Jitkapat Sawatphol

A master's student specializing in NLP at @VISTEC_Thailand. Focuses on robustness, authorship analysis, and sentence representation learning.

[Alexander Panfilov @kotekjedi_ml]

We can finally talk about it:

We found a way to extract hidden reasoning of frontier models using a vulnerability in the APIs of every frontier AI company.

We verified that our reasoning token count matches billed API thinking tokens 1:1 for most of the prompts we queried.

[UKP Lab @UKPLab]

A picture of our colleague @KexinWang2049 explaining his work on DAPR – Document-Aware Passage Retrieval 🕵️‍♂️ at #ACL2024NLP, a collaboration with @Nils_Reimers (@Cohere) and @Iryna_Gurevych (@UKPLab) at Poster Session B.

https://t.co/FGn7djSGWC

[Yangsibo Huang @YangsiboHuang]

Check out our codebase if you're working on LM unlearning - it'll save you time! https://t.co/MUDxsye4Ja

We support 8 mainstream unlearning methods, including the OG gradient ascent, and new ones like NPO (@RuiqiZhang0614 Licong Lin, @yubai01, @Song__Mei), WHP (@EldanRonen @markrussinovich), & Task Vectors

[Niklas Muennighoff @Muennighoff]

Information Retrieval (IR) is entering a new era where docs are retrieved based on intensive reasoning:
Lvl 1: Keywords - BM25..
Lvl 2: Semantics - SBERT..
Lvl 3: 🧠 Reasoning - ❓
We release BRIGHT ✨to test reasoning-intensive retrieval; Lvl 1/2 methods underperform.
🧶 1/2

[Zhijing Jin @ZhijingJin]

Check out our #ACL2024NLP paper on Mechanistic Interpretability➡️"Competition of Mechanisms", where we interpret how #LLMs handle tradeoffs between multiple mechanisms (e.g., a factual memory vs. a counterfactual statement in context). Led by our awesome student @francescortu!

[Sahar Abdelnabi 🕊 @sahar_abdelnabi]

Very excited about this work!!

LLMs in applications process inputs from many sources, making them vulnerable to prompt injections.

We look into models' internals (activations) to catch if models drifted from users' instructions after processing supposedly data-only sources. 1/

[Anna Rogers @annargrs]

🥳 our position paper with @SashaMTL got into ICML! arXiv abs/2308.07120 now hosts camera-ready hot takes on:
- how we even define "LLM"
- 'LLMs are robust'
- 'LLMs are SOTA'
- 'Scale is all you need'
- 'LLMs are general-purpose technologies'
- 'LLMs have emergent properties'
/1

[Sasha Rush @srush_nlp]

I think grad students should work on Language Models! They're wildly important, super weird, and we understand next to nothing about them. I sometimes find it hard to think about anything else.

(That being said whenever I doubt Yann, I'm almost always wrong)

[Niloofar @niloofar_mire]

Submit your GenAI/Law/Policy work to our GenLaw@ICML 2024 workshop!!

Ddl is *June 10th* and submissions format is 1-2 page abstracts: https://t.co/G9yK7IQuAs

Accepted papers and spotlights from our 2023 event: https://t.co/G9yK7IQuAs

[Zhiyu Zoey Chen @ZhiyuChen4]

📈🏥⚖️How to build LLM applications in vital sectors such as finance, healthcare, and law?

📜Arxiv: https://t.co/2eRMT1d1pm

Sharing our new survey paper: A Survey on Large Language Models for Critical Societal Domains: Finance, Healthcare, and Law. We explore the LLM research landscape, applications, transformative impact, and future prospects in these vital sectors.
#AIResearch #LLMs #Finance #Healthcare #Law

[David Stutz @davidstutz92]

A topic I regularly talk about with doctoral students is how to organize a PhD. It can be overwhelming to keep track of all the papers, research ideas, reviews, different collaborations, and so on. So I took some notes on how I organized my PhD: https://t.co/QpZzHKnJxd

[Niloofar @niloofar_mire]

In🇦🇹this week @iclr_conf to present:

1⃣Privacy-preserving in-context learning with DP few-shot generation https://t.co/2JawiGp8bn
🖼️Tue 10:45 B-229

2⃣Can LLMs Keep a Secret? Testing Privacy Implications via Contextual Integrity https://t.co/CY7ATF6VS2
🖼️ Wed 16:30 B-215

[Workshop on Human-Centered Large Language Modeling @HuCLLM]

🚨Exciting news! #NLProc

📢The call for papers is now open for the Workshop on Human-Centered Large Language Modeling (#HuCLLM2024) at #ACL2024NLP.

See CFP: https://t.co/HdgqZUvHvl
Direct Submission Deadline: May 10, 2024
ARR-reviewed papers Commitment Deadline: May 17, 2024

[Zhijing Jin @ZhijingJin]

Really enjoyed the twice-per-year webinar by @InstMathStat @eth_math @YoungStatS2 w/ Nicolai Meinshausen @mo_lotfollahi Max Simchowitz (and thx to all organizers!)
🎙️My talk “A Paradigm Shift in Addressing Distribution Shifts: Insights from LLMs”
📜Slides: https://t.co/p9zXIy3tac

[Shi Feng @ihsgnef]

Many AI safety methods rely on AIs to provide oversight and supervision. The fact that models are interacting with one another alone should be a safety concern. We provide some initial evidence to support this intuition. More to follow soon!

[Chenhao Tan @ChenhaoTan]

I am quite excited about this work! Ideation and hypothesis generation are challenging tasks for scientists. It is quite plausible that LLMs will contribute to the next breakthrough in science.

[Niloofar @niloofar_mire]

I will be talking about what differential privacy is, what it is not and what some common misconceptions are in privacy for generative AI in a couple hours @genlawcenter in DC!

Join us on the live stream: https://t.co/BcbmNReXtB
Slides: https://t.co/xN3dx5uV05

[Niloofar @niloofar_mire]

Can we uncover memorization of pre-training data in LLMs, using other LLMs?

Our iterative prompt optimization method finds prompts that propel an LM to output training data using other LMs. We show higher avg. data reconstruction & extract 1.4X more PII!
https://t.co/KGKuZtjk3f

[Niloofar @niloofar_mire]

I'll be @uclanlp tmw & @ucsbNLP the next day to talk about how the "emergent" capabilities of LLMs create emergent inference-time privacy risks, and how membership inference attacks can be inconclusive in current setups! Hit me up if you wanna chat!

https://t.co/h5Il2eOKE6

[Alice Oh @aliceoh]

Let's be honest. For everyone in ML/NLP, it's a really exciting time but also very stressful with so many new papers, models, benchmarks, deadlines, reviews, talks, workshops, conferences...

How do you keep up and stay sane?

For me, the only solution is collaboration. 1/n