🔥The 9th Round of Easy Loan, Earn $40 Reward is in progress❗️ ⏰ Promotion Period: January 15th - Feburary 15th, 2025 👉 Register now and check more details at gate.io/campaigns/358

New paper🚨: We introduce POISONBENCH, a benchmark for assessing LLM vulnerabilities to data poisoning during preference learning. Key finding: Even 3% poisoned data can cause up to 80% performance deviation when triggered. 🧵

New paper🚨 Enhancing Interpretability with Feature-Aligned Sparse Autoencoders SAEs help us understand NNs by learning sparse representations of features, but they can learn features not in the neural network they were trained on Mutual Feature Regularization mitigates this!

🧵 Excited to share our new work "Video Motion Transfer with Diffusion Transformers"! We present DiTFlow, a training-free approach to transfer motion between videos using Diffusion Transformers (DiT). Paper: huggingface.co/papers/2412.07… Project page: ditflow.github.io

TVG are excited to work with Prof Xiaowei Zhao University of Warwick on the Engineering and Physical Sciences Research Council funded SuperAIRE project to Revolutionize Renewable Energy with AI. Read more about it here: gtr.ukri.org/projects?ref=E…

🤩We are interested in taking visitors in the fields in both AI for drug discovery and for AI Safety. Please get in touch 🤩

🚨 New Paper Alert: Open Problem in Machine Unlearning for AI Safety 🚨 Can AI truly "forget"? While unlearning promises data removal, controlling emergent capabilities is a inherent challenge. Here's why it matters: 👇 Paper: arxiv.org/pdf/2501.04952 1/8

🧑‍🎓Apply for our UNIQ+ Research Internship - Navigating the Genetic Perturbation Landscape: Multi-modal causal representation learning for target discovery - tinyurl.com/ybejp4wc This is aimed at talented undergraduates from under-represented groups - Deadline 19 Feb

New paper alert! 🚨 Important question: Do SAEs generalise? We explore the answerability detection in LLMs by comparing SAE features vs. linear residual stream probes. Answer: probes outperform SAE features in-domain, out-of-domain generalization varies sharply between

🤩We are looking for a ML researcher in one of the following areas: Foundational work on agents & agentic safety, AI in the field of genomics to aid drug discovery or Explainable AI for law. Closing 24th March Apply here: tinyurl.com/wre8jus8

Introducing MALT: Improving Reasoning with Multi-Agent LLM Training🫡 We present a new multi-agent post-training method that uses credit assigned synthetic data to improve the reasoning capabilities and self-correction rates of a generator, critic, and refinement model working

New work with Lars Schaaf, Chen Lin, Guangrun Wang, and philip: We optimize neural networks to smoothly represent minimum energy paths and predict transition states for chemical reactions. Compared to the traditional approach, our method shows (i) improved resilience to

🚀 LLMs are hard to interpret? 🧠✨ Our solution: Mixture of Experts (MoE)! Built for interpretability, they create sparse, powerful models 👉 No more post-training hacks (Bye-bye sparse auto-encoders) 🔗 Read the paper: arxiv.org/abs/2503.07639 huggingface.co/papers/2503.07…

⚠️Beware: Your AI assistant could be hijacked just by encountering a malicious image online! Our latest research exposes critical security risks in AI assistants. An attacker can hijack them by simply posting an image on social media and waiting for it to be captured. [1/6] 🧵

Thrilled to share that our former postdoc, Prof X Qi xjqi.github.io has been named one of IEEE AI’s 10 to Watch – a well‐deserved honour! Proud to see you pushing the boundaries in AI. Congratulations and keep inspiring! computer.org/publications/t… #IEEEAI #Oxford #HKU

The amazing philip and I have a visiting research fellowship open now at the University of Oxford in collaboration with Pillar VC *£115k salary *£100k compute *The chance to work on cutting-edge AI Read more: lnkd.in/eRSP5pvV Apply: lnkd.in/eSizGd93

🤩Applications are open for a one year Fellowship in AI for Drugs Discovery This project is with Pillar VC @aria_research More info here: encode.pillar.vc/projects/drug-… Closing: 30 April

A huge congratulations to our award winners and thank you to the applicants, reviewers, and panellists who made this programme possible 👏🎉 aisi.gov.uk/work/strengthe…

LLMs are continuous models, but language is discrete. What happens when a continuous model approximates a discrete sequence? Spoiler: weird stuff! Glad to announce that we’ll be presenting “LLMs Are Implicitly Continuous” at ICLR 2025’s Main Track!

🚨 New paper alert: Our recent work on LLM safety has been accepted to ICLR 2025 🇸🇬 We propose a new framework for LLMs safety. 🧵 (1/7) #LLM #AISafety #ICLR2025 #Certification #AdversarialRobustness #NLP #Shhhhhh #DomainCertification #AI