Weijia Xu
@weijiavxu
Followers
623
Following
318
Media
28
Statuses
105
Researcher at @MSFTResearch. Natural Language Processing and Machine Learning. Formerly a PhD student @UMDCS advised by @MarineCarpuat. 🏳️🌈 she/her
Joined August 2017
Our PNAS paper is finally OUT🚨 🤖LLMs have become popular tools for creative writing and ideation. But can they truly spark creativity? 💡 We introduce Sui Generis score to quantify the uniqueness of high-level elements (e.g. plot) in AI storytelling ✍️
1
2
10
This work is done in collaboration with Nebojsa Jojic, @raosudha89, @chris_brockett, and Bill Dolan at @MSFTResearch
0
0
1
The metric also reveals other characteristics of AI-generated stories, e.g. overly fast pacing without fully resolving a surprising plot. Check our paper for more interesting findings:
pnas.org
With rapid advances in large language models (LLMs), there has been an increasing application of LLMs in creative content ideation and generation. ...
1
0
1
The Sui Generis score also aligns moderately with human perceptions of surprise. This suggests that the score can be a proxy measure of surprise or interestingness, which may find uses in both model improvements and collaborative writing tools.
1
0
1
Evaluating on a hundred stories, we find that model-generated stories often contain “echoes” of plot elements that repeat across generations and even across LLMs, while plots from the original human-written stories are rarely echoed.
1
0
1
The Echoes in AI paper showed quite the opposite with also a story continuation setup. Additionally, we present evidence that both *syntactic* and *discourse* diversity measures show strong homogenization that lexical and cosine used in this paper do not capture.
People assume that AI homogenizes creative writing, producing much less diverse work than groups of humans This paper finds this isn’t true: given stories to complete, GPT-4o writes as diversely as humans (stylistic, lexical, & semantic) when prompted with context & randomness
2
14
38
Jiatao has been such an inspiring and insightful mentor to me! I learned a lot from him during my internship. Highly recommended to potential PhD students!
Life update: Excited to share that I will be joining @CIS_Penn @PennEngineers as an Assistant Professor in Fall 2025!🤯 I’m also seeking multiple PhD students passionate about Generative Intelligence and leveraging it to empower AI agents to interact with the Physical World🌟
1
0
1
We further evaluated our algorithm on 15 additional tasks in the latest version. Reprompting improves over manual CoT by +9.4 points on average over all 20 tasks. Come to our poster to discuss more!
1
0
1
The paper will be presented at ICML TODAY at Hall C 4-9 #606!
New #NLP paper! #GPT #ChatGPT Reprompting: Automated Chain-of-Thought Prompt Inference Through Gibbs Sampling Sparks of slow thinking: Do LLMs really need human-written CoT for slow thinking and multi-step reasoning? We show how to infer CoT recipes without human intervention ⬇️
1
0
2
I’m thrilled that this Human-Centered MT paper was recognized with an outstanding paper award at #EMNLP2023. Congratulations to lead authors Nikita Mehandru (@ucberkeley iSchool) and @swetaagrawal20 (@umdclip @istecnico) for making this interdisciplinary collaboration a success!
2/8 "Physician Detection of Clinical Harm in Machine Translation: Quality Estimation Aids in Reliance and Backtranslation Identifies Critical Errors" with @nikita_mehandru @swetaagrawal20 @elainekhoong Niloufar Salehi among others https://t.co/zAkgavVJhA
https://t.co/kwYqSkyTqU
4
20
114
I'm at #EMNLP2023 with many collaborators from @umdclip and beyond to share recent work in human-centered NLP. Some themes: reliance and trust, explainability, building common ground, measuring harms.
1
11
54
What does the future of machine translation research look like in the age of LLMs? At AMTA today, Ge Gao (@INFOCollegeUMD), Sharon O’Brien (@DCU_Research), Michel Simard (@NRC_CNRC), and I argued for broadening our view of MT by centering people throughout the MT R&D lifecycle.
1
14
40
Happening now!
With LLMs gaining popularity, it is important to discuss how underrepresented researchers are affected by them. We can't wait to see you all at our panel at #ACL2023NLP with @sarahookr, @davlanade and @ovalle_elia! #LLMs #NLProc #WiNLP (1/n)
0
0
0
Come join us on July 11! And share your opinions and questions with us on Twitter! We will collect them before the panel.
With LLMs gaining popularity, it is important to discuss how underrepresented researchers are affected by them. We can't wait to see you all at our panel at #ACL2023NLP with @sarahookr, @davlanade and @ovalle_elia! #LLMs #NLProc #WiNLP (1/n)
0
0
2
With LLMs gaining popularity, it is important to discuss how underrepresented researchers are affected by them. We can't wait to see you all at our panel at #ACL2023NLP with @sarahookr, @davlanade and @ovalle_elia! #LLMs #NLProc #WiNLP (1/n)
1
16
58
With a call for papers, there comes a call for reviewers! We are looking for reviewers for the WiNLP 2023 workshop at EMNLP 2023. We appreciate your willingness to participate and help! Please fill in the form in the attached tweet :)
3
4
3
The @WiNLPWorkshop 2023 call for papers is officially out! To be co-located with @emnlpmeeting, we invite researchers to submit a 2-pages abstract to be considered for a poster presentation. The early visa-friendly deadline is July 1 (AOE).
1
24
29
LLMs exhibit translation capabilities despite having never seen intentionally-included translation examples, so... where do those capabilities come from? 🚨We show that incidental bilingualism connects to the machine translation capabilities of PaLM. 📜 https://t.co/6lhv6ccSyH
13
94
512
Also we show that LLM comparisons can be highly sensitive to the choice of CoT, further emphasizing the need for automatic prompt discovery and optimization using algorithms like Reprompting
0
0
2