Large language models (LLMs) hold great promise for medical applications and are evolving rapidly, with new models being released at an accelerated pace. However, benchmarking on large-scale real-world data such as electronic health records (EHRs) is critical, as clinical decisions are directly informed by these sources, yet current evaluations remain limited. Most existing benchmarks rely on medical exam-style questions or PubMed-derived text, failing to capture the complexity of real-world clinical data. Others focus narrowly on specific application scenarios, limiting their generalizability across broader clinical use. To address this gap, we present BRIDGE, a comprehensive multilingual benchmark comprising 87 tasks sourced from real-world clinical data sources across nine languages. It covers eight major task types spanning the entire continuum of patient care across six clinical stages and 20 representative applications, including triage and referral, consultation, information extraction, diagnosis, prognosis, and billing coding, and involves 14 clinical specialties. We systematically evaluated 95 LLMs (including DeepSeek-R1, GPT-4o, Gemini series, and Qwen3 series) under various inference strategies. Our results reveal substantial performance variation across model sizes, languages, natural language processing tasks, and clinical specialties. Notably, we demonstrate that open-source LLMs can achieve performance comparable to proprietary models, while medically fine-tuned LLMs based on older architectures often underperform versus updated general-purpose models. The BRIDGE and its corresponding leaderboard serve as a foundational resource and a unique reference for the development and evaluation of new LLMs in real-world clinical text understanding. The BRIDGE leaderboard: https://huggingface.co/spaces/YLab-Open/BRIDGE-Medical-Leaderboard

Leveraging inference-time search in large language models has proven effective in further enhancing a trained model's capability to solve complex mathematical and reasoning problems. However, this approach significantly increases computational costs and inference time, as the model must generate and evaluate multiple candidate solutions to identify a viable reasoning path. To address this, we propose an effective approach that integrates search capabilities directly into the model by fine-tuning it on unpaired successful (learning) and failed reasoning paths (forgetting) derived from diverse search methods. A key challenge we identify is that naive fine-tuning can degrade the model's search capability; we show this can be mitigated with a smaller learning rate. Extensive experiments on the challenging Game-of-24 and Countdown arithmetic puzzles show that, replacing CoT-generated data with search-generated data for offline fine-tuning improves success rates by around 23% over inference-time search baselines, while reducing inference time by 180$\times$. On top of this, our learning and forgetting objective consistently outperforms both supervised fine-tuning and preference-based methods.

Alignment of large language models (LLMs) has predominantly relied on pairwise preference optimization, where annotators select the better of two responses to a prompt. While simple, this approach overlooks the opportunity to learn from richer forms of human feedback, such as multiwise comparisons and top-$k$ rankings. We propose Ranked Choice Preference Optimization (RCPO), a unified framework that bridges preference optimization with (ranked) choice modeling via maximum likelihood estimation. The framework is flexible, supporting both utility-based and rank-based choice models. It subsumes several existing pairwise methods (e.g., DPO, SimPO), while providing principled training objectives for richer feedback formats. We instantiate this framework with two representative ranked choice models (Multinomial Logit and Mallows-RMJ). Empirical studies on Llama-3-8B-Instruct and Gemma-2-9B-it across AlpacaEval 2 and Arena-Hard benchmarks show that RCPO consistently outperforms competitive baselines. RCPO shows how directly leveraging ranked preference data, combined with the right choice models, yields more effective alignment. It offers a versatile and extensible foundation for incorporating (ranked) choice modeling into LLM training.

Matrix-based preconditioned optimizers, such as Muon, have recently been shown to be more efficient than scalar-based optimizers for training large-scale neural networks, including large language models (LLMs). On the other hand, recent benchmarks on optimizers for LLM pre-training have demonstrated that variance-reduction techniques such as MARS can achieve substantial speedups over standard optimizers that do not employ variance reduction. In this paper, to achieve the best of both worlds, we introduce MARS-M, a new optimizer that integrates the variance reduction technique in MARS with Muon. Under standard regularity conditions, we prove that Muon-M converges to a first-order stationary point at a rate of $\tilde{\mathcal{O}}(T^{-1/3})$, which improves upon $\tilde{\mathcal{O}}(T^{-1/4})$ rate attained by Muon. Our empirical results on language modeling and computer vision tasks demonstrate that MARS-M consistently yields lower losses and improved performance across various downstream benchmarks. The implementation of MARS-M is available at https://github.com/AGI-Arena/MARS/tree/main/MARS_M.

Earlier today, OpenAI completed a controversial restructuring of its for-profit arm into a public benefit corporation: the latest gust in a whirlwind that has swept up hundreds of billions of dollars of global investment for artificial intelligence (AI) tools. But even as the AI company--founded as a nonprofit, now valued at 500 billion--completes its long-awaited restructuring, a nagging issue with its core offering remains unresolved: hallucinations. Large language models (LLMs) such as those that underpin OpenAI's popular ChatGPT platform are prone to confidently spouting factually incorrect statements. These blips are often attributed to bad input data, but in a preprint posted last month, a team from OpenAI and the Georgia Institute of Technology proves that even with flawless training data, LLMs can never be all-knowing--in part because some questions are just inherently unanswerable. However, that doesn't mean hallucinations are inevitable.

An illustration photo shows the OpenAI logo displayed on a smartphone with the Microsoft logo in the background in Chongqing, China on Aug. 27, 2025. An illustration photo shows the OpenAI logo displayed on a smartphone with the Microsoft logo in the background in Chongqing, China on Aug. 27, 2025. OpenAI has completed a restructuring, dividing itself into a nonprofit and for-profit entity, the company announced on Tuesday. The nonprofit arm, now called the OpenAI Foundation, will have a $130 billion stake in the for-profit enterprise, a public benefit corporation called OpenAI Group PBC. "The OpenAI Foundation and OpenAI Group will work in concert to advance solutions to hard problems and opportunities posed by AI progress," the company said in its blog post announcing the restructuring. "This includes making intelligence a tool that everyone can benefit from, building safe and aligned systems, turbocharging scientific discovery, and strengthening global cooperation and resilience."

When you purchase through links in our articles, we may earn a small commission. Hundreds of thousands of users are also showing signs of strong emotional attachment to ChatGPT, with many exhibiting signs of psychosis or mania. OpenAI has published new data showing that a small but significant percentage of ChatGPT users aren't just talking to the chatbot about mental health issues but engaging in suicidal planning. About 0.15 percent of active weekly users "have conversations that include explicit indicators of potential suicidal planning or intent." A similar proportion of users are reportedly showing signs of strong emotional attachments to ChatGPT, with hundreds of thousands of users exhibiting signs of psychosis or mania.

Microsoft and OpenAI have reached a deal to allow the ChatGPT maker to restructure itself into a public-benefit corporation, valuing OpenAI at $500bn and giving it more freedom in its business operations. The deal, unveiled on Tuesday, removes a major constraint on raising capital for OpenAI that has existed since 2019. As its ChatGPT service exploded in popularity, those limitations had become a notable source of tension between the two companies. Microsoft will still hold a stake of about $135bn, or 27 percent, in OpenAI Group PBC, which will be controlled by the OpenAI Foundation, a nonprofit, the companies said. Microsoft, based in Redmond, Washington in the United States, has invested $13.8bn in OpenAI, with Tuesday's deal implying that the firm had generated a return of nearly 10 times its investment.

Sam Altman speaks in San Francisco on 2 June 2025. Sam Altman speaks in San Francisco on 2 June 2025. OpenAI said on Tuesday it had converted its main business into a for-profit corporation, the conclusion of a lengthy and fraught legal saga. A crucial regulator, Kathy Jennings, the Delaware attorney general, said she approved the plan for the startup, which began as a non-profit in 2015, to change to a public benefit corporation, a type of for-profit entity that expresses commitment to bettering society. The company also said it had reorganized its ownership structure and signed a new agreement with its longtime backer Microsoft that gives the software giant a roughly 27% stake in OpenAI's new for-profit corporation, but changes some of the details of their close partnership.

Hurricane tracker shows Melissa is now stronger than Katrina as'storm of the century' closes in on Jamaica: Live updates Hurricane tracker reveals Melissa's path over Jamaica and threat to US coast as it becomes stronger than Katrina I know the pathetic truth about Kristen Bell's'cry for help' that will settle this domestic violence scandal once and for all: KENNEDY LIZ JONES: Why I believe ruthless Kate's the driving force behind Andrew's eviction - and why no one now dares cross her William'threatened to strip Eugenie and Beatrice of their titles unless Andrew and Fergie left Royal Lodge' Raunchy photos and violent death: Unraveling of famous life coach's spoilt daughter who decided to drive mom's Lexus at 124mph Rare earthquake hits Maryland's ancient fault, exposing hidden seismic risks along the East Coast'Humiliating' truth about influencer TooTurntTony and his extreme stunts: He's ripped, makes $3m a year and has all the hottest girls... but a dark reality lies beneath Beyonce's girl Blue Ivy, 13, looks just like her famous mom as she supports grandma Tina Knowles at Angel Ball A torrid affair with his friend's wife, a teenage model and'problems' with alcohol...the past that could ruin Gavin Newsom's White House bid Frustrated Facebook boss Mark Zuckerberg asked AG Pam Bondi's advice on how to talk to Trump, new book claims'She hasn't told the full story. This is typical her': How David Harbour is'looking after' Lily Allen's daughters despite'victim' singer publicly humiliating him... as insider tells DOLLY BUSBY what's REALLY going on Bill Maher eats up Charlie Sheen's idea to create a'special place' for repeat offenders: 'That's very good' Heidi Klum, 52, and daughter Leni, 21, brush off backlash towards their'inappropriate' lingerie shoots as they slip into matching nude knitwear for new Intimissimi campaign Why I now fear my daughter's rare genetic condition is linked to me becoming a father later in life. This is the evidence you can't ignore: PROF ROB GALLOWAY READ MORE: Amazon's delivery drivers will be forced to wear AI GLASSES In a recent blog post, the AI giant warned that 0.07 per cent of its weekly users showed signs of serious mental health emergencies. While this figure might sound small, with over 800 million weekly users according to CEO Sam Altman, that adds up to 560,000 users. Meanwhile, 1.2 million users - 0.15 per cent - send messages that contain'explicit indicators of potential suicidal planning or intent' each week.