🔥The 9th Round of Easy Loan, Earn $40 Reward is in progress❗️ ⏰ Promotion Period: January 15th - Feburary 15th, 2025 👉 Register now and check more details at gate.io/campaigns/358

A nice example and useful discussion here! Here are a few reactions:

Do Intermediate Tokens Produced by LRMs (need to) have any semantics? Our new study "Beyond Semantics: The Unreasonable Effectiveness of Reasonless Intermediate Tokens" lead by kstechly, Karthik Valmeekam Atharva & Vardhan Palod dives into this question 🧵 1/

Keep stumbling on older gems doing sparse dictionary learning for concept extraction. This one’s from 2018, tldr: transform deep features into sparse convex combos of archetypes. arxiv.org/pdf/1805.11155

BREAKING: Trump administration bans Harvard from enrolling international students

Without its international students, Harvard is not Harvard. hrvd.me/IntStudents25t

What are the organizing dimensions of language processing? We show that voxel responses are organized along 2 main axes: processing difficulty & meaning abstractness—revealing an interpretable, topographic representational basis for language processing shared across individuals.

Shreyas Gite Ville 🤖 the robotics field needs an observability layer to scan models and datasets. the most convenient way would be an extension of CRAFT for multimodal inputs. huggingface.co/papers/2211.10… Thomas Fel

My research group is recruiting MSc and PhD students at the University of Alberta in Canada. Research topics include generative modeling, representation learning, interpretability, inverse problems, and neuroAI. If interested, check my website and consider applying.

The Worldwide LeRobot hackathon is in 2 weeks, and we have been cooking something for you… Introducing SmolVLA, a Vision-Language-Action model with light-weight architecture, pretrained on community datasets, with an asynchronous inference stack, to control robots🧵

I'm thrilled to announce the release of FastPlaid ! 🚀🚀 FastPlaid is a high-performance engine for multi-vector search, built from the ground up in Rust (with the help of Torch C++)⚡️ You can view FastPlaid as the counterpart of Faiss for multi vectors.

Looking forward to speaking today at Harvard’s Kempner Institute at Harvard University NeuroAI conference on theories of learning, creativity and reasoning. Looks like a great set of speakers. Virtual attendance is possible: kempnerinstitute.harvard.edu/frontiers-in-n…

🚨 New paper alert! Linear representation hypothesis (LRH) argues concepts are encoded as **sparse sum of orthogonal directions**, motivating interpretability tools like SAEs. But what if some concepts don’t fit that mold? Would SAEs capture them? 🤔 1/11

We then analyze what the hierarchical concepts look like, finding an intriguing set of **shared** concepts across modality-specific backbones to enable hierarchical structures. This stands in contrast to the “usual understanding” of disparate representations spaces in VLMs! 8/11

Around CVPR for the next 2 days—if you're into interpretability, SAEs, complexity, or just wanna know how cool Kempner Institute at Harvard University is, hit me up 👋

NEW: Asma Ghandeharioun (Asma Ghandeharioun) of Google DeepMind demonstrates techniques to enhance interpretability of large language models, including the Patchscopes framework. Watch the video: youtu.be/Og8FTqUrvtA #NeuroAI2025 #AI #ML #LLMs #NeuroAI

*Universal Sparse Autoencoders* by Harry Thasarathan Thomas Fel Matthew Kowal Kosta Derpanis They train a shared SAE latent space on several vision encoders at once, showing, e.g., how the same concept activates in different models. arxiv.org/abs/2502.03714

🧵Why are linear properties so ubiquitous in LLM representations? We explore this question through the lens of 𝗶𝗱𝗲𝗻𝘁𝗶𝗳𝗶𝗮𝗯𝗶𝗹𝗶𝘁𝘆: “All or None: Identifiable Linear Properties of Next-token Predictors in Language Modeling” Published at #AISTATS2025🌴 1/9

NEW: George Alvarez (George Alvarez) of Harvard Psychology and the #KempnerInstitute shows how long-range feedback projections can enhance the alignment between #ML models and human visual processing. Watch the video: youtu.be/Ju_eD0Jwa8Q #NeuroAI20205 #AI #neuroscience #NeuroAI

Check out our ACL paper! We use shapley interactions to see which words (and phones) interact non-linearly -- what we lose when we assume linear relationships between features. Chat to Diganta in Vienna!

The Kempner Institute congratulates its research fellows Isabel Papadimitriou (Isabel Papadimitriou) and Jenn Hu (Jennifer Hu) for their faculty appointments (UBC Linguistics & JHU Cognitive Science) and celebrates their innovative research. Read more here: bit.ly/448heBy #AI #LLMs