🧵 1/N) Excited to share our recent work at Scale AI, "Jailbreaking to Jailbreak (J2)".😈 We present a novel LLM-as-red-teamer approach in which a human jailbreaks a refusal-trained LLM to make it willing to jailbreak itself or other LLMs. We refer to this process as

🔍 Are Verifiers Trustworthy in RLVR? Our paper, Pitfalls of Rule- and Model-based Verifiers, exposes the critical flaws in reinforcement learning verification for mathematical reasoning. 🔑 Key findings: 1️⃣ Rule-based verifiers miss correct answers, especially when presented in

frontier ai today

"Tell, Don't Show" was accepted to #ACL2025 Findings! Our simple approach for literary topic modeling combines the new (language models) with the old (classic LDA) to yield better topics. A possible addition to your CSS/DH research 🛠️ box ✨📚 arxiv.org/abs/2505.23166

this is pretty egregious

Very well-said as always! There's lots of tricks to make a paper perceived good by reviewers, but that doesn't mean it's actually good science Really interested in seeing how we can measure what "good" means :)

🧵 Multilingual safety training/eval is now standard practice, but a critical question remains: Is multilingual safety actually solved? Our new survey with Cohere Labs answers this and dives deep into: - Language gap in safety research - Future priority areas Thread 👇

I'm excited to share that I’ll be joining Univ. of Maryland as an Assistant Professor in Computer Science, where I’ll be launching the Resilient AI and Grounded Sensing Lab. The RAGS Lab will build AI that works in chaotic environments. If you would like to partner, please DM me!

RewardBench 2 is here! We took a long time to learn from our first reward model evaluation tool to make one that is substantially harder and more correlated with both downstream RLHF and inference-time scaling.

🙋 Can LLMs reliably detect PII such as person names? ‼️ Not really, especially if the context has ambiguity. 🖇️ Our work shows that LLMs can struggle to recognize person names in barely ambiguous contexts.

Do you like trivia? Can you spot when AI is feeding you BS? Or can you make AIs turn themselves inside out? Then on June 14 at College Park (or June 21 online), we have a competition for you.

I’m thrilled to share RewardBench 2 📊— We created a new multi-domain reward model evaluation that is substantially harder than RewardBench, we trained and released 70 reward models, and we gained insights about reward modeling benchmarks and downstream performance!

🤔 What if you gave an LLM thousands of random human-written paragraphs and told it to write something new -- while copying 90% of its output from those texts? 🧟 You get what we call a Frankentext! 💡 Frankentexts are surprisingly coherent and tough for AI detectors to flag.

This is such a clever way to make a benchmark, cool work!

𝐖𝐡𝐚𝐭 𝐇𝐚𝐬 𝐁𝐞𝐞𝐧 𝐋𝐨𝐬𝐭 𝐖𝐢𝐭𝐡 𝐒𝐲𝐧𝐭𝐡𝐞𝐭𝐢𝐜 𝐄𝐯𝐚𝐥𝐮𝐚𝐭𝐢𝐨𝐧? I'm happy to announce that the preprint release of my first project is online! Developed with the amazing support of Abhilasha Ravichander and Ana Marasović (Full link below 👇)

Excited to share that I'll be joining the CS department at University of Alberta as an Assistant Professor in January 2026 where I will be affiliated with Amii. I'll be recruiting 2-3 PhD/MSc students and establishing a research lab on AI in Education and Human–AI Interaction.

Evaluating long-form answers to complex technical questions is very challenging. Existing methods fall short in this setting. At Samaya, we built Criteria-Eval, a checklist-based evaluation that aligns with how domain experts judge answers. 🧵samaya.ai/blog/criteria-… ✍️

I'm now a Ph.D. candidate! 🎉🥳 A few weeks ago, I proposed my thesis: "Teaching AI to Answer Questions with Reasoning that Actually Helps You". Thanks to my amazing committee + friends UMD CLIP Lab! 🫶 I won't be back in Maryland for a while, some exciting things coming soon 👀

When it comes to text prediction, where does one LM outperform another? If you've ever worked on LM evals, you know this question is a lot more complex than it seems. In our new #acl2025 paper, we developed a method to find fine-grained differences between LMs: 🧵1/9

What if LLMs could learn your habits and preferences well enough (across any context!) to anticipate your needs? In a new paper, we present the General User Model (GUM): a model of you built from just your everyday computer use. 🧵