Xuanli He

[Xuanli He @zodiacJRH]

🚨 New AI Threat Alert
Multilingual LLMs can secretly transfer backdoors from one language to many others.
Spanish in, Chinese out—maliciously.
Come see how at our poster:
🗓 Today (07/28), 18:00–19:30
📍 Hall 4/5
#ACL2025 #AIsecurity #LLMsafety

[Xuanli He @zodiacJRH]

🚨 New Paper! (https://t.co/no2gxd8i5E)🚨 We uncover significant vulnerabilities in Multilingual LLMs (MLLMs) (e.g., BLOOM, Llama2, Llama3, Gemma, and GPT-3.5-turbo) to cross-lingual transferable backdoor attacks. #AIsafety #LLMs #backdoors

[Joshua Ong @ ICML @joshuaongg21]

'Theorem Prover as a Judge for Sythetic Data Generation' has been accepted to ACL (Main) 🚀. Do check us out at July 30th (Wednesday) 11:00- 12:30pm at Hall 4/5!

A huge thank you to my amazing collaborators: Shay @GiwonHong413849 @WendaLi8

📝: https://t.co/Lq7zZeBiVw

[Xuanli He @zodiacJRH]

@abeirami We show that backdoors in LLMs can transfer during post-training across 25+ languages, even when only one language was attacked. Works on lexical & semantic levels. Cross-lingual power = new security risk. (ACL’25)
PDF: https://t.co/no2gxd8i5E

[Yu Zhao @yuzhaouoe]

NAACL 2025 Oral Presentation💥

Our work about using Sparse AutoEncoder to resolve knowledge conflict will present on 30 Apr 11:30–11:45 AM • Ballroom C

Thank Hongru for presenting our work!!!

[Xuanli He @zodiacJRH]

@OwainEvans_UK Our work (https://t.co/no2gxd8PVc) demonstrated the same trend and vulnerability can be transferred across languages.

[Xuanli He @zodiacJRH]

@giffmana Our work (https://t.co/no2gxd8i5E) demonstrated that the same trend and vulnerability can be transferred across languages.

[Aryo Pradipta Gema @aryopg]

Cool work from @CohereForAI ! 🎉 However, this highlights a concern raised by our MMLU-Redux team: error propagation to many languages. Issues in MMLU (e.g., "rapid intervention to solve ebola") seem to persist in many languages. Let's solve the root cause first?

[Minqi Jiang @MinqiJiang]

OpenAI is nothing without AIDE.

Really cool to see @WecoAI’s agent framework give o1 such a huge uplift.

[Xuanli He @zodiacJRH]

Excited about using synthetic natural language explanations to boost ICL's robustness against adversarial samples? Join us at our poster session at 10:30 AM (14/08)! #ACL2024 #ICL #AdversarialRobustness

[Xuanli He @zodiacJRH]

🚀 Exciting News from Our Latest Research! 📄

🤔💡 Can Natural Language Explanations (NLEs) 📚 enhance the robustness of LLMs?

🌟 Yes! We've significantly enhanced the robustness of popular LLMs (GPT3.5-turbo, LLaMa2, Vicuna, Mistral, Zephyr)! 🔗 [https://t.co/hg0XzHqQPb]

[Xuanli He @zodiacJRH]

knock! knock! Who are you? MMLU-Redux! RUN RUN!🏃🏃‍♀️🏃‍♂️

[Bill Yuchen Lin @billyuchenlin]

A quick independent evaluation of Llama-3.1-405B-Instruct-Turbo (on @togethercompute) ⬇️
1️⃣ It ranks 1st on GSM8K!
2️⃣ Its logical reasoning ability on ZebraLogic is quite similar to Sonnet 3.5, and much better than the others. (note that ZebraLogic is a very new dataset).
3️⃣ It has worse performance on MMLU-Redux (a cleaner subset of MMLU): 76.53 vs 88.01 (gpt4o).
4️⃣ Its overall performance is between Sonnet 3.5 and GPT4o on the tested tasks. Note that this is just the "Turbo" version (fp8; not even the full precision one).

More results (e.g., WildBench) and analysis coming soon! Please stay tuned! :D

Why ZeroEval? Existing comparisons between Llama-3.1 and Sonnet 3.5, GPT4o, and other LLMs may use different prompting templates, answer extraction methods, or sampling strategies. We aim to unify these factors and do fair evaluation!

More on 🔗 https://t.co/dsUh9Ze6go

[David Ifeoluwa Adelani 🇳🇬 @davlanade]

This is a great collaboration with many authors @JessicaOjo__ @Israel_Abebe , @alabi_jesujoba @zodiacJRH @iam_OchiengM @sarahookr @AndieBukula @happybuzaaba1 @banQomania @jonathanmukiibi @SalomonKabongo @yuehgoh @RoowyM @Shmuhammadd @NanaYaaSally @SambSokhar

[Xuanli He @zodiacJRH]

joint work with @Ansh7xpex @maximilianmozes @LookFor42947068 , Mark Dras, Srinibas Swain

[Xuanli He @zodiacJRH]

🎉Excited to announce our paper got accepted at Findings of ACL 2024!🎉
🧑‍🔧Mitigate backdoor attacks—merge backdoored models with other homogeneous ones even if they aren’t fully secure—at no extra cost!
🚨Check it out: https://t.co/kxjtQWpNMH🚨
#NLP #AISafety #ACL2024 #Backdoor

[Xuanli He @zodiacJRH]

Our method works for classification (left figure) and instruction-tuning tasks (right figure) without needing training data, attack specifics, or training procedures 🧙‍♂️ Highly effective against unknown threats! ✨👀🎯 #LLMSafety

[Xuanli He @zodiacJRH]

🚀 Excited to share that our paper made it to #ACL2024 main conference! 📄 Dive into the newest insights here: https://t.co/hg0XzHqQPb #NLP #Research Big cheers to @YuxiangJWu @oanacamb @PMinervini and Pontus Stenetorp🎉

[Xuanli He @zodiacJRH]

🚀 Exciting News from Our Latest Research! 📄

🤔💡 Can Natural Language Explanations (NLEs) 📚 enhance the robustness of LLMs?

🌟 Yes! We've significantly enhanced the robustness of popular LLMs (GPT3.5-turbo, LLaMa2, Vicuna, Mistral, Zephyr)! 🔗 [https://t.co/hg0XzHqQPb]

[Xuanli He @zodiacJRH]

Sorry @Meta and @Google, Llama 2,3 and Gemma are also susceptible! 🤒

Our study dives into the vulnerabilities of major open-source LLMs, including BLOOM, Llama2, Llama3, and Gemma. Turns out, the more powerful the model, the greater its risks for cross-lingual backdoor attacks.

[Xuanli He @zodiacJRH]

🚨 New Paper! (https://t.co/no2gxd8i5E)🚨 We uncover significant vulnerabilities in Multilingual LLMs (MLLMs) (e.g., BLOOM, Llama2, Llama3, Gemma, and GPT-3.5-turbo) to cross-lingual transferable backdoor attacks. #AIsafety #LLMs #backdoors

[Xuanli He @zodiacJRH]

Plus, our method applies to cross-lingual responses and achieves a 50% average success rate across 25 languages🌍 ! #GPT3.5-turbo