A renowned researcher in the field just stopped by my poster and we chatted. One of the best moment of my career so far.

RT @c_voelcker: We often use #VAML/ #MuZero losses with deterministic models. But if we want stochastic models to measure uncertainty or to….

Paper: Code: Dataset: (6/n).

🔎While we successfully jailbroke the model, it is more important to understand why model fails. Through attention analysis, we investigate how LLM's long-context capabilities are exploited and how the instruction-following pattern is reinforced through PANDAS. (5/n).

🎯Additionally, we introduce an 𝗔daptive 𝗦ampling method to optimally select malicious dialogues during jailbreakings. Together, PANDAS significantly improves jailbreaking effectiveness and sets a new SOTA for long-context attacks. (4/n).

🐼We introduce 𝗣𝗔𝗡𝗗𝗔𝗦, a jailbreaking method that reinforces this instruction-following pattern using: .✅𝗣ositive 𝗔ffirmation: encouraging the model to continue with unsafe compliance, .❌𝗡egative 𝗗emonstrations: explicitly showing that refusal should be avoided. (3/n).

Current safety-aligned LLMs typically refuse direct malicious prompts. However, by prefixing these prompts with hundreds of malicious question-answer pairs, we can establish an instruction-following pattern that deceives the model into compliance. (2/n).

🚀Our paper on LLM jailbreaking has been accepted as a spotlight poster at ICML2025! . 🐼PANDAS: Improving Many-shot Jailbreaking via.Positive Affirmation, Negative Demonstration, and Adaptive Sampling. Collaboration with Yangchen Pan and Amir massoud Farahmand @sologen. (1/n)

🔎We also conduct an attention analysis to understand long-context vulnerabilities and how PANDAS reinforces the instruct-following behaviours in many-shot jailbreaking.

We introduce PANDAS🐼—a jailbreaking method that exploits LLMs' long-context capabilities! PANDAS significantly outperforms many-short jailbreaking by the introduction of:.✅Positive affirmations.❌Negative demonstrations.🎯Adaptive demo sampling.Paper:

RT @SoloGen: 🎉Good news, everyone! 🎉.I will recruit graduate students on the algorithmic and theoretical aspects of Reinforcement Learning.….

Huge thanks to my advisor Amir-massoud Farahmand @SoloGen and collaborators Yangchen Pan, Philip Torr, and Jindong Gu @Jindong73504766. Paper: Code:

Looking to improve the transferability of adversarial perturbations? Join us at our poster session (Thursday 10:30-12:30, #31) to explore how we transform any source model into one that generates more transferable attacks. #ECCV2024.

I’ll be presenting our work on understanding the robustness difference between models trained via different optimizers at @iclr_conf. Visit our poster (Friday 4:30-6:30 Halle B #101) to learn about the pitfall of adaptive gradient methods. #ICLR2024.Paper:

RT @SoloGen: "Without a perfect model, model-based RL is hopeless!". Our paper at #ICLR2024 challenges this belief! Even an inaccurate mode….

RT @SoloGen: Blog: Is Your Neural Network at Risk? The Pitfall of Adaptive Gradient Optimizers. Summary: Models trained using SGD exhibit s….

Another paper rejected,.CVPR review, GPT-suspected,.AC inaction, disappointed,.Innovation, undetected,.To ECCV, resubmitted.

RT @SoloGen: Did you know that the optimizer significantly affects the robustness of NN? And Adam is the wrong answer!😯."Understanding the….