LLMs can appear to reason well, but a single wrong token can derail the whole output. Our new work shows that token-level memorization is a key cause of failure, especially under distribution shift. Introducing: STIM 🔍🧠 🧵 #NLProc

Excited to be speaking at Berkeley #SBC2025 #BASS2025 about convergence of AI x Web3 and how we could create an open, collaborative economy for an AI-driven future. Our blockchain product lead shares the tech architecture for how we achieve on-chain monetization flows for AI assets. @BerkeleyRDI @StanfordSBA

I passed my thesis defense today! 😊 It's been a rewarding journey at @nlp_usc @CSatUSC. Many thanks to @xiangrenNLP @robinomial @swabhz @avestime, and everyone who supported me!

I’ll be at @aclmeeting next week to present this work! 🇦🇹 Excited to meet old friends and make new ones. Let’s catch up if you like thinking more about the future of human-centred NLP, personalization and multi-turn interactions or just wanna get some nice Viennese coffee ☕️

1+1=3 2+2=5 3+3=? Many language models (e.g., Llama 3 8B, Mistral v0.1 7B) will answer 7. But why? We dig into the model internals, uncover a function induction mechanism, and find that it’s broadly reused when models encounter surprises during in-context learning. 🧵

Data Services Platform (DSP) is LIVE! 🔆 Now anyone, anywhere in the world, can contribute to AI development and earn real rewards for their work. 🔆 $450K+ in $SAHARA + partner rewards available day one! Get started today → #AIforALL

We @Zai_org are thrilled to open-source GLM-4.1V-9B-Thinking, a VLM that can think with long CoTs. SoTA in <10B VLMs, comparable to Qwen-2.5-VL-72B in 18 tasks. One RL to rule them all! Details - Tech report: - Code:

I didn't believe when I first saw, but: We trained a prompt stealing model that gets >3x SoTA accuracy. The secret is representing LLM outputs *correctly* 🚲 Demo/blog: 📄: 🤖: 🧑‍💻:

Here's a recent talk I gave recapping the last 6-12 months of AI progress, why getting perfect models is hard, how labs are likely approaching the next phase of training (for agents), and other interesting tidbits across the reasoning landscape. Topics: 00:00 Introduction & the state of reasoning 05:50 Hillclimbing imperfect evals 09:18 Technical bottlenecks 13:02 Sycophancy 18:08 The Goldilocks Zone 19:28 What comes next? (hint, planning) 26:40 Q&A YouTube etc in replies. Thanks @corbtt and @OpenPipeAI for hosting me.

Can LLMs appropriately explain “Why is the sky blue?” …to a 10-year-old 👶🏽 vs. someone with a PhD in physics 👩🏽‍🔬? In our #ACL2025 paper, we evaluate how well how well LLMs can tailor their explanations to different people.✍️ 🔗 🧵 (1/n)