Simulating step-wise human behavior with Large Language Models (LLMs) has become an emerging research direction, enabling applications in various practical domains. While prior methods, including prompting, supervised fine-tuning (SFT), and reinforcement learning (RL), have shown promise in modeling step-wise behavior, they primarily learn a population-level policy without conditioning on a user's persona, yielding generic rather than personalized simulations. In this work, we pose a critical question: how can LLM agents better simulate personalized user behavior? We introduce Customer-R1, an RL-based method for personalized, step-wise user behavior simulation in online shopping environments. Our policy is conditioned on an explicit persona, and we optimize next-step rationale and action generation via action correctness reward signals. Experiments on the OPeRA dataset emonstrate that Customer-R1 not only significantly outperforms prompting and SFT-based baselines in next-action prediction tasks, but also better matches users' action distribution, indicating higher fidelity in personalized behavior simulation.

In just seven minutes, the thieves took off with crown jewels containing with thousands of diamonds along with other precious gems. Police stand outside the Louvre after a brazen theft. Could the French TV series have been prophetic? The show envisioned a heist at the Louvre, an event that became reality on the morning of October 19, when a group of professional thieves managed to break into the world-famous Paris museum . In just seven minutes, they stole a host of priceless French crown jewels.

Large language models (LLMs) are moving beyond static uses and are now powering agents that learn continually during their interaction with external environments. For example, agents can learn reusable skills while navigating web pages or toggling new tools. However, existing methods for skill learning often create skills that are over-specialized to a single website and fail to generalize. We introduce PolySkill, a new framework that enables agents to learn generalizable and compositional skills. The core idea, inspired by polymorphism in software engineering, is to decouple a skill's abstract goal (what it accomplishes) and its concrete implementation (how it is executed). Experiments show that our method (1) improves skill reuse by 1.7x on seen websites and (2) boosts success rates by up to 9.4% on Mind2Web and 13.9% on unseen websites, while reducing steps by over 20%. (3) In self-exploration settings without specified tasks, our framework improves the quality of proposed tasks and enables agents to learn generalizable skills that work across different sites. By enabling the agent to identify and refine its own goals, the PolySkill enhances the agent's ability to learn a better curriculum, leading to the acquisition of more generalizable skills compared to baseline methods. This work provides a practical path toward building agents capable of continual learning in adaptive environments. Our findings show that separating a skill's goal from its execution is a crucial step toward developing autonomous agents that can learn and generalize across the open web continuously.

All products featured on WIRED are independently selected by our editors. However, we may receive compensation from retailers and/or from purchases of products through these links. Fall is for nesting--and for feathering your nest with whatever will keep you sane during the winter. Which is why a number of retailers, including The Home Depot, drop prices on home goods with big fall deals. The Home Depot fall savings event for 2025 is unusually broad, because The Home Depot itself is unusually broad--the store that first brought the home improvement superstore nationwide.

Gear Home This 297-piece Kobalt Mechanics Tool Kit is just $99 at Lowe's with an included tool box This kit is typically $150, but it's just $99 at Lowe's, which makes it a fantastic gift for just about anyone. We may earn revenue from the products available on this page and participate in affiliate programs. I truly believe that a big tool kit with a dedicated carrying case is one of the best gifts you can give. It looks really impressive, it's useful for every type of person, and it's easy to wrap because it's usually rectangular (though, I recommend ditching wrapping paper this year). Right now, Lowe's has this 297-piece Mechanics Tool Set for just $99, which is a total sweet spot for gift buying.

The Bose QuietComfort Ultra get a rare discount until the end of the day. All products featured on WIRED are independently selected by our editors. However, we may receive compensation from retailers and/or from purchases of products through these links. Bose is well known for its noise-canceling headphones and earbuds, and the high-end QuietComfort Ultra (9/10, WIRED Recommends) are currently marked down to just $329 on Amazon, with the same discount at Best Buy . You'll have to move fast, though, as both sites feature countdown timers with less than 24 hours remaining as I write this.

Cory Doctorow's theory of "enshittification" explains how tech platforms rot from within. As AI grows more profitable--and powerful--it risks the same fate. Cory Doctorow speaks onstage during Unfinished Live at The Shed in New York City. As one does these days, I ran my itinerary past GPT-5 for sightseeing suggestions and restaurant recommendations. The bot reported that the top choice for dinner near our hotel in Rome was a short walk down Via Margutta.

Blueprint combines AI and curated designs to get your website up and running in a few minutes. All products featured on WIRED are independently selected by our editors. However, we may receive compensation from retailers and/or from purchases of products through these links. Squarespace wants you to build a website with AI. Its Blueprint AI feature began life as a guided website design system in 2023, but, like many things these days, it was revamped with artificial intelligence.

China's biggest shopping event starts five weeks early to revive spending It's known to be China's biggest online shopping event - taking place on 11 November each year. But this year, Single's Day sales have already begun in mid-October, as part of efforts by Chinese retailers to boost spending in a sluggish market. China has been plagued with issues like growing youth unemployment, a prolonged property crisis, steep government debt and an ongoing trade war with the US - all of which is making the country's consumers cut back on spending. The Chinese government has been spending billions - through family subsidies, more wages and discounts for consumer goods in a bid to counter this, but retail sales growth is still failing to meet expectations. Originally created by Alibaba as a Chinese shopping festival, Singles' Day is akin to Amazon's Prime Day or Black Friday promotions elsewhere in the world.

Model distillation has become essential for creating smaller, deployable language models that retain larger system capabilities. However, widespread deployment raises concerns about resilience to adversarial manipulation. This paper investigates vulnerability of distilled models to adversarial injection of biased content during training. We demonstrate that adversaries can inject subtle biases into teacher models through minimal data poisoning, which propagates to student models and becomes significantly amplified. We propose two propagation modes: Untargeted Propagation, where bias affects multiple tasks, and Targeted Propagation, focusing on specific tasks while maintaining normal behavior elsewhere. With only 25 poisoned samples (0.25% poisoning rate), student models generate biased responses 76.9% of the time in targeted scenarios - higher than 69.4% in teacher models. For untargeted propagation, adversarial bias appears 6x-29x more frequently in student models on unseen tasks. We validate findings across six bias types (targeted advertisements, phishing links, narrative manipulations, insecure coding practices), various distillation methods, and different modalities spanning text and code generation. Our evaluation reveals shortcomings in current defenses - perplexity filtering, bias detection systems, and LLM-based autorater frameworks - against these attacks. Results expose significant security vulnerabilities in distilled models, highlighting need for specialized safeguards. We propose practical design principles for building effective adversarial bias mitigation strategies.