🔥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

Qualcomm AI Research is looking for interns (PhD/Master's) in Toronto, in the area of LLMs, multi-modality and agents. Job posting: tinyurl.com/3tyavey5

Excited to present our #ICLR2024 paper “Look, Remember and Reason: Grounded Reasoning in Videos with Language Models” (arxiv.org/pdf/2306.17778…). Our method: LRR, is current ranked 1st on the STAR leaderboard: eval.ai/web/challenges… 1/3

A widely held believe is that difficult vision tasks require vision components like object detectors during inference. We show that we can push any of that stuff into the training stage and get SOTA on challenging visual reasoning tasks. End-to-end visual reasoning on raw pixels.

I remember that well, have been looking for it recently but couldn't find it. Also features Jitendra Malik as far as I remember.

Even finnicky visual reason tasks are best solved end-to-end. By distilling visual subroutines, like object detection, into the model during training... Check out our poster at #ICLR2024

I'll be giving a talk online next week on language models that "see" and interact with you in real-time.

It is tempting to view the context window of an LLM like an "array" - with elements you can access like in an addressable memory. In this work, we argue that this entrenched, but wrong, view may be at the heart of problems like the inability of LLMs to length-generalize.

#COLM24 is live! Head over to paper-central to browse the proceedings and access each paper's 🤗 paper-page, where you can check out discussions, open-source resources, GitHub links, and OpenReview peer review conversations—all in one spot. By the way, this new conference

Yes! And there are many more.

🚨 Don't miss out our #NeurIPS2024 (D&B track) poster "Live Fitness Coaching as a Testbed for Situated Interaction". arXiv: arxiv.org/abs/2407.08101 Dataset: qualcomm.com/developer/soft… Code: github.com/Qualcomm-AI-re… (coming shortly) 📅 Fri 13 Dec 4:30 p.m. PST — 7:30 p.m. PST

Join us at the CVPR 2025 Workshop on Vision-based Assistants in the Real-world (VAR) and tackle one of AI's biggest challenges: building systems that can comprehend and reason about dynamic, real-world scenes. Workshop Page: varworkshop.github.io #CVPR2025 1/2

There is a very simple possible explanation: The transformer architecture lacks a simple inductive bias - that of inductive ("step-by-step") inference. The lack of that bias leads to insane data inefficiency. If true, humans + RNNs will learn this task easily with much less data.

A few years ago I lost a bet, saying by 2021 you could have a video call with someone without being able to tell if it's AI or human. It's 2025 and we're half-way there (the missing half is making the AI model see properly through the camera, which is still an unsolved problem).

No existing AI model can talk to a user in the real world and understand what's happening _right now_. This is called situated AI and it's still a wide open problem.

Binary parity ("count the number of 1s in a bit-string") is a common task to show that transformers cannot generalize. Turns out, a random(!) RNN (train only readout) can learn the task easily, and it can do so with as few as 2 (two) training examples...: arxiv.org/pdf/2505.21749