What am I seeing right now?

Rethinking the Problem of Collaboration in Language Models Large language models (LLMs) have demonstrated remarkable capabilities in single-agent tasks such as question answering and structured reasoning. However, the ability to reason collaboratively—where multiple agents interact, disagree, and align on solutions—remains underdeveloped. This form of interaction is central to many human tasks, from academic collaboration to decision-making in professional contexts. Yet, most LLM training pipelines and benchmarks focus on isolated, single-turn outputs, overlooking the social dimensions of problem-solving such as assertiveness, perspective-taking, and persuasion. One primary challenge in advancing collaborative capabilities is the lack of scalable, high-quality multi-turn dialogue datasets designed for reasoning tasks. To address this limitation, Meta AI introduces Collaborative Reasoner (Coral)—a framework specifically designed to evaluate and enhance collaborative reasoning skills in LLMs. Coral reformulates traditional reasoning problems into multi-agent, multi-turn tasks, where two agents must not only solve a problem but reach consensus through natural conversation. These interactions emulate real-world social dynamics, requiring agents to challenge incorrect conclusions, negotiate conflicting viewpoints, and arrive at joint decisions. The framework spans five domains, including mathematics (MATH), STEM multiple-choice (MMLU-Pro, GPQA), and social cognition (ExploreToM, HiToM). These tasks serve as testbeds for evaluating whether models can apply their reasoning abilities in a cooperative, dialogue-driven context. Methodology: Synthetic Collaboration and Infrastructure Support Coral defines new evaluation metrics tailored to multi-agent settings. At the conversation level, agreement correctness measures whether the agents converge on the correct solution. At the turn level, social behaviors such as persuasiveness (the ability to influence another agent) and assertiveness (the ability to maintain one’s position) are explicitly quantified. To address the data bottleneck, Meta AI proposes a self-collaboration approach, where a single LLM plays both roles in a conversation. These synthetic conversations are used to generate training data through a pipeline involving tree sampling, belief filtering, and preference fine-tuning using Direct Preference Optimization (DPO). To support data generation at scale, Meta introduces Matrix, a high-performance serving framework. Matrix supports a variety of backends, employs gRPC for efficient networking, and integrates with Slurm and Ray for large-scale orchestration. Empirical comparisons show that Matrix achieves up to 1.87x higher throughput than comparable systems like Hugging Face’s llm-swarm, making it suitable for high-volume conversational training. Empirical Results: Performance Gains and Generalization Evaluation across five benchmarks reveals that collaboration, when properly modeled and trained, yields measurable gains. Fine-tuned Coral models significantly outperform baseline single-agent chain-of-thought (CoT) approaches. For instance, Llama-3.1-8B-Instruct shows a 47.8% improvement on ExploreToM after Coral+DPO training. The Llama-3.1-70B model fine-tuned on Coral surpasses GPT-4o and O1 on key collaborative reasoning tasks such as MMLU-Pro and ExploreToM. Notably, models trained via Coral exhibit improved generalization. When tested on unseen tasks (e.g., GPQA and HiToM), Coral-trained models demonstrate consistent gains—indicating that learned collaborative behaviors can transfer across domains. Despite the improvements, Coral-trained models still underperform CoT-trained baselines on complex mathematical problems (e.g., MATH), suggesting that collaboration alone may not suffice in domains requiring deep symbolic reasoning. Collaborative Reasoner provides a structured and scalable pathway to evaluate and improve multi-agent reasoning in language models. Through synthetic self-dialogue and targeted social metrics, Meta AI presents a novel approach to cultivating LLMs capable of effective collaboration. The integration of Coral with the Matrix infrastructure further enables reproducible and large-scale experimentation. As LLMs become increasingly embedded in human workflows, the ability to collaborate—rather than simply perform—is likely to be a defining capability. Coral is a step toward that direction, offering a foundation for future research on social agents capable of navigating complex, multi-agent environments. Here is the Paper, Download the Collaborative Reasoner code and Download the MATRIX code. Also, don’t forget to follow us on Twitter and join our Telegram Channel and LinkedIn Group. Don’t Forget to join our 90k+ ML SubReddit. Asif RazzaqWebsite |  + postsBioAsif Razzaq is the CEO of Marktechpost Media Inc.. As a visionary entrepreneur and engineer, Asif is committed to harnessing the potential of Artificial Intelligence for social good. His most recent endeavor is the launch of an Artificial Intelligence Media Platform, Marktechpost, which stands out for its in-depth coverage of machine learning and deep learning news that is both technically sound and easily understandable by a wide audience. The platform boasts of over 2 million monthly views, illustrating its popularity among audiences.Asif Razzaqhttps://www.marktechpost.com/author/6flvq/NVIDIA Introduces CLIMB: A Framework for Iterative Data Mixture Optimization in Language Model PretrainingAsif Razzaqhttps://www.marktechpost.com/author/6flvq/OpenAI Releases a Technical Playbook for Enterprise AI IntegrationAsif Razzaqhttps://www.marktechpost.com/author/6flvq/Meta AI Released the Perception Language Model (PLM): An Open and Reproducible Vision-Language Model to Tackle Challenging Visual Recognition TasksAsif Razzaqhttps://www.marktechpost.com/author/6flvq/An In-Depth Guide to Firecrawl Playground: Exploring Scrape, Crawl, Map, and Extract Features for Smarter Web Data Extraction

Star Wars Celebration not only revealed that Volume 3 of Star Wars: Visions will be released on October 29, 2025, but also that a new spin-off series is in the works that will debut with the next chapter of The Ninth Jedi story that began in Volume 1. Star Wars: Visions Volume 3 will feature nine short films from different Japanese anime studios, including Studio Trigger (Cyberpunk: Edgerunners), WIT Studio (Attack on Titan), David Production, Kamikaze Douga, ANIMA, Kinema citrus Co., Polygon Pictures, Production I.G., and Project Studio Q.It was also confirmed that three of the episodes will be a continuation of stories from previous seasons, and those are Kamikaze Douga's The Duel, Kinema citrus Co.'s The Village Bride, and Production I.G.'s The Ninth Jedi.Speaking of The Ninth Jedi, writer and director Kenji Kamiyama stopped by Star Wars Celebration to share that Kara's journey will be continued in this new spin-off series that will allow for longer stories from the larger Star Wars: Visions universe.While we didn't get many more details, we do know that Kara will appear alongside Juro in the upcoming 'Child of Hope' episode in Volume 3. For more, check out our review of Star Wars: Visions Volume 1 and Volume 2, as well as the news that you'll soon be caring for Grogu on Millennium Falcon: Smuggler's Run, our chat about the future of Disney Parks experiences, and all the biggest news from The Mandalorian & Grogu, Ahsoka, Andor panels.Adam Bankhurst is a writer for IGN. You can follow him on X/Twitter @AdamBankhurst and on TikTok.

Apr 20, 2025Ravie LakshmananCyber Espionage / Malware The Russian state-sponsored threat actor known as APT29 has been linked to an advanced phishing campaign that's targeting diplomatic entities across Europe with a new variant of WINELOADER and a previously unreported malware loader codenamed GRAPELOADER. "While the improved WINELOADER variant is still a modular backdoor used in later stages, GRAPELOADER is a newly observed initial-stage tool used for fingerprinting, persistence, and payload delivery," Check Point said in a technical analysis published earlier this week. "Despite differing roles, both share similarities in code structure, obfuscation, and string decryption. GRAPELOADER refines WINELOADER's anti-analysis techniques while introducing more advanced stealth methods." The use of WINELOADER was first documented by Zscaler ThreatLabz in February 2024, with the attacks leveraging wine-tasting lures to infect diplomatic staff systems. While the campaign was first attributed to a threat activity cluster named SPIKEDWINE, a subsequent analysis by Google-owned Mandiant connected it to the APT29 (aka Cozy Bear or Midnight Blizzard) hacking group, which is affiliated with Russia's Foreign Intelligence Service (SVR). The latest set of attacks entails sending email invites impersonating an unspecified European Ministry of Foreign Affairs to targets for wine-tasting events, coaxing them into clicking a link that triggers the deployment of GRAPELOADER by means of a malware-laced ZIP archive ("wine.zip"). The emails were sent from the domains bakenhof[.]com and silry[.]com. The campaign is said to have mainly singled out multiple European countries with a specific focus on Ministries of Foreign Affairs, as well as other countries' embassies in Europe. There are indications that diplomats based in the Middle East may also have been targeted. The ZIP archive contains three files: A DLL ("AppvIsvSubsystems64.dll") that serves as a dependency for running a legitimate PowerPoint executable ("wine.exe"), which is then exploited for DLL side-loading to launch a malicious DLL ("ppcore.dll"). The sideloaded malware functions as a loader (i.e., GRAPELOADER) to drop the main payload. The malware gains persistence by modifying the Windows Registry to ensure that the "wine.exe" executable is launched every time the system is rebooted. GRAPELOADER, in addition to incorporating anti-analysis techniques like string obfuscation and runtime API resolving, is designed to collect basic information about the infected host and exfiltrate it to an external server in order to retrieve the next-stage shellcode. Although the exact nature of the payload is unclear, Check Point said it identified updated WINELOADER artifacts uploaded to the VirusTotal platform with compilation timestamps matching that of "AppvIsvSubsystems64.dll." "With this information, and the fact that GRAPELOADER replaced ROOTSAW, an HTA downloader used in past campaigns to deliver WINELOADER, we believe that GRAPELOADER ultimately leads to the deployment of WINELOADER," the cybersecurity company said. The findings come as HarfangLab detailed Gamaredon's PteroLNK VBScript malware, which is used by the Russian threat actor to infect all connected USB drives with VBScript or PowerShell versions of the malicious program. The PteroLNK samples were uploaded to VirusTotal between December 2024 and February 2025 from Ukraine, a primary target of the hacking group. "Both tools, when deployed on a system, repeatedly attempt to detect connected USB drives, in order to drop LNK files and in some cases also a copy of PteroLNK onto them," ESET noted in September 2024. "Clicking on a LNK file can, depending on the particular PteroLNK version that created it, either directly retrieve the next stage from a C2 server, or execute a PteroLNK copy to download additional payloads." The French cybersecurity firm described PteroLNK VBScript files as heavily obfuscated and responsible for dynamically constructing a downloader and an LNK dropper during execution. While the downloader is scheduled to execute every 3 minutes, the LNK dropper script is configured to run every 9 minutes. The downloader employs a modular, multi-stage structure to reach out to a remote server and fetch additional malware. The LNK dropper, on the other hand, propagates through local and network drives, replacing existing .pdf, .docx, and .xlsx files in the root of the directory with deceptive shortcut counterparts and hiding the original files. These shortcuts, when launched, are engineered to run PteroLNK instead. "The scripts are designed to allow flexibility for their operators, enabling easy modification of parameters such as file names and paths, persistence mechanisms (registry keys and scheduled tasks), and detection logic for security solutions on the target system," HarfangLab said. It's worth noting that the downloader and the LNK dropper refer to the same two payloads that the Symantec Threat Hunter team, part of Broadcom, revealed earlier this month as part of an attack chain distributing an updated version of the GammaSteel stealer - NTUSER.DAT.TMContainer00000000000000000001.regtrans-ms (Downloader) NTUSER.DAT.TMContainer00000000000000000002.regtrans-ms (LNK dropper) "Gamaredon operates as a critical component of Russia's cyber operations strategy, particularly in its ongoing war with Ukraine," the company said. "Gamaredon's effectiveness lies not in technical sophistication but in tactical adaptability." "Their modus operandi combines aggressive spearphishing campaigns, rapid deployment of heavily obfuscated custom malware, and redundant C2 infrastructure. The group prioritizes operational impact over stealth, exemplified by pointing their DDRs to long-standing domains publicly linked to their past operations." Found this article interesting? Follow us on Twitter  and LinkedIn to read more exclusive content we post. SHARE    

Here are the answers for The New York Times Mini Crossword for April 20.

While capabilities like dancing can be fun and eyecatching, they don’t actually show how useful humanoid robots are in real-world situations, says Fern. Even being able to run a half marathon isn’t a very useful benchmark for their skills—it’s not like there’s market demand for robots that can compete with human runners. The benchmarks that Fern says matter to him are how well they can handle diverse real-world tasks without step-by-step human instructions. “But I would expect to see China shifting this year to focusing more on doing useful things, because people are going to be bored of dancing and karate,” Fern says.The robots who participated in the race came in a variety of forms. The shortest one was only 2 feet and 5 inches tall. Sporting a blue and white tracksuit and waving to onlookers every few seconds, it was probably the crowd favorite. The tallest, at five feet nine inches, was the winner Tiangong Ultra.What all of the robots have in common is that they are bipedal instead of running on wheels, a requirement to participate in the race. As long as the robots met that requirement, they were free to get creative, and the companies behind them adopted a wide range of strategies to try to get an advantage over their competitors. Some were wearing kid-sized sneakers (though screwed to their pedals to avoid falling off). Others were equipped with knee pads to protect their delicate parts from damage when they fell. Most of the robots had their fingers removed and some were even missing heads—you don't need such parts for running, after all, and taking them off reduces a robot’s weight and the amount of burden placed on their motors.Tiangong Ultra and another model, the N2 robot made by Chinese company Noetix Robotics, which won second place in the race, stood out for their consistent, albeit slow pace. The performance of the other humanoids was mostly disastrous. One robot called Huanhuan, which has a human-like head, only moved at the speed of a snail for a few minutes while its head shook uncontrollably—as if it could fall off any time.Another robot named Shennong looks like a real Frankenstein’s monster, with the head that resembles Gundam and four drone propellers that face backwards. It sits on a foundation with eight wheels, and it’s not clear how that alone wasn’t disqualifying. But that wasn’t even Shennong’s biggest problem, as the robot immediately twirled in two circles after taking off from the starting line, hit the wall, and dragged down its human operators with it. It was painful to watch.Duct tape proved to be the most effective problem-solving tool. Not only did the accompanying humans make makeshift robot shoes with duct tape, they also used it to adhere the head of a robot back onto its body after it repeatedly fell off during the run, making for some very jarring scenes.Every robot had human operators, often two or three running beside them. Some held control panels that allowed them to give the robot instructions, including how fast to go, while other operators led the way for their robots and tried to clear potential obstacles on the ground. Quite a few of the humanoids were being held on what looked like, well, pet leashes. “You wanna think of these robots more like running a remote control car through the race. But the robots don't have wheels,” says Fern.

WrestleMania 41 continues Sunday with Cena vs. Cody, two open match slots, and a stacked card featuring title fights, legends, and wildcards. Here's what time it starts.

According to Amex CEO Stephen Squeri, Generation Z and millennial Amex cardholders have an average FICO score of 750 and lower delinquency rates, bucking industry-wide trends. Silas Stein/picture alliance via Getty Images 2025-04-19T21:18:33Z Save Saved Read in app This story is available exclusively to Business Insider subscribers. Become an Insider and start reading now. Have an account? Data shows Gen Zers are racking up more credit card debt than older generations. But younger American Express cardholders seem to be telling a different story. Amex CEO Stephen Squeri said Gen Z and millennial customers have better-than-average FICO scores and spend less. Younger American Express credit card holders are bucking industrywide trends regarding fiscal responsibility, said Amex CEO Stephen Squeri.Squeri said during an earnings call on Thursday, in which the company surpassed Wall Street expectations,Studies have shown that the younger generation is racking up historic levels of credit card debt. A study by TransUnion showed that the average credit card debt held by 22- to 24-year-olds was $2,834, a 26% increase over millennials when they were the same age a decade ago.Christophe Le Caillec, Amex's Chief Financial Officer, said the average FICO score of its Gen Z and millennial clients is 750.In February of last year, Intuit Credit Karma data found that one in three Gen Z and millennials had a subprime credit score below 600. This is coupled with the younger generation's increasing willingness to open up more lines of credit."It's fairly normal for young people to borrow a lot during the early years of their careers, and we certainly see that happening right now with Gen Z and millennials," Rich Franks, head of Credit Karma's Light Box, previously told Business Insider.Gen Z and millennials have been a boon to Amex's business."As in past quarters, millennial and Gen Z consumers made up over 60% of new consumer accounts acquired globally in Q1," Squeri said during the earnings call, driving up revenue from fees.But where the trend diverges for younger Amex holders compared to the rest of the industry is how they use their cards and their willingness to hold onto debt, both the CEO and the CFO said.Squeri said the millennial and Gen Z segments comprised about 35% of overall spending. Part of that comes from restaurant spending, since Amex offers a competitive point rewards system for dining.But, according to Le Caillec, millennial and Gen Z customers "combined" still spent about 20% less than their older Amex counterparts.He added that they're also revolving a bit less, which is industry-speak for paying off their balance in full each billing cycle.The financial health of younger Amex cardholders could be partly explained by the type of clients the card attracts.Amex's lowest card offering requires a $95 annual fee after an initial $0 fee for the first year. Firms like Chase and Citi offer cards with cash back on purchases and no annual fee.An Amex spokesperson did not respond to a request for comment. Recommended video

Slow Rhythm Apartment / KC Design StudioSave this picture!© Yi-Hsien Lee Architects: KC Design Studio Area Area of this architecture project Area:  159 m² Year Completion year of this architecture project Year:  2024 Photographs Photographs:Yi-Hsien Lee Lead Architects: Lin Jo-yu More SpecsLess Specs Save this picture! Text description provided by the architects. Considering the living habits of the owner, the original three rooms are transformed into four. To tackle the lower pillars, we also need to consider how to ensure overall comfort and functionality within the limited space.Save this picture!Save this picture!In spatial operations, different arcs are used to replace the sharp lines of the ceiling and façade, allowing curved elements to be incorporated throughout the space. This further weakens the sense of oppression, naturally guides the sight, and makes the overall space look more open. The arcs not only extend the structure but also serve as key elements in connecting the spatial atmosphere.Save this picture!In spatial arrangements, the study room serves as the core, extending the circular flow and connecting different functional areas. The master bedroom features an elevated floor and a linear hallway, smartly dividing the space. This design not only maintains openness but also effectively separates different functions, increasing flexibility and improving the connection between spaces.Save this picture!Save this picture!The contours of life are outlined through the curved lines. The soft, warm colors dominate the overall tone. The overlapping of different chromas creates rich layers, endowing the space with a flowing sense of tranquility.Save this picture! Project gallerySee allShow less Project locationAddress:Taoyuan District, Taoyuan City, TaiwanLocation to be used only as a reference. It could indicate city/country but not exact address.About this officeKC Design StudioOffice••• Published on April 20, 2025Cite: "Slow Rhythm Apartment / KC Design Studio" 20 Apr 2025. ArchDaily. Accessed . <https://www.archdaily.com/1029223/slow-rhythm-apartment-kc-design-studio&gt ISSN 0719-8884Save世界上最受欢迎的建筑网站现已推出你的母语版本!想浏览ArchDaily中国吗?是否 You've started following your first account!Did you know?You'll now receive updates based on what you follow! Personalize your stream and start following your favorite authors, offices and users.Go to my stream

Nature, Published online: 16 April 2025; doi:10.1038/s41586-025-08915-1Simultaneous inhibition of LSD1 and GSK3 kinase promotes cell differentiation, providing a therapeutic strategy for treating acute myeloid leukaemia.