Is it possible to infer the information content of LLMs' activations w/o intervention experiments? 🤔 Absolutely! ✅ We propose a method to map activations to specific input subsets, revealing the encoded information w/o altering the model. 🌟 See the linked🧵for more

Saarland is a beautiful region of Germany with an absolutely stellar NLP community. Check out these exciting opportunities in neurosymbolic modeling!

❗Just updated for ICML: - General theorem statements (beyond S4 / Mamba) - Experiments showing SSM extensions *can* solve state tracking - ♟️♟️Fun analysis of chess state tracking complexity (@ chess heads) Paper: arxiv.org/pdf/2404.08819 w/ jackson petty Ashish Sabharwal

What does it mean for language comprehension to be “good-enough”? With Richard Futrell, we present a computational formalization of model of shallow and deep processing using rate-distortion theory in our new #CogSci2024 paper: arxiv.org/abs/2405.08223 (1/n)

Thanks Michael Hahn for hosting me at Universität Saarland for a talk spanning LLMs, Neurosymbolic Reasoning, TALMs, and their challenges. Was amazing to interact with student researchers and of course Michael himself! Hope to have more of these discussions offline! #saarland #reasoning #llm

Although lots of work on sentence processing has been devoted to predictability effects and locality effects, work on how these two factors has been limited and produced mixed results so far. In two new #CogSci2024 papers, we address this question with different methods and data

Need two emergency reviews for EMNLP. Its urgent and Please RT! Topics are math reasoning, application of VLMS. #NLProc #VLM #MathReasoning

Thank you ACL Anthology! In the meantime, Behzad Shayegh has very helpfully compiled a list of the missing papers with as many preprint links as possible: theshayegh.github.io/ACL2024Missing…