Current parameter-efficient fine-tuning methods for adapting pre-trained language models to downstream tasks are susceptible to interference from noisy data. Conventional noise-handling approaches either rely on laborious data pre-processing or employ model architecture modifications prone to error accumulation. In contrast to existing noise-process paradigms, we propose a noise-robust adaptation method via asymmetric LoRA poisoning experts (LoPE), a novel framework that enhances model robustness to noise only with generated noisy data. Drawing inspiration from the mixture-of-experts architecture, LoPE strategically integrates a dedicated poisoning expert in an asymmetric LoRA configuration. Through a two-stage paradigm, LoPE performs noise injection on the poisoning expert during fine-tuning to enhance its noise discrimination and processing ability. During inference, we selectively mask the dedicated poisoning expert to leverage purified knowledge acquired by normal experts for noise-robust output. Extensive experiments demonstrate that LoPE achieves strong performance and robustness purely through the low-cost noise injection, which completely eliminates the requirement of data cleaning.

The demand for high-fidelity test data is paramount in industrial settings where access to production data is largely restricted. Traditional data generation methods often fall short, struggling with low-fidelity and the ability to model complex data structures and semantic relationships that are critical for testing complex SQL code generation services like Natural Language to SQL (NL2SQL). In this paper, we address the critical need for generating syntactically correct and semantically relevant high-fidelity mock data for complex data structures that includes columns with nested structures that we frequently encounter in Google workloads. We highlight the limitations of existing approaches used in production, particularly their inability to handle large and complex data structures, as well as the lack of semantically coherent test data that lead to limited test coverage. We demonstrate that by leveraging Large Language Models (LLMs) and incorporating strategic pre- and post-processing steps, we can generate syntactically correct and semantically relevant high-fidelity test data that adheres to complex structural constraints and maintains semantic integrity to the SQL test targets (queries/functions). This approach supports comprehensive testing of complex SQL queries involving joins, aggregations, and even deeply nested subqueries, ensuring robust evaluation of SQL code generation services, like NL2SQL and SQL Code Assistant. Our results demonstrate the practical utility of an LLM (\textit{gemini}) based test data generation for industrial SQL code generation services where generating high-fidelity test data is essential due to the frequent unavailability and inaccessibility of production datasets for testing.

Understanding generalization in overparameterized neural networks hinges on the interplay between the data geometry, neural architecture, and training dynamics. In this paper, we theoretically explore how data geometry controls this implicit bias. This paper presents theoretical results for overparameterized two-layer ReLU networks trained below the edge of stability. First, for data distributions supported on a mixture of low-dimensional balls, we derive generalization bounds that provably adapt to the intrinsic dimension. Second, for a family of isotropic distributions that vary in how strongly probability mass concentrates toward the unit sphere, we derive a spectrum of bounds showing that rates deteriorate as the mass concentrates toward the sphere. These results instantiate a unifying principle: When the data is harder to "shatter" with respect to the activation thresholds of the ReLU neurons, gradient descent tends to learn representations that capture shared patterns and thus finds solutions that generalize well. On the other hand, for data that is easily shattered (e.g., data supported on the sphere) gradient descent favors memorization. Our theoretical results consolidate disparate empirical findings that have appeared in the literature.

Vision-Language Models (VLMs) extend large language models with visual reasoning, but their multimodal design also introduces new, underexplored vulnerabilities. Existing multimodal red-teaming methods largely rely on brittle templates, focus on single-attack settings, and expose only a narrow subset of vulnerabilities. To address these limitations, we introduce VERA-V, a variational inference framework that recasts multimodal jailbreak discovery as learning a joint posterior distribution over paired text-image prompts. This probabilistic view enables the generation of stealthy, coupled adversarial inputs that bypass model guardrails. We train a lightweight attacker to approximate the posterior, allowing efficient sampling of diverse jailbreaks and providing distributional insights into vulnerabilities. VERA-V further integrates three complementary strategies: (i) typography-based text prompts that embed harmful cues, (ii) diffusion-based image synthesis that introduces adversarial signals, and (iii) structured distractors to fragment VLM attention. Experiments on HarmBench and HADES benchmarks show that VERA-V consistently outperforms state-of-the-art baselines on both open-source and frontier VLMs, achieving up to 53.75% higher attack success rate (ASR) over the best baseline on GPT-4o.

OpenAI has unveiled ChatGPT Atlas, a long-anticipated artificial intelligence-powered web browser built around its popular chatbot, in a direct challenge to Google Chrome's dominance. OpenAI on Tuesday unveiled ChatGPT Atlas, a long-anticipated artificial intelligence-powered web browser built around its popular chatbot, in a direct challenge to Google Chrome's dominance. The launch marks OpenAI's latest move to capitalize on 800 million weekly active ChatGPT users, as it expands into more aspects of users' online lives by collecting data about consumers' browser behavior. It could accelerate a broader shift toward AI-driven search, as users increasingly turn to conversational tools that synthesize information instead of relying on traditional keyword-based results from Google -- intensifying competition between OpenAI and Google. Shares of Alphabet, which owns the Chrome browser, were down 1.8% in afternoon trading.

OpenAI's Atlas Browser Takes Direct Aim at Google Chrome The new ChatGPT-powered web browser is OpenAI's boldest play yet to reinvent how people use the web. OpenAI announced on Tuesday it's rolling out a new internet browser called Atlas that integrates directly with ChatGPT . Atlas includes features like a sidebar window people can use to ask ChatGPT questions about the web pages they visit. "We think that AI represents a rare, once a decade opportunity to rethink what a browser can be about," OpenAI CEO Sam Altman said during a livestream announcing Atlas. "Tabs were great, but we haven't seen a lot of browser innovation since then."

OpenAI's CEO, Sam Altman, testifies on Capitol Hill in Washington DC on 8 May. OpenAI's CEO, Sam Altman, testifies on Capitol Hill in Washington DC on 8 May. Company's AI-powered browser built around marquee bot is designed to provide more personalized web experience OpenAI on Tuesday launched an AI-powered web browser built around its marquee chatbot. The browser is designed to provide a more personalized web experience and includes a ChatGPT sidebar that enables users to asks questions about or engage with various aspects of each website they visit, as demonstrated in a video posted alongside the announcement. Atlas is now available globally on Apple's Mac operating system and will soon be made available on Windows, iOS and Android, according to OpenAI's announcement.

Pillay is an editorial fellow at TIME. Pillay is an editorial fellow at TIME. Every week, over 800 million people use ChatGPT to answer questions, complete tasks, and make decisions. AI systems are being rapidly adopted in schools, universities, and workplaces worldwide. Meanwhile, with billions of dollars being invested in building better systems, the technology itself continues to advance--and the future is set to be weirder than ever.

This holiday season, more shoppers are expected to use chatbots to figure out what to buy. This holiday season, rather than searching on Google, more Americans will likely be turning to large language models to find gifts, deals, and sales. Retailers could see up to a 520 percent increase in traffic from chatbots and AI search engines this year compared to 2024, according to a recent shopping report from Adobe . OpenAI is already moving to capitalize on the trend: Last week, the ChatGPT maker announced a major partnership with Walmart that will allow users to buy goods directly within the chat window. As people start relying on chatbots to discover new products, retailers are having to rethink their approach to online marketing.

Upon its release earlier this month, OpenAI's Sora 2 model took the Internet by storm, thanks to its ability to generate realistic videos from just a text prompt. But Sora is about more than just capturing eyeballs with viral content. "On the surface, Sora, for example, does not look like it is AGI-relevant," OpenAI CEO Sam Altman said on a podcast earlier this month. "But I would bet that if we can build really great world models, that will be much more important to AGI than people think." Altman was speaking to a growing belief inside the AI industry at large: that if you can simulate the world with enough accuracy, you could drop AI agents into those simulations. There, they could learn more skills than they currently can from just text, photos, and videos--because they could interact with a simulated world. That form of training could be highly efficient, in part because simulated time can be accelerated, and because many simulations can be run in parallel.