Minnesota NLP

We warmly welcome you, @AlexanderSpangh ! Excited to have you join our community.

Farewell to our graduating students with good food! 💕🔥

🚨 New paper! 🚨 Legal work isn’t a checklist. It’s full of ambiguity, changing goals, and evolving info. So why is legal AI mostly built for static tasks like clause selection or doc search? 💡 We argue: model the process, not just the product. ArXiv : arxiv.org/abs/2504.18942

🚨 New Paper Alert! 🚨 How can we align language models without drowning in prompt engineering or falling into reward hacking traps? We introduce Meta Policy Optimization (MPO)—a new reinforcement learning framework that evolves its own reward model rubrics through meta-level reflection. Inspired by metacognition and evaluative thinking, MPO trains models to think about how they evaluate, not just what they generate. 🔥 Why it matters: ✔️ Boosts stability and robustness in RLAIF ✔️ Reduces human labor in prompt crafting ✔️ Generalizes across tasks: essays, summarization, ethical and mathematical reasoning Check it out: huggingface.co/papers/2504.20… Big thanks to co-authors @chanwoopark20 (MIT), @_vipulraheja (Grammarly), and @dongyeopkang (UMN)! #AI #LLMs #ReinforcementLearning #MetaLearning #NLP #Alignment #RLHF #RLAIF #EvaluativeThinking #PromptEngineering

A New Paper Alert ! 🎉🚨 How do researchers write their papers? Introducing 𝐒𝐜𝐡𝐨𝐥𝐚𝐖𝐫𝐢𝐭𝐞: the first-of-its-kind keystroke dataset of an 𝐞𝐧𝐝-𝐭𝐨-𝐞𝐧𝐝 𝐬𝐜𝐡𝐨𝐥𝐚𝐫𝐥𝐲 𝐰𝐫𝐢𝐭𝐢𝐧𝐠 𝐩𝐫𝐨𝐜𝐞𝐬𝐬, with thorough annotations of 15 cognitive writing intentions.

Two exciting projects will be presented at #EMNLP2024 (1) @shirley_ah_ 's diversity extraction from LLMs (2) Karin de Langis's multi-style/reward control in RLHF. Shirley & Karin (& alumni @Ruyuan_Wan @MinnesotaNlp ) are amazingly fun to talk to! Below are links to papers:

Arriving at Miami for #EMNLP2024 🏝️Happy to catch up with old friends and meet new ones 😄✨ Also do check out my work tomorrow & @customnlp4u workshop on Saturday! 👋🏻 📍Poster Session B (Riverfront Hall) 📅 Tue, Nov 12 ⏰ 2pm Workshop: 📍Merrick 1 📅 Sat, Nov 16 ⏰ 9am-5pm

How can we produce generations that are well-balanced in irony, formality, and positivity? Or between negativity and formality? In our recent #EMNLP2024 work, we observe style control as a reinforcement learning problem and introduce a new reward-shaping formulation via dynamic weighting with discriminator gradient magnitudes.

Won't be at #ACL2024, but check out the following work from @MinnesotaNlp , including topics on 🧵 - video alignment with AI feedback - human-vs-machine text detection using discourse motifs - benchmarking cognitive biases in LLM-as-judge - multi-modal, multi-hop reasoning

Just finished my oral presentation at #IC2S2 in Philly :) We propose motif and image modalities as an initial step towards human-friendly encodings of graph data and our methods can be extended to multiple social systems! #NLP #multimodality #NLProc #CSS

Had a lot of fun presenting our paper on LLM modalities for graph reasoning at #NAACL2024 today! 🇲🇽 Please PM me if you'd like to collaborate on future directions of this exciting work! :) Paper link : aclanthology.org/2024.findings-… #nlpproc #llms #graphs

Thank you to @dongyeopkang @snchancellor and all of @MinnesotaNlp and @UMNComputerSci for hosting me today for my talk. It was a pleasure meeting everyone!

Check out our work in the poster session 2 today at #MSLD2024! And the talk tomorrow at 2pm!

Instructions may compose of subtasks: the output from a subtask becomes the input for the next subtask, or called as “chain-of-instructions” (CoI)⛓️ CoI-tuned LLMs perform better on complex tasks & downstream tasks of multilingual summarization! 📎: arxiv.org/pdf/2402.11532…

We're #UMNProud of @dongyeopkang for developing an AI-based text editing system to assist with complex writing tasks! Read more here: z.umn.edu/DKAIEditing

"Paraphrasing attacks" can compromise the effectiveness of AI content detectors. 🙀 Can hierarchical structures in texts help build a more robust detector? Our research reveals a resounding💡YES!💡Delighted to share our work on merging discourse frameworks with graph analysis.

🤖Excited to share @MinnesotaNlp Labs latest work on Data Efficient Instruction Tuning: "SelectLLM: Can LLMs Select Important Instructions to Annotate?". Leveraging LLMs to select high quality instructions. Arxiv: arxiv.org/abs/2401.16553 #LLM #NLP #AI #ArtificialIntelligence

Check out our work on investigating artifacts in LLM-generated data! 🧐 We also provide suggestions to mitigate these issues for safer usage of LLM for generating data🙂

Consider submitting your work to #In2Writing workshop! ✍🏻📑📝

My first undergraduate, Ryan Ko @im_koogusry, received an honorable mention in CRA URA. He co-authored multiple interesting papers, including CoEdit, Cobbler, Writing-Decoding, Meta-crafting, etc. ⚠️ Currently, he is applying for CS PhD programs on RLHF, LLM alignment, etc!! ⚠️