Letiția Pârcălăbescu (@letiepi) header
Letiția Pârcălăbescu (@letiepi) profile picture

👩‍🎓 PhD, 🧠 AI Researcher, 📺 ML videos creator on YouTube: https://nitter.cf/t.co/yd92HLx342 ▶ @AICoffeeBreak

Joined August 2020
Photo by Letiția Pârcălăbescu (@letiepi)Photo by Letiția Pârcălăbescu (@letiepi)Photo by Letiția Pârcălăbescu (@letiepi)Photo by Letiția Pârcălăbescu (@letiepi)Photo by Letiția Pârcălăbescu (@letiepi)Photo by Letiția Pârcălăbescu (@letiepi)Photo by Letiția Pârcălăbescu (@letiepi)Photo by Letiția Pârcălăbescu (@letiepi)Photo by Letiția Pârcălăbescu (@letiepi)Photo by Letiția Pârcălăbescu (@letiepi)
Replying to @letiepi
@letiepi has a new Twitter handle: @AICoffeeBreak. Keeping the professional profile separated from the private one. Once again, I would like to express my gratitude for being on this amazing list!
My favorite ML educators: [1/n] - @JayAlammar: visual guides on NLP - @fishnets88: intuition & real-world anecdotes - @CShorten30, @ecsquendor, @ykilcher : netflix of research papers - @letiepi: bite-size animated videos on ML - @nlpguy_: experience on NLP in industry
2
16
Proud to have helped build Kolibri 🐦 We brought our Merlin-Arthur work on grounding and hallucination reduction into the model. Iteration one: scale brings its challenges, but the training pipeline lets us learn and iterate fast.
Small bird, fast wings, Kolibri is here. 78B parameters. 3.46B active. Up to 1M tokens of context. Built in Europe. Now the weights are yours. Run it on your own hardware, under Apache 2.0.
3
1
68
5,300
It was really great talking to many people at #ICLR in Singapore about my last paper written during my PhD! 📜
5
8
184
10,219
Quick thread with paper summary: nitter.cf/letiepi/status/1882148…:
The last paper of my PhD @HD_NLP is accepted at ICLR 2025!🙌 We investigate the reliance of modern Vision & Language Models (VLMs) on image🖼️ vs. text📄 inputs when generating answers vs. explanations, revealing fascinating insights into their modality use and self-consistency.👇
1
3
623
The last paper of my PhD @HD_NLP is accepted at ICLR 2025!🙌 We investigate the reliance of modern Vision & Language Models (VLMs) on image🖼️ vs. text📄 inputs when generating answers vs. explanations, revealing fascinating insights into their modality use and self-consistency.👇
4
1
90
4,747
The last paper of my PhD @HD_NLP is accepted at ICLR 2025!🙌 We investigate the reliance of modern Vision & Language Models (VLMs) on image🖼️ vs. text📄 inputs when generating answers vs. explanations, revealing fascinating insights into their modality use and self-consistency.👇
1
9
2
54
12,985
Successfully defended my PhD dissertation today (summa cum laude)! 🎓
Submitted! ✅ Next up: PhD defence! 🛑
39
16
2
253
25,620
❣️So happy to announce that our paper "On Measuring Faithfulness or Self-consistency of Natural Language Explanations" arxiv.org/abs/2311.07466 was accepted to #ACL2024 main! Can't wait to discuss our findings in Bangkok! Meanwhile, here's a quick overview: 👇🧵
2
7
2
39
3,541
1️⃣ We clarify the status of existing LLM explanation faithfulness tests and characterise them as self-consistency tests instead: Comparing these tests across models and data reveals widely different outcomes. This questions the view that they measure the same thing (faithfuln.).
1
2
131
Submitted! ✅ Next up: PhD defence! 🛑
23
4
2
220
34,569
Replying to @ReviewAcl
@ReviewAcl Why did you change the form where one enters the maximum load for reviewing? Now one cannot choose a maximum load of 2-3 papers (a great number for busy times. 4 papers is just not manageable for me in the following two months).
1
5
276
Finally, the ACL anonymity policy has been lifted!
ACL announcement: "The ACL Executive Committee has voted to significantly change ACL's approach to protecting anonymous peer review. The change is effective immediately. (1/4) #NLPRoc
1
8
327
Dear @ReviewAcl , is maintaining a GitHub repo against the anonymity policy? Say, I publish a paper on arxiv where I link the code (commited before the anonymity period), but then discuss an Issue that somebody else opened, or make a commit. Is it against the anonymity policy?
3
1
13
6,079
Of course the submitted paper to ARR would show an anonymous placeholder instead of the actual link to the code. My concern is that the policy is for radio silence on social media. Is GitHub activity bad too?
3
471
Yay, our work on MM-SHAP got accepted at #ACL2023nlp 🥳. MM-SHAP quantifies at dataset, instance and token-level how much multimodal encoders use each modality for their prediction. 📺 youtu.be/RLaiomLMK9I 📜 arxiv.org/abs/2212.08158
5
28
1
175
33,294
At corpus-level, we found that even on the same task, different models rely on modalities differently: They can be balanced or can show either exacerbated visual or textual focus.
1
8
522
In case you missed our VALSE 💃 poster at #acl2022nlp, you can still talk to us at our oral presentation, today at 3:15 PM (Dublin time). See you there! Paper link 📜: aclanthology.org/2022.acl-lo…
5
1
49
Congrats, Dr. Paul! 👨‍🎓
Today, I successfully defended my Ph.D. thesis! I’m incredibly grateful to my advisor @AnetteMFrank. And big thanks to my family and all the people supporting this journey!
1
2
In case you are at ALVR at #NAACL2021, we would love to see you this Friday (tomorrow) at our poster discussing “What is Multimodality?”. It’s the easy sounding questions that have the hardest answer. Joint work with @TrostNils and @AnetteMFrank @HD_NLP 💪 underline.io/events/122/sess…
2
3
1
11
1-min trailer here: 📺 youtu.be/0Mcu7MmlJMc If you’re not at #NAACL2021, stay tuned, something could jump out of @AICoffeeBreak's roaster about this topic soon! 😉
1
7
This visualization package is amazing. Now available in Python! 🔥
#Babyplots now has a #Python package! Create amazing interactive 3D plots directly in your #Jupyter notebook! Get started here: bp.bleb.li/documentation/pyt…
3
Looks like Ms. Coffee Bean has finished working this year earlier than me. 🤪
Only one week left until vacation. @YouTubeCreators already sent me AI Coffee Break's year review! Two things that really surprised me this year: * I joined @YouTube in the first place. Didn't even imagine it before 2020. * You seem to like the content and to interact with it?🤗
1
Looking forward to meet again! This time, virtually. 🙃
The COST Action Multi3Generation on multi-task, multimodal, multi-lingual language generation will have its 2nd MC Meeting on December 11th 2020 under the auspices of INESC-ID as Grant Holder and Anabela Barreiro as the Action Chair. #multi3generation #naturallanguagegeneration
1
Can't get enough of multimodal machine learning!🤯 However, I can already reveal that Ms. Coffee Bean's next @AICoffeeBreak video will not be about Multimodality. 😱
Second episode following the quest of Ms. Coffee Bean to reveal what is going wrong in multimodal vision and language fusion. This time, we are looking at the causes of the problem: youtu.be/9puivTkDaQk Link to the first part: youtu.be/P23EWdiPWDw We hope you enjoy!😌
3