Feeling hard generating challenging evaluation data for LLMs? Check our work👇! Introducing LINK🔗, the first framework for systematically generating data in the long-tail distribution, guided by symbolic rules arxiv.org/abs/2311.07237 w/USC NLP MOSAIC 🧵⬇️ #NLProc [1/n]

We're excited to attend #SocalNLP today! ICYMI, sunny southern California is a fantastic place to do #NLProc, come check out what USC NLP [nlp.usc.edu] has been working on lately! And did we say we're hiring PhD students this fall? 🌴🏖️☀️

Throwback to when Sean Ren 🔆 and our lab made our wishlist and dream research directions to discuss in our lab meeting — very helpful in contextualising our work in the age of LLMs!! 🙌🏼 USC NLP is such a great place to do research 🫶

How good are LMs at inductive reasoning? How are their behaviors similar to/contrasted with those of humans? We study these via iterative hypothesis refinement. We observe that LMs are phenomenal hypothesis proposers, but they also behave as puzzling inductive reasoners: (1/n)

Arrived at NOLA for #NeurIPS2023🔥 Exciting time to chat about limits/science of LLMs, “slow” reasoning & explainability. Join our posters for a fun disucssion🍻 Ads: USC CS is hiring tenured track AI faculty + USC NLP is looking for strong PhD students. Talk to us!

To detect if your data was used for LLM pretraining, consider using data watermarks: arxiv.org/pdf/2402.10892… Detection can be framed as hypothesis testing (statistical guarantees!), if you contributed multiple training documents and watermarked them before public release. 🧵

Absolutely thrilled to receive this honor. Rarely for a researcher could have their first PhD publication win a Test of Time Award (for 10 years of its cumulative impact). I’m super grateful for the chance to collaborate with Xiao on this fun project — turns out to be a

Wanna know gpt-3.5-turbo's embed size? We find a way to extract info from LLM APIs and estimate gpt-3.5-turbo’s embed size to be 4096. With the same trick we also develop 25x faster logprob extraction, audits for LLM APIs, and more! 📄 arxiv.org/abs/2403.09539 Here’s how 1/🧵

New paper 🚨 Looking for a strong, open-sourced entailment-verification model to verify your model generations for consistency? ✅ You can now use the 🤗model huggingface.co/soumyasanyal/n… for this! Our FlanT5-xxl finetuned model can predict entailment errors better than GPT3.5 and

🧐LMs forget upstream knowledge when continuously fine-tuned. When fine-tuned on new data, can we forecast what upstream examples will be forgotten? 🥳Excited to share our #ICML Spotlight paper on forecasting example forgetting! 🔗Project page: inklab.usc.edu/lm-forgetting-…

Congratulations to the GDM Google DeepMind team on their best paper award at #ICML2024 & Appreciate @afedercooper's shout out to our concurrent paper 🙌 If you are into the topic of recovering model info through just its output logits, check out our paper led by Matthew Finlayson too!

Introducing 𝗟𝗶𝗳𝗲𝗹𝗼𝗻𝗴 𝗜𝗖𝗟 and 𝗧𝗮𝘀𝗸 𝗛𝗮𝘆𝘀𝘁𝗮𝗰𝗸, a new approach for evaluating long-context LMs, featuring ever-changing task streams that controllably fill the context window, and NIAH-style visualization for easy diagnosis. 📜 arxiv.org/abs/2407.16695 🧵

Excited to see everyone soon at #acl2024 in Bangkok! I'll be presenting our work, Relying on the Unreliable: The Impact of Language Models' Reluctance to Express Uncertainty arxiv.org/abs/2401.06730 Poster session 3 on Aug 12 at 16:00! W/ Maarten Sap (he/him) Jena Hwang Sean Ren

Arriving in Bangkok for ACL 2025! 😃 Will be sharing our recent work on logical scaffolding, model uncertainty expression & multi-hop entailment inference w/ folks USC NLP + Kaitlyn Zhou ✈️ CSCW, EMNLP! +friends Ai2 I'm also helping on the <AI / ALL> summit w/ Sahara AI 🔆 👇👇

Find us at the posters! Can LLMs Reason with Rules? Logic Scaffolding for Stress-Testing and Improving LLMs w/ Siyuan Wang Yejin Choi et al Relying on the Unreliable: The Impact of Language Models' Reluctance to Express Uncertainty w/ Kaitlyn Zhou ✈️ CSCW, EMNLP!, Maarten Sap (he/him) et al.

Join us at the co-located <AI / ALL> summit on Aug 15, with the social party in the evening! lu.ma/mxcx5bia co-hosted with SCB 10X SambaNova Systems sponsored by Amazon Web Services participated by folks AI at Meta @google Cohere For AI Together AI

Proud moment seeing our CEO & Co-Founder Sean Ren 🔆 alongside his USC NLP students at ACL 2025. Supporting the next generation of thought leaders in AI is exactly what drives us forward.

Heading to #EMNLP2024, down to chat! Excited to present our work (Wed 10:30am) on systematic data generation in long-tail (low confidence) distribution for more challenging evaluation. 🧵👇 📰: arxiv.org/abs/2311.07237 💻: github.com/INK-USC/LINK 🔖: zenodo.org/records/101179…

Proud of my student Huihan Li and intern Arnav presenting their #ICLR2025 work on attributing culture-conditioned generation to LLM’s training corpora. Fun time meeting many friends. Ping me if you want to chat about model security, interpretability and human-LM interaction!

Thrilled for the Best Paper Award runner-up at #NAACL2025! 🥳 Even when answers are incorrect, people may rely more on LLMs if they use warm and emphatic expressions! We analyze the risks of human over-reliance on LLM expressions of uncertainty: arxiv.org/pdf/2407.07950 w/