🔥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

Agents research with LMs right now feels like Deep RL in the late 2010s. Tons of new algorithms on narrow domains, so I expect most of these results to not transfer at all. The thing is, its still early, so its going to get worse still before it gets better.

Personalization in LLMs is crucial for meeting diverse user needs, yet collecting real-world preferences at scale remains a significant challenge. Introducing FSPO, a simple framework leveraging synthetic preference data to adapt new users with meta-learning for open-ended QA! 🧵

Prompt caching lowers inference costs but can leak private information from timing differences. Our audits found 7 API providers with potential leakage of user data. Caching can even leak architecture info—OpenAI's embedding model is likely a decoder-only Transformer! 🧵1/9

As AI systems increasingly simulate human behavior, we must ask: How do we ensure they don’t amplify bias, deceive, or manipulate? This paper lays out a much-needed framework for responsible AI design. Its REALLY good. arxiv.org/abs/2503.02250

people love and dream about movies like The Martian and Interstellar but then celebrate defunding basic science research to save a few dollars in taxes per year

(1/n) Fine tuning APIs create significant security vulnerabilities, breaking alignment in frontier models for under $100! Introducing NOICE, a fine-tuning attack that requires just 1000 training examples to remove model safeguards. The strangest part: we use ONLY harmless data.

Google, Anthropic, xAI etc should have a Model Spec. Would help them with all of these if done right: Developers: Know what future models will become Internal: Focus to define and to deliver your goals Regulators: Transparency into wtf frontier labs care about

An LLM generates an article verbatim—did it “train on” the article? It’s complicated: under n-gram definitions of train-set inclusion, LLMs can complete “unseen” texts—both after data deletion and adding “gibberish” data. Our results impact unlearning, MIAs & data transparency🧵

Incredibly relevant work— suggesting the real bottleneck for effective misinformation isn’t technical detection (watermarking) but purely capabilities (how persuasive and fluent models become)… a chilling shift for AI safety efforts

Thoughtful retrospective on where RL folks went wrong (focus on algorithms, ignore priors should have been flipped) and where to go next. Domains w dense rewards seem ~solved (verified reasoning, RLHF) so more open ended evals are next (chatbot arena)

Let's open-source GPT-4. If the non-profit genuinely controlled OpenAI, it's hard to see why they wouldn't release the model. Tons of science that would be unlocked with this. Tons of papers that become reproducible.

Excited to announce our work got accepted as a spotlight to ICML!! If people have recs for Vancouver LMK!