Francisco Girbal Eiras

Improving AI's academic abilities may not markedly improve user experience as in the past. In LMSYS rankings, GPT-4o and GPT-4o mini rank 6th and 7th, despite a large academic gap (MMLU: 88.7% v. 82%). o1 may have underwhelmed because most people can't appreciate Olympiad skills.

@AiEleuther See also their related paper: arxiv.org/abs/2405.14782

This tutorial on the challenges of LM evaluations from @AiEleuther was one of the highlights of #ICML2024 for me, I highly recommend it! If you think evaluating these systems is easy, you probably have not spent enough time trying to do it 🤷 icml.cc/virtual/2024/t…

#ICML2024 was great! 🇦🇹 I had the opportunity to present our oral position paper on open-source Gen AI! And it was fantastic to connect with so many talented people working on important problems related to trustworthy and safe ML 🥳

The poster session for this work is today, 3:30 - 4:30pm in Hall A1 at #icml2024! x.com/fgirbal/status…

@fgirbal and @AleksPPetrov presenting our new Open Source position paper in #ICML24. The paper is accepted as Oral and the talk is today at 4.30PM in Hall A1.

3. "Mimicking User Data: On Mitigating Fine-Tuning Risks in Closed Large Language Models" (arxiv.org/abs/2406.10288) 🗓️ Poster at @NG_AI_Safety workshop, Fri 26/07, 9am - 6pm (Hall A1) #icml2024

2. "Efficient Error Certification for Physics-Informed Neural Networks" (arxiv.org/abs/2305.10157) 🗓️ Poster session, Tue 23/07, 1:30 - 3pm (Hall C 4-9, #816) #icml2024

I'm excited to be in beautiful Vienna 🇦🇹 for #icml2024! 👋 Let's grab a coffee and chat about trustworthy and safe AI 👇 See below the details of the 3 papers I'm presenting

🧵 See 👇 a thread on the main ideas of the expanded version of the paper (arxiv.org/abs/2405.08597). twitter.com/fgirbal/status…

The impact GenAI will have is highly dependent on the ability to open-source these models. Do the benefits provided outweigh the marginal risks incurred? 👉 Our #icml2024 Oral position paper argues they overwhelmingly do! 🗓️ Drop by Oral 2B (Hall A1) on 23rd of July @ 5pm.

🤗 Work developed with @Adel_Bibi @philiptorr @OxfordTVG @BunelR @DjDvij and M. Pawan Kumar @GoogleDeepMind 📆 Come to our poster on 23rd of July, 1:30-3pm at Hall C 4-9! 🧵 See below our previous thread on it: twitter.com/fgirbal/status…

After winning an Outstanding Paper Award at the 2nd Workshop of Formal Verification for ML at #icml2023 (📷 🥂), I am excited to present our work "Efficient Error Certification for Physics-Informed Neural Networks" as a poster at #icml2024! 🔗 Paper: arxiv.org/abs/2305.10157

🔗 Paper: arxiv.org/abs/2406.10288 📍 If you're at #icml2024, pass by our poster at the @NG_AI_Safety workshop on Friday, 26th of July (Hall A1) 🎉 🤗 Joint work with @AleksPPetrov @OxfordTVG, M. Pawan Kumar and @Adel_Bibi

To mitigate it, we propose during fine-tuning to mix-in safety data that mimics the format/style of the user data by paraphrasing it using an LLM. We show this simple strategy reduces harmfulness while maintaining task performance more consistently than existing baselines!

📈 Task-specific fine-tuning allows LLMs to solve tasks more efficiently. ❌ Recent work shows fine-tuning on benign/adversarial benign-looking instruction-following data increases harmfulness. 🤔 Does this happen in the task-specific setting? If so, how can we mitigate it? 🧵

Wanna know how finetuning makes a foundation model forget and how to avoid it? Have a look at our paper accepted @TmlrOrg! 📜 Paper: arxiv.org/abs/2308.13320 Work done with amazing co-authors: @yaringal @philiptorr @puneetdokania @OATML_Oxford @OxfordTVG!