neural networks

[Sonia Murthy @soniakmurthy]

Excited to be presenting our work on using cognitive models to interpret pluralistic values in LLMs once again as a spotlight talk 🌟 at the NeurIPS CogInterp workshop!

Come by upper level room 5AB today and check out the paper here: https://t.co/feDaH3RvKY