Yufei Tian

post-training @openai | prev: PhD @UCLA | NLP, creativity, unconventional reasoning | undergrad @Tsinghua_Uni

[Yufei Tian @yufei_t]

@mohsen_fayyaz @uclanlp @NeurIPSConf What a niceee group photo! Our group is growing!

[Yufei Tian @yufei_t]

@sanerc110 It’s open now!

[Yufei Tian @yufei_t]

👋👋Like everyone, I’m heading to SD for #NeurIPS! First ML conference since wrapping up my PhD and joining OpenAI.

These days I think about:

- post-training for personal & private context
- long-horizon, proactive, personalized agents
- memory, skills, cognition & context engineering
- maintaining diversity & creativity for LLMs

Please please overwhelm me with interesting papers! If any of the above resonates (or you’re curious about life/work at OpenAI), come say hi! ☕️🌮 Would love to meet old and new friends.

[Alex Spangher @AlexanderSpangh]

✨✨✨Hello everyone, I’m on the faculty job market this year.✨✨✨ I’m completing my PhD at USC, where I study agentic planning in creative contexts.

But before I get deeper into my research, I really want to tell you a little bit about myself :)

[Nanyun (Violet) Peng @ ACL26 @VioletNPeng]

Thrilled, grateful, and humbled to have won 3 outstanding paper awards at #EMNLP2024!!! Not even in my wildest dreams. Immense thanks to my amazing students and collaborators!
All three works are on evaluating LLM’ abilities in creative narrative generation. 🧵👇

[Yufei Tian @yufei_t]

Happening today 📆 10:45-11:00 at Flagler! #EMNLP Drop by if you want to hear about more about LLM storytelling!

[Yufei Tian @yufei_t]

🤔Could LLMs one day win the Nobel Prize 🏆in literature?
🚀 Thrilled to share our latest paper at #EMNLP2024: "Are Large Language Models Capable of Generating Human-Level Narratives?" 📚✨

If you have ever asked LLMs for an interesting story and felt something was off, you’re not alone! We compare LLM storytelling to human expert storytelling by analyzing narrative development and plot progression. Our computational framework focuses on three key aspects of discourse:
1. Story Arcs 📈📉
2. Turning Points 🔄
3. Affective Dimensions (Arousal 😲 & Valence 😊)

TLDR: We find important discrepancies in narrative development. While human-written stories are more suspenseful, arousing, and diverse, LLM-generated stories tend to be homogeneously positive and lacking in tension.

Joint work with @TenghaoHuang45 @AlexanderSpangh @muhao_chen @jonathanmay @VioletNPeng at UCLA, USC, and UCD.

[Yufei Tian @yufei_t]

#EMNLP Zero-shot AI content detectors based on token distributions struggle with real-world scenarios like high decoding T, complex prompts, and adversarial attacks.

Tmr 2pm: I will present a more robust method that incorporates discourse elements, e.g., event transitions.

[Yufei Tian @yufei_t]

What's Next:
We hope this work sparks further research into discourse studies, including but not limited to better narrative generation, comprehension, and evaluation. Our code and expert-annotated dataset are available here: https://t.co/4ORMGPSGwf

We look forward to your thoughts and feedback!
#NLP #AI #Storytelling #MachineLearning #EMNLP2024

[Yufei Tian @yufei_t]

🤔Could LLMs one day win the Nobel Prize 🏆in literature?
🚀 Thrilled to share our latest paper at #EMNLP2024: "Are Large Language Models Capable of Generating Human-Level Narratives?" 📚✨

If you have ever asked LLMs for an interesting story and felt something was off, you’re not alone! We compare LLM storytelling to human expert storytelling by analyzing narrative development and plot progression. Our computational framework focuses on three key aspects of discourse:
1. Story Arcs 📈📉
2. Turning Points 🔄
3. Affective Dimensions (Arousal 😲 & Valence 😊)

TLDR: We find important discrepancies in narrative development. While human-written stories are more suspenseful, arousing, and diverse, LLM-generated stories tend to be homogeneously positive and lacking in tension.

Joint work with @TenghaoHuang45 @AlexanderSpangh @muhao_chen @jonathanmay @VioletNPeng at UCLA, USC, and UCD.

[Yufei Tian @yufei_t]

Why does this matter?
Narratives shape our world, and as AI becomes more integrated into daily life, the way they tell stories will impact not only how we interact with machines but also how we communicate as a society. By understanding, debiasing, and improving LLMs' ability to craft engaging and diverse narratives, we can enhance human-to-human interactions, making them more meaningful and fostering a more connected society.

[Honghua Zhang @HonghuaZhang2]

Proposing Ctrl-G, a neurosymbolic framework that enables arbitrary LLMs to follow logical constraints (length control, infilling …) with 100% guarantees.

Ctrl-G beats GPT4 on the task of text editing by >30% higher satisfaction rate in human eval.
https://t.co/U6oz3bc935

[Yufei Tian @yufei_t]

Heading to Mexico city and excited to meet new and old friends @NAACL!! 🇲🇽🇲🇽 Let's grab ☕️🌮😇

Also I will be presenting MacGyver on Monday afternoon, Poster Session #2. Please drop by if you are interested in reasoning and/or creativity!

[Yufei Tian @yufei_t]

💡Can LLMs like GPT-4 reason creatively? Excited to share our latest research on AI and creativity! 🚀 Introducing MacGyver: a new playground for everyday innovation and physical reasoning --we collect problems to trigger unconventional usage of objects and innovative solutions.

[Yufei Tian @yufei_t]

@faeze_brh @lasha_nlp @Lianhuiq @Ronan_LeBras @RajaMarjieh @VioletNPeng @YejinChoinka @cocosci_lab Our work is accepted to NAACL'24. See you in Mexico City! 🇲🇽🇲🇽🇲🇽

[Yufei Tian @yufei_t]

💡Can LLMs like GPT-4 reason creatively? Excited to share our latest research on AI and creativity! 🚀 Introducing MacGyver: a new playground for everyday innovation and physical reasoning --we collect problems to trigger unconventional usage of objects and innovative solutions.

[Yufei Tian @yufei_t]

7/N
Finally, huge thanks to all my amazing mentors and collaborators: @faeze_brh @lasha_nlp @Lianhuiq @Ronan_LeBras @RajaMarjieh @VioletNPeng @YejinChoinka and Tom Griffiths (@cocosci_lab)!!