Our LabelBench work has been accepted to the DMLR journal🎉 Super smooth experience and highly recommended for anyone working on data centric ML work. Check out labeltrain.ai: our broader set of label efficient learning work More on LabelBench: x.com/jifan_zhang/st…

Check out our latest paper in data selection for #CLIP pretraining! by simply switching from CLIPScore filter to negCLIPLoss filter, we have achieved universal improvements. Our additional techniques have also led us to set a new SOTA on #Datacomp-medium.

Check out our new paper on adapting language models to new user preferences in decoding time! Key idea: Optimality analysis based on Legendre transform shows that greedy logits merging is better!

Sorry for raining on o1’s parade, but the more we train LLMs, the worse they get on generating and understanding captions for the New Yorker cartoons. 🧵on our latest study on humor in AI: •Large gap between AI & top human submissions •Dataset with 250M+ human ratings •New

Microsoft Research Deep Learning Group and Microsoft GenAI are hiring self-motivated part-time research interns working on long sequence modeling. We have hundreds of H100/A100 dedicated to this project. Please send CV to [email protected] and [email protected].

Happy to share that our paper was accepted to #NeurIPS2024. Hope to see you in Vancouver (if my visa goes smoothly...😀)

Excited to announce that our work has been accepted by #NeurIPS2024 as a spotlight! 🤩🤩Hope to talk with people in Vancouver!

🚨Curious how LLMs deal with uncertainty? In our new #EMNLP2024 Findings paper, we dive deep into their ability to abstain from answering when given insufficient or incorrect context in science questions 💡arxiv.org/pdf/2404.12452 Joint work w/ Bill Howe Lucy Lu Wang UW iSchool

I will be at #NeurIPS2024! 🚨On academic/industry job market this year 🚨and excited to catch up in person! My research focused on data-efficient learning algorithms across model lifecycle (pre/post-training & test-time). Bridging classical theories (data selection, active

I will be at #NeurIPS2024 from Dec. 10 - 15. Looking forward to catching up with new and old friends and talking anything interesting! I will also present our negCLIPLoss & NormSim paper (spotlight): 📍West Ballroom A-D #7205 📅Dec. 12, 4.30 - 7.30p.m. Hope to see you all!

2022: I never wrote a RL paper or worked with a RL researcher. I didn’t think RL was crucial for AGI Now: I think about RL every day. My code is optimized for RL. The data I create is designed just for RL. I even view life through the lens of RL Crazy how quickly life changes

East Exhibit Hall A-C #3505 Dec.11 4:30-7:30pm Come to check out!

1/3 Today, an anecdote shared by an invited speaker at #NeurIPS2024 left many Chinese scholars, myself included, feeling uncomfortable. As a community, I believe we should take a moment to reflect on why such remarks in public discourse can be offensive and harmful.

feeling a bit under the weather this week … thus an increased level of activity on social media and blog: kyunghyuncho.me/i-sensed-anxie…

Introducing StoryEval: our new video generation benchmark! Can a model present short stories like 'How to put an elephant in a refrigerator'? arXiv: arxiv.org/abs/2412.16211

We released Phi-4-mini (3.8B base in LLM), a new SLM excelling in language, vision, and audio through a mixture-of-LoRA, uniting three modalities in one model. I am so impressed with its new audio capability. I hope you can play with it and share with us your feedback. We also

🧠 Ever notice how LLMs struggle with familiar knowledge in unfamiliar formats? Our new paper "CASCADE Your Datasets for Cross-Mode Knowledge Retrieval of Language Models" tackles this head-on! 🔍 Our findings: - Created a qualitative pipeline demonstrating problem we call

Previous works study the sample complexity of DPO and emphasize the role of samplers in online DPO. What about its role in optimization convergence rates? Check out our paper at #ICLR2025 on convergence rates of online DPO with various samplers! ArXiv: arxiv.org/pdf/2409.19605.

🧠 Your LLM should model how you think, not reduce you to preassigned traits 📢 Introducing LoRe: a low-rank reward modeling framework for personalized RLHF ❌ Demographic grouping/handcrafted traits ✅ Infers implicit preferences ✅ Few-shot adaptation 📄 arxiv.org/abs/2504.14439