UNC NLP

🚨🎓 We have several PhD (and postdoc) openings in NLP+CV+ML+AI in beautiful Chapel Hill 👇 Please RT+apply & ping our faculty for any questions (application-fee waivers & no GRE requirement)! @mohitban47 @gberta227 @snigdhac25 @TianlongChen4 @shsriva @HuaxiuYaoML + others 🧵

So excited to share this work (which received very high reviewer scores!) led by Sapir! Why wait for a full sentence to find a hallucination? We built a model that detects factual errors on prefixes as they’re generated. This leads to huge faithfulness gains (our 3B model…

🚨 New paper alert! We’re thrilled to share our new preprint “PrefixNLI: Detecting Factual Inconsistencies as Soon as They Arise” ✨ LLMs generate text one token at a time, but factuality checks still wait for a full sentence. We extend NLI to text prefixes, enabling the…

[2/2] 🎤 We’re excited to host an amazing lineup of keynote speakers for #NewSumm2025 at #EMNLP2025! Mohit Bansal (UNC) @mohitban47 Greg Durrett (NYU) @gregd_nlp Arman Cohan (Yale) @armancohan Alexander R. Fabbri (Scale AI) Jey Han Lau (Univ. of Melbourne) #NLProc

🚨 Introducing MIRA — a challenging Visual-CoT benchmark that requires multimodal LLMs to draw to think. Even the strongest models (GPT-5, Gemini 2.5 Pro, o3) score > 80% on popular benchmarks, yet collapse to < 20% on MIRA, where reasoning demands visual imagination. Text-only…

🚨 Excited to share Gistify! Often the easiest way to understand large/complicated repos is by playing around with test cases and tracing back through the code that is executed. Gistify tasks models with turning a codebase and an entry-point (e.g. command, unit test) into a…

I'll be presenting ✨MAgICoRe✨ virtually tonight at 7 PM ET / 8 AM CST (Gather Session 3)! I'll discuss 3 key challenges in LLM refinement for reasoning, and how MAgICoRe tackles them jointly: 1⃣ Over-correction on easy problems 2⃣ Failure to localize & fix its own errors 3⃣…

Last week, I gave a talk on Ethical Issues in LLMs at UNC's Parr Center for Ethics. The extended Q&A session was the best part! Thoughtful questions about bias, responsibility & accountability of AI judgment. Inspiring to see people thinking so deeply about ethics in tech. @unccs

don’t forget to stop by my poster during Gather Session 3 at 8 AM Nov 6 (CST) / 7 PM Nov 5 (ET)! #EMNLP2025

🚨 Proud to share our #TACL work on localizing factual inconsistencies in attributable text generation! To find where LLMs hallucinate, we need to get granular. We introduce QASemConsistency, a new method that decomposes text into simple question-answer pairs to precisely…

Even though I cannot attend EMNLP in person, plz enjoy our work Video-Skill-CoT at Suzhou 🇨🇳 Appreciate all of my collaborators and @mohitban47! 😆

🎉Thanks for the shoutout! I’ll be virtually presenting our new work Video-RTS at #EMNLP2025 (my co-lead @jaeh0ng_yoon will present in person). If you’re into advanced video-reasoning frameworks, check it out: - No SFT, pure RL: trains with simple output-based rewards (GRPO)—no…

Excited to be at #EMNLP2025 in Suzhou! I’ll present our work: (1) MEXA (Fri 12:30 PM CST) about general multimodal reasoning with dynamic multi-expert aggregation and (2) RACCooN (Wed 4:30 PM CST) about editing videos via auto-generated narratives. Please stop by our…

🚨 I'll be presenting virtually tonight at 7PM ET/ 8AM CST on Gather! I'll be talking about how strong LLMs can exploit loopholes introduced by ambiguous instructions, and what that means for safety! P.s. I am hiring Ph.D. students for my lab at UT Austin CS, applications due…

unfortunately I won’t be in Suzhou but I’ll be presenting our paper at Gather Session 3 on Nov 6 at 8 AM CST. save the date! #EMNLP2025

Is language a "terrible abstraction" for video understanding? Many in the video community often dismiss language-driven approaches in favor of complex, video-native solutions. However, I believe this resistance stems more from internal bias—validating a research identity as a…

🔥The deadline (Nov 3, 2025 AoE) for 𝐍𝐞𝐮𝐫𝐈𝐏𝐒 𝟐𝟎𝟐𝟓 𝐖𝐨𝐫𝐤𝐬𝐡𝐨𝐩 𝐨𝐧 𝐒𝐨𝐜𝐢𝐚𝐥𝐥𝐲 𝐑𝐞𝐬𝐩𝐨𝐧𝐬𝐢𝐛𝐥𝐞 𝐚𝐧𝐝 𝐓𝐫𝐮𝐬𝐭𝐰𝐨𝐫𝐭𝐡𝐲 𝐅𝐨𝐮𝐧𝐝𝐚𝐭𝐢𝐨𝐧 𝐌𝐨𝐝𝐞𝐥𝐬 (𝐑𝐞𝐬𝐩𝐨𝐧𝐬𝐢𝐛𝐥𝐞𝐅𝐌) is approaching!🔥 📍 Hybrid (Hilton Mexico City Reforma +…

🚨 BREAKING: AI Can't Actually See Videos. New benchmark shows mainstream LVLMs barely hit 60% accuracy—while humans reach 94.82%. This isn’t a glitch—it’s a fundamental failure in video understanding. LVLMs are doing visual theater, not real comprehension.

🚨 #EMNLP2025 Oral: AI Can’t Actually “See” Videos - New Benchmark Exposes the Truth LVLMs aren’t thinking with video - they’re performing it. What looks like understanding is just visual theater. Introducing GLIMPSE - a benchmark revealing how today’s models fail when…

🚨 Excited to announce Gistify!, where a coding agent must extract the gist of a repository: generate a single, executable, and self-contained file that faithfully reproduces the behavior of a given command (e.g., a test or entrypoint). ✅ It is a lightweight, broadly applicable…

🚨 Check out our awesome students/postdocs' papers at #EMNLP2025 and say hi to them 👋! Also, I will give a keynote (virtually) on "Attributable, Conflict-Robust, and Multimodal Summarization with Multi-Source Retrieval" at the NewSumm workshop. -- Jaehong (in-person) finished…