🔥The 9th Round of Easy Loan, Earn $40 Reward is in progress❗️ ⏰ Promotion Period: January 15th - Feburary 15th, 2025 👉 Register now and check more details at gate.io/campaigns/358

Newish #EMNLP2021 work w/ Arturs Backurs & Karl Stratos: we try to generate text (in a data-to-text setting) by splicing together pieces of retrieved neighbor text. Paper: arxiv.org/pdf/2101.08248… 1/3

The DukeNLP group is hiring PhD students in all areas of natural language processing! Apply at gradschool.duke.edu/admissions/app… by Dec 15 to work with Sam Wiseman or Bhuwan Dhingra.

See a glimpse👇of how beautiful @unc +research triangle fall colors are 😍 Come join our awesome group of UNC NLP UNC Computer Science students+staff+faculty (& great neighbors eg. DukeNLP). We are hiring at all levels (phd, postdocs, faculty); feel free to ping any of us with questions 🙏

🤔 When does a factoid question need a *long* answer? 🤖 "Long" could mean multiple things: either you ask for a city with a very long name or … Read Ivan Stelmakh's internship paper to get the second part of the answer! arxiv.org/abs/2204.06092

Increasing the input length often increases accuracy on NLP tasks like summarization. But given limited time and a fixed number of GPUs, is it better to increase model size or input sequence length? Find the answer in our latest work: arxiv.org/abs/2204.07288 1/3

New Preprint from Yukun Huang! Can an LLM determine when its responses are incorrect? Our latest paper dives into "Calibrating long-form generations from an LLM". Discover more at arxiv.org/abs/2402.06544 (1/n)

150+ people registered for the SouthNLP 2024 at Emory University on 4/5. The schedule is available on our website: southnlp.github.io/southnlp2024/ Registration is open until March 10th. If you plan to attend, please register by completing the form here: forms.gle/NBWrgtgM5KgUq3…

🧐 Can we generate *LLM-proof* math problems❓ 👉Check out the new preprint from @ruoyuxyz , Chengxuan Huang and Junlin Wang : arxiv.org/abs/2402.17916 #LLMs #NLProc 🧵(1/6)

Welcome to Duke!

🦝Excited to announce our work on robustness & security of LLM systems! 𝐑𝐚𝐜𝐜𝐨𝐨𝐧: Prompt Extraction Benchmark of LLM-Integrated Applications Prompt extraction from LLM-integrated apps like GPT-s is a critical security concern. ‼️

🚨 Breaking: >90% AUC on the WikiMIA dataset for membership inference! Want to know if your data is in LLM's training set?🔍 Check out our latest work "ReCaLL: Membership Inference via Relative Conditional Log-Likelihoods" ✨ royxie.com/recall-project… 🧵1/6

🎉 Excited to share that IsoBench has been accepted at Conference on Language Modeling! IsoBench features isomorphic inputs across Math/Graph problems, Chess games, and Physics/Chemistry questions. Check out the dataset here: huggingface.co/datasets/isobe…

🧵When should LLMs trust external contexts in RAG? New paper from Yukun Huang and Sanxing Chen enhances LLMs’ *situated faithfulness* to external contexts -- even when they are wrong!👇

📢 New preprint on a benchmark for multimodal information extraction! Structured data extraction from long documents consisting of interconnected data in text, tables, and figures remains a challenge. MatViX aims to fill this gap. matvix-bench.github.io

**New paper from Roy Xie ** Do LLMs know when they have read enough to answer a question? We show how language models can STOP processing input text early without losing accuracy – Why waste 40,000 tokens when 500 suffice? 🧵 📄 Paper: arxiv.org/abs/2502.01025

Excited to share work from my Together AI internship—a deep dive into inference‑time scaling methods 🧠 We rigorously evaluated verifier‑free inference-time scaling methods across both reasoning and non‑reasoning LLMs. Some key findings: 🔑 Even with huge rollout budgets,

At #ICLR2025? Check out Yukun Huang ‘s poster tomorrow on when to trust external contexts using LLMs..

📢 New Preprint from Raghuveer @ NAACL25 on Multimodal Contrastive Learning: Breaking the Batch Barrier (B3) 📢 TL;DR: Smart batch mining based on community detection achieves state of the art on the MMEB benchmark. Preprint: arxiv.org/pdf/2505.11293 Code: github.com/raghavlite/B3

Glad to share a new ACL Findings paper from @MaxHolsman and Yukun Huang! We introduce Fuzzy Speculative Decoding (FSD) which extends speculative decoding to allow a tunable exchange of generation quality and inference acceleration. Paper: arxiv.org/abs/2502.20704

Can we train reasoning LLMs to generate answers as they think? Introducing 𝐈𝐧𝐭𝐞𝐫𝐥𝐞𝐚𝐯𝐞𝐝 𝐑𝐞𝐚𝐬𝐨𝐧𝐢𝐧𝐠! We train LLMs to alternate between thinking & answering 🚀 Reducing Time-to-First-Token (TTFT) by over 80% ⚡AND improving Pass@1 accuracy up to 19.3%!📈 🧵 1/n