@thomaspalmeira
Working with #LLMs #NLP #Speech for Multilingual, Low-Resource, and Efficient ML PhD Track at @Polytechnique and @TelecomParis. Science saves lifes! #EMNLP2024 #LoG2024 #NeurIPS2024 http://thomas-ferraz.github.io https://linktr.ee/_thomaspf
I’M BACK TO THE ALPS! 🏔️⛷️🇫🇷
I am excited to share that I have started a position as a Research Scientist at @naverlabseurope.bsky.social . I'll be working on exciting topics around Memory and Reasoning in LLMs and AI Agents.
Was I too lenient, or did I just get lucky with the papers I was assigned to review for #ICLR2025 ?
The result is that you often need to double or triple-check their responses to ensure accuracy because it’s easy to be convinced by an answer that sounds good at first glance but, in detail, makes no sense at all…
However, the way this feedback is gathered and knowledge limitations of LLMs, has often led these models to learn how to seem useful and relevant, even when they are not.
A report about judges from Portuguese Courts using LLMs for rulings that cited laws or precedents that don’t exist…
An interesting fact about LLMs is that they are trained using RLHF, a technique to make model learn to generate more relevant and useful responses.
www.publico.pt/2024/11/24/s...
🚀 Don’t miss our #LoG2024 poster on Tue, Nov 28 @ 15:00 GMT!
Join us to discuss GNNs + LLMs for disinformation detection. The conference registration is free!
#FakeNewsDetection #GNNs #AI #LLMs @logconference.bsky.social
🙌 Collaborators:
This research is part of @BrunoCroso’s undergrad internship at POLI-USP at @roseli_deus lab.
📄 Paper: arxiv.org/pdf/2410.19193
📂 Code: github.com/BrunoCroso/C...
📜 More details: linktr.ee/_thomaspf
📊 Key findings:
1️⃣ Contextual text representations (BERTweet) improve Macro F1 by 33.8% over GNNs without text.
2️⃣ Retweets provide richer signals than user bios, but combined reach the best result.
3️⃣ Noise injection augmentation (commonly used in LLMs) reduces training stability.
📰 Why?
We know disinformation campaigns erode public trust on information.
On the technical side, few works leverage Transformers/LLMs advances to encode information in GNNs. Our work explore using contextual representations (from user bios & replies) to enhance fake news detection.
Happy to share our paper "Enriching GNNs with Text Contextual Representations for Detecting Disinformation Campaigns on Social Media" being presented tomorrow at #LoG2024!
🗓 Tue, Nov 28
⏰ 15h GMT
Virtual
Paper: arxiv.org/pdf/2410.19193
Join us to discuss combating disinformation with GNNs! 🧵 (1/5)
Response from HF ethicist about the HF Bluesky scrapped dataset everyone is freaking about
Photo with some of my amazing co-authors!
From left to right: Haw-Shiuan, Thomas, Violet and Sijia standing in front of the poster.
Thomas standing in front of his poster in the conference room
Excited to kick off my Bluesky sharing that I just presented our LLM Self-Correction paper at #EMNLP2024! 🎉 We propose a benchmark and a solution for LLMs on multi-constrained instruction following.
Check it: bit.ly/DeCRIM
Super exciting discussion there and lots of new ideas coming out from it!