Saurabh Srivastava

#NLProc Ph.D. student at @GeorgeMasonU. Prev intern at @AmazonScience and @Deloitte. Formerly NLP Researcher @TCSResearch; exIIT'ian @IITPAT; NLP, ML, DL

[Saurabh Srivastava @salokr_deep]

🚨New Preprint🚨

(1/n) Do SOTA #LRMs like #DeepSeekR1 and #o1 still need Prompt Optimization?

We put them to the test on a structured task, Event Extraction, and did the first deep dive into prompt optimization.

We found: Yes, they do benefit from it.

#NLProc #LLMs #LRMs A🧵

[Saurabh Srivastava @salokr_deep]

@thetaytay @ZiyuYao Hi,
Thank you for appreciating our work 🙏 Yes, we plan to release the code soon. Hopefully, by the mid of next week. I’ll share the link soon.

[Saurabh Srivastava @salokr_deep]

🚨New Preprint🚨

(1/n) Do SOTA #LRMs like #DeepSeekR1 and #o1 still need Prompt Optimization?

We put them to the test on a structured task, Event Extraction, and did the first deep dive into prompt optimization.

We found: Yes, they do benefit from it.

#NLProc #LLMs #LRMs A🧵

[Saurabh Srivastava @salokr_deep]

(8/8) Shoutout to my advisor @ZiyuYao for the steady support and thoughtful encouragement throughout! 🙌

Looking forward to more insights, feedback, and future collabs on this wild EE + LLM journey! 🚀✨

📚 Preprint: https://t.co/iCeMI2zMKN

[Saurabh Srivastava @salokr_deep]

(7/n) 🔬Our error analysis shows that Span Overprediction and Coreference errors remain the most common across all models.

However, prompts generated by LRMs often include rules that
🔹 Reduce argument-related mistakes
🔹 encourage identification of multiple events per input

[Saurabh Srivastava @salokr_deep]

Please come and chat with us If you like our paper, at Convention Center A1 between 12:45 PM - 1:45 PM 🤝

[Saurabh Srivastava @salokr_deep]

🚨ACL 2024 Poster Alert🚨"Rewriting Prompts for Instances with LLMs in the Loop Yields Better Zero-Shot Performance" with @ZiyuYao on supervising LLMs to elicit correct and safe zero-shot responses.
📚 https://t.co/zXoWspmBYx
⏰ 12:45 PM - 1:45 PM Convention Center A1
#NLProc

[Saurabh Srivastava @salokr_deep]

Similar performance improvements were observed using cross-family LLMs such as Open-Source and weaker LLMs acting as supervisors for closed and stronger LLMs. These are the direct consequences of the recent "Evaluation is easier than generation" trend.

[Saurabh Srivastava @salokr_deep]

📢 Excited to present our paper "MailEx: Email Event and Argument Extraction" (id: 77) at #EMNLP2023! 📚 Join us at the poster session to dive into the world of extracting events and arguments from emails.🚀
Location and time: East Foyer 2PM
#NLProc #EMNLP2023 #MailEx

[Ziyu Yao @ZiyuYao]

I'm at #EMNLP23. Welcome ☕️ & chat!

1. MailEx dataset for conv. event extraction (w/ my stu @salokr_deep) https://t.co/bN0zKjkFrT. Dec 8, 2pm, Poster 2, EastFoyer

2. Gentopia @GentopiaAI open framework for augmented LLMs https://t.co/Z3gbTKIA0g. Dec 10, 9am, Demo 6, EastFoyer

[Saurabh Srivastava @salokr_deep]

@GeorgeMasonU @UofMaryland @ZiyuYao Kudos to our annotators! Their meticulous work and invaluable contributions were pivotal in constructing MailEx. Every label and every annotation brought this dataset to life. 🙏

Link to the paper: https://t.co/KvR28X2LKY

[Saurabh Srivastava @salokr_deep]

🎉 Excited to share our paper "MailEx" accepted at #EMNLP2023! 🚀 We introduce the first-ever dataset for event extraction from conversational email threads. 📧✨

arXiv: https://t.co/KvR28X2LKY
Code: https://t.co/79g0oeKc6Z
Dataset: https://t.co/rde0XKoF5M
#NLProc #LLMs #email

[Saurabh Srivastava @salokr_deep]

Thankful to our teams at C4I @GeorgeMasonU & ARLIS @UofMaryland for the collaboration. A massive shoutout to my advisor @ZiyuYao for constant support and encouragement. Looking forward to more insights, feedback, and potential collaboration in this venture!🙌 🌟