Yiming Cui

We are extremely overwhelmed and honored to receive the IEEE Signal Processing Society Best Paper Award (2025) for our paper "Pre-Training with Whole Word Masking for Chinese BERT", published in IEEE/ACM TASLP (2021). 🎉🎉🎉 @IEEEsps #IEEE signalprocessingsociety.org/newsletter/202…

@anton_reshetov @OpenAI same here. really frustrating. i am on an annual subscription, which is even worse. 🥲

Hey @OpenAI I received the “Codex for Open Source” email offering 6 months of ChatGPT Pro, but when I try to activate it I get: “This promotion isn’t available for your plan.” I currently have an active ChatGPT Plus subscription.

Good News: I am enrolled in #Codex for Open Source. 🎉 Bad News: I cannot redeem this untill my Plus subscription ends. What even worse, I am on an annual subscription, ends 6 months+. 😭 @OpenAI please help.

AI vs. the Olympiad: Can Multimodal LLMs Truly 'See' Chemistry? communities.springernature.com/posts/ai-vs-th…

(5/x) Occlusion-based saliency analysis. By masking image regions and measuring confidence drops,baseline MLLMs often rely on spurious visual cues, while CoT shifts attention toward chemically meaningful structures, improving both grounding and confidence. nature.com/articles/s4200…

Excited to share our first paper in @CommsChem (Nature Portfolio) 🎉🎉🎉 We systematically evaluate multimodal large language models on chemistry Olympiad–level problems, revealing where current models succeed and where they still struggle. #AI4Chemistry #LLM #MultimodalAI #NLP

(4/x) When seeing hurts: visual input can degrade performance. Adding visual input sometimes reduces accuracy, especially in smaller models. Larger models tend to balance textual and visual signals better, which may be key to achieving strong performance. nature.com/articles/s4200…

(3/x) Task-type breakdown. Current multimodal LLMs perform well on tables and charts, but struggle with molecular structures and experimental apparatus, which require chemistry-specific visual understanding and domain knowledge. nature.com/articles/s4200…

(2/x) CoT generally improves chemical reasoning performance. Analysis show that CoT is especially helpful for mid-tier models. For e.g., GPT-4.1-mini achieves 20~26 accuracy improvement with CoT, while less significant for small/large-scale models. nature.com/articles/s4200…

(1/x) we curate a chemistry benchmark based on USNCO exams, spanning over two decades, consisting of 473 real multimodal QA problems. It covers a broad spectrum of chemistry topics, including general, physical, organic, inorganic, and analytical chemistry. nature.com/articles/s4200…

Our paper "Self-Evolving GPT: A Lifelong Autonomous Experiential Learner" is accepted at #ACL2024 main! We propose a framework for LLMs to autonomously learn and apply experience, boosting GPT-3.5 and GPT-4 performance. Stay tuned for the paper and code release! #NLP #LLM #GPT

Happy to introduce Chinese-LLaMA-Alpaca-3, which is our 3rd open-source projects on #Llama series. We release Llama-3-Chinese-8B and Llama-3-Chinese-8B-Instruct with continual PT/SFT on Chinese corpora. Check our project: github.com/ymcui/Chinese-… #nlproc #llama3

Through our empirical experiments on creating Chinese Mixtral, we find that extending vocabulary might NOT be a necessity for LLM language transfer. As usual, we open-source Chinese-Mixtral(-Instruct) at GitHub/HF: github.com/ymcui/Chinese-… arXiv Paper: arxiv.org/abs/2403.01851

We release Chinese-LLaMA-2-7B and Chinese-Alpaca-2-7B based on #Llama-2, which achieved significant improvements over our first-gen Chinese-LLaMA/Alpaca, even surpass 13B models on some metrics. Check our GitHub repo: github.com/ymcui/Chinese-… #llm #NLProc

@joemkwon Sorry for the late reply. Regarding your question, our main motivation is to add more trainable parameter (qkvo and mlp) within LoRA scheme. Recent research QLoRA also shows that adapting qkvo/mlp is essential to achieve a better performance. Maybe you can check the QLoRA paper.

@KCrosner I’m curious how you decided which parameter to freeze and wrap with LoRA, vs just freezing entirely or tuning entirely

Excited to release our Chinese 🦙#LLaMA and #Alpaca LLMs (7B for now), extended with an additional 20k Chinese vocabulary, trained with alpaca-lora. Our model works seamlessly with the wonderful llama.cpp on CPU. Give it a try at github.com/ymcui/Chinese-… #nlproc #llm #AI

Update 13B Chinese #LLaMA and #Alpaca. Better quality compared to 7B. GPT-4 rates 13B model 71/100 while 49 for 7B version. We also provide a Colab notebook for fast conversion, and of course it is fully compatible with llama.cpp. Try: github.com/ymcui/Chinese-… #nlproc #llm #ai

Happy to release our multimodal pre-trained model VLE, which achieved top performance on VCR. We also set up a pipeline with captioning model and LLM to generate much user-friendly answers for VQA. Resources, code, and demo are available through: github.com/iflytek/VLE 🎉🎉🎉

@cryptexcode @SemEvalWorkshop @naacl Thank you. The live session (mainly for task organizers) is hosted via Zoom, and all system papers are presented as posters (no oral). I'm not sure if the video will be made public by official. If you are interested in best paper list, it will be posted on SemEval website soon.

@KCrosner @SemEvalWorkshop @naacl Congratulations! Are the recordings available offline?

We are happy to announce that our SemEval-2022 system description paper is recognized as "best paper honorable mention award". 🎉🎉🎉 Paper and code: github.com/GeekDream-x/Se… @SemEvalWorkshop @NAACL #nlproc #naacl2022 #semeval