Announcing GPT-4, a large multimodal model, with our best-ever results on capabilities and alignment: openai.com/product/gpt-4

📢 Our #ACL2023 paper "Fact-Checking Complex Claims with Program-Guided Reasoning" is now publically available. A framework for fact-checking complex claims with the guide of auto-generated reasoning programs. 📜Paper: arxiv.org/abs/2305.12744 💻Codes: github.com/teacherpeterpa…

📢 Exciting! 📚 Introducing our preprint on memory-efficient fine-tuning! 🧠💡 Make your pre-trained model reversible without any additional pre-training!🔍 Work with Christof Monz and Shaomu Tan! preprint: arxiv.org/pdf/2306.00477… repos: github.com/baohaoLiao/mef…

Checking for our paper! 🥳

Excited to share our #AACL paper: "FOLLOWUPQG: Towards Information-Seeking Follow-up Question Generation" ✨ Thanks for collaborators Liangming Pan , Yixin Cao, and Min-Yen Kan! 📖BLOG: yanmeng-nlp.com/261-2/ 📑PAPER: arxiv.org/pdf/2309.05007… #NLProc

Enjoyed visiting UC Berkeley’s Machine Learning Club yesterday, where I gave a talk on doing AI research. Slides: docs.google.com/presentation/d… In the past few years I’ve worked with and observed some extremely talented researchers, and these are the trends I’ve noticed: 1. When

Glad to share that our paper is accepted by the main conference in #EACL2024🎉 Thanks for my supervisor Christof Monz! Check out our paper and blog for more details on the interplay between transfer and regularization in MMT. Blog Link: yanmengnlp.com/target-side-tr… arxiv.org/abs/2402.01772

The Chat Shared Task (WMT2024) is live! 💥💥 Happy to announce this year’s Chat Shared Task which aims to translate a corpus composed of genuine bilingual conversations from the customer support domain!

Tomorrow I will present our paper for the oral session at 9:00 AM in Marie Louise 1 eaclmeeting #EACL2024 Happy to meet you there :)

Circuits are a hot topic in interpretability, but how do you find a circuit and guarantee it reflects how your model works? We (Sandro Pezzelle, Yonatan Belinkov, and I) introduce a new circuit-finding method, EAP-IG, and show it finds more faithful circuits arxiv.org/abs/2403.17806 1/8

Next week I'll be in Vienna at ICML Conference! Want to learn more on how to explicitly model embeddings on hypersphere and encourage dispersion during training? Come to the Gram Workshop poster session 2 on 27.07 Shoutout to my collaborators Hua Chang Bakker and timorous bestie 😷 💫

1/4 #ACL2024 Excited to share our new paper on the impact of fine-tuning on the qualitative advantages of LLMs in machine translation! 🤖 Our work highlights the importance of preserving LLM capabilities during fine-tuning. arxiv.org/abs/2405.20089

LTL will present two papers at ACL 2024: 1. The Fine-Tuning Paradox: Boosting Translation Quality Without Sacrificing LLM Abilities by David David Stap 2. How Far Can 100 Samples Go? Unlocking Overall Zero-Shot Multilingual Translation via Tiny Multi-Parallel Data by Di Di Wu

LLMs often "hallucinate". But not all hallucinations are the same! This paper reveals two distinct types: (1) due to lack of knowledge and (2) hallucination despite knowing. Check out our new preprint, "Distinguishing Ignorance from Error in LLM Hallucinations"

Memorization is NOT merely detrimental for reasoning tasks - sometimes, it’s surprisingly helpful. I’m really enjoying this project, as we work toward a more rigorous definition and understanding of reasoning and memorization (albeit in a controlled synthetic setting):

I am happy to share that our paper "How to Learn in a Noisy World? Self-correcting the Real-World Data Noise in Machine Translation" was accepted by NAACL 2025. Paper Link: arxiv.org/pdf/2407.02208 #NLP #NAACL

Key Finding 1: We simulate the primary source of noise in the parallel corpus, i.e., semantic misalignment, and show the limited effectiveness of widely-used sentence-level pre-filters for detecting it. This underscores the necessity of handling data noise in a fine-grained way.

Key Finding 2: With an observation of the increasing reliability of the model's self-knowledge for distinguishing misaligned and clean data at the token level, we propose self-correction to leverage the model's self-knowledge to correct the training supervision.

Tomorrow I will present our paper about data quality for MT at 9:00 AM in Hall 3 at #NAACL2025. Happy to meet you there :)