Do all dims in the hiddn repr. inside transformer contribute twrds decision? Our #NeurIPS2025 work studies the geometry of hiddn repr. (& decision-making dynamics) in LLMs thrgh lens of intrinsic dimension
tinyurl.com/id-llm
Awesm work by @iabhinavjoshi @divyanshubhatt0
Associate Prof@CSE, @IITKanpur. Researching on NLP, ML, RL, AI, Affect, LegalText,...... Tweets are personal views.
Joined May 2014
- In the current Mechanistic Interpretability research atten. heads and MLP are considered atomic units, however we do not think so. Curious to know what is more foundational? Check our #NeurIPS2025 paper: arxiv.org/abs/2511.20273 Awesome work by Abhinav Joshi & Areeb Ahmad
- @ANRFIndia @shivkuma_k The ANRF submission portal is crashing, please help and resolve this at the earliest.
- Is @openreviewnet down? I am not able to access it. @ReviewAcl how do we put the rebuttal for May cycle?
- LLMs are talk of the town but if you always wonder how LLMs are able to perform so well on common sense reasoning, then check out our paper: "Towards Quantifying Commonsense Reasoning with Mechanistic Insights" (arxiv.org/pdf/2504.10077)

