December 2023 - nlp_reading_group

[nlplab] Reading Discussion on Friday Dec 29
by Zhiyang Xu 27 Dec '23

27 Dec '23

Hi Everyone, In the reading discussion this week, we will discuss: Kosmos-G: Generating Images in Context with Multimodal Large Language Models <https://arxiv.org/abs/2310.02992> This paper develops a training framework to efficiently bootstrap pretraining from LLMs and diffusion models for image editing. It effectively tackles the challenge of interleaved multimodal inputs involving multiple images and text pieces. The results demonstrate the high quality of generated images and their fidelity towards source images. The reading discussion is at *4-5 pm on Dec 29 *this Friday. You may attend the reading discussion in Room 366 in the Data & Decision Building, or on Zoom *(https://virginiatech.zoom.us/j/8297378667 <https://virginiatech.zoom.us/j/8297378667>)*. Enjoy reading! Best wishes, Zhiyang Xu _______________________________________________ nlp_reading_group mailing list nlp_reading_group(a)cs.vt.edu https://mailman.cs.vt.edu/mailman/listinfo/nlp_reading_group

1 0

[nlplab] Reading Discussion on Friday Dec 22
by Zihao Lin 20 Dec '23

20 Dec '23

Hi Everyone, In the reading discussion this week, we will discuss two papers: 1. Label Words are Anchors: An Information Flow Perspective for Understanding In-Context Learning (*EMNLP'23 Best Paper*) 2. In-Context Learning Creates Task Vectors (*EMNLP'23*) These two papers mainly provide explanations of how LLM works in In-Context Learning. The reading discussion is at *4-5 pm on Dec 22* this Friday. You may attend the reading discussion in Room 336 in the Data & Decision Building, or on Zoom (*https://virginiatech.zoom.us/j/7048491561 <https://virginiatech.zoom.us/j/7048491561>*). Enjoy reading! Best wishes, Zihao Lin

1 0

[nlplab] Reading Discussion on Friday Dec 15
by Minqian Liu 12 Dec '23

12 Dec '23

Hi Everyone, In the reading discussion this week, we will discuss: Counterfactual Memorization in Neural Language Models <https://arxiv.org/abs/2112.12938> (*NeurIPS'23 Spotlight*) This paper thoroughly analyzed how documents in the training set influence the model's predictions. You may think about the following questions during reading: 1. What are the implications of the findings in this paper, especially in topics such as factual checking and alignment? 2. What are the limitations of this work? The reading discussion is at *4-5 pm on Dec 15 *this Friday. You may attend the reading discussion in Room 366 in the Data & Decision Building, or on Zoom *(https://virginiatech.zoom.us/j/8297378667 <https://virginiatech.zoom.us/j/8297378667>)*. Enjoy reading! Best wishes, Minqian Liu

1 0

[nlplab] Reading Discussion on Friday Dec 8
by Barry Yao 05 Dec '23

05 Dec '23

Hello everyone, In this week's reading discussion, we will talk about: Skill-it! A Data-Driven Skills Framework for Understanding and Training Language Models <https://openreview.net/forum?id=IoizwO1NLf> (*NeurIPS 2023 spotlight*) This paper proposes a novel data selection approach to train language models efficiently. I highly recommend you read the paper in advance and think about: *1. Could we extend this approach to train a large multimodal model?* *2. The current data selection approach is in a static manner, a.k.a., the skill associated with each data point is predefined. Could we extend the approach to a dynamic manner? For example, could we leverage an LMM to generate feedback or analyze student model errors to select data in a dynamic manner?* You could attend the meeting in the Data & Decision Building, Room 366, or on Zoom ( https://virginiatech.zoom.us/j/4506040955?pwd=TGFtTUVWR2tuYmVlS3JraTQzbjgrU…) at 4-5 pm on Friday. Happy reading! Best regards, Barry Menglong Yao Ph.D. student in Computer Science Virginia Tech Homepage <https://barry-yao.netlify.app/>

1 0