Ever wondered if model merging works at scale? Maybe the benefits wear off for bigger models? Maybe you considered using model merging for post-training of your large model but not sure if it generalizes well? cc: Google AI Google DeepMind UNC NLP 🧵👇 Excited to announce my

🚨 New paper 🚨 Excited to share my first paper w/ my PhD students!! We find that advanced LLM capabilities conferred by instruction or alignment tuning (e.g., SFT, RLHF, DPO, GRPO) can be encoded into model diff vectors (à la task vectors) and transferred across model

Cool work on model merging showing results in some practical settings

Check out our new paper: just apply a diff vector from an instruction-tuned model to another non-instruction-tuned model. That’s it. ✅ Instruction-following transfers ✅ Multilingual improves ✅ Alignment (SFT, DPO, RLVR, GRPO) carries over Huge thx to all my collaborators!

Our paper is now available on arXiv: arxiv.org/abs/2503.20110

Almost 7 years ago, Tu Vu and I wrote our first paper together, one of few. It is fantastic to see the first paper by Tu’s students this time. Congratulations and looking forward to many such great works from Tu’s group!

My first PhD paper is out 😆 took 7 months and lots of back-and-forth. Learned so much from Tu — sharp thinking, real feedback, and always pushing the idea further. Also, shoutout to my collaborators and the folks at Virginia Tech Computer Science!

prompt engineering -> thought engineering :-) arxiv.org/abs/2503.24370

Introducing the DeepSeek-R1 Thoughtology -- the most comprehensive study of R1 reasoning chains/thoughts ✨. Probably everything you need to know about R1 thoughts. If we missed something, please let us know.

📢 Research internship Google📢 I am looking for a PhD student researcher to work with me and my colleagues on advanced reasoning and/or RAG factuality this summer Google Mountain View, CA. We will focus on open-source models and benchmarks, and aim to publish our findings.

How does RL improve performance on math reasoning? Studying RL from pretrained models is hard, as behavior depends on choice of base model. 🚨 In our new work, we train models *from scratch* to study the effect of the data mix on the behavior of RL. arxiv.org/abs/2504.07912

✨ New paper ✨ 🚨 Scaling test-time compute can lead to inverse or flattened scaling!! We introduce SealQA, a new challenge benchmark w/ questions that trigger conflicting, ambiguous, or unhelpful web search results. Key takeaways: ➡️ Frontier LLMs struggle on Seal-0 (SealQA’s

More thinking power at test-time doesn't fix noisy-search problems—SealQA proves it. AI's reasoning capabilities fall flat when web search turns messy, and SealQA quantifies that. SealQA introduces an exceptionally challenging benchmark for search-augmented language models,

Excited to share that our paper on model merging at scale has been accepted to Transactions on Machine Learning Research (TMLR). Huge congrats to my intern Prateek Yadav and our awesome co-authors Jonathan Lai, Alexandra Chronopoulou, Manaal Faruqui, Mohit Bansal, and Tsendsuren 🎉!!

This work got accepted at Transactions on Machine Learning Research (TMLR). Congratulations to Prateek Yadav and my co-authors. Also, thank you to the reviewers and editors for their time.

We just evaluated Grok 4 on our SEAL-0 dataset 👍Try it: huggingface.co/datasets/vtllm…

🎉🎉

DeepSeek achieved a strong result on SEAL0, a challenging benchmark for reasoning with conflicting search results. 🎊

LoRA makes fine-tuning more accessible, but it's unclear how it compares to full fine-tuning. We find that the performance often matches closely---more often than you might expect. In our latest Connectionism post, we share our experimental results and recommendations for LoRA.