Microsoft's subscription service, Xbox Game Pass, has grown into a must-have this year thanks to a whole host of big releases, but is it enough to help Xbox catch up to PlayStation?Tech13:00, 23 May 2025Game Pass is arguably more important than an Xbox console these days(Image: Future Publishing via Getty Images)It’d be fair to say at times that Xbox’s strategy this console generation has been clear as mud.From its opening salvo of multiplatform releases on PlayStation and Switch, to other games being released day and date on other systems, and some even looking better on PS5 Pro, you’d be reasonably wondering if you need an Xbox at all.‌In fact, Microsoft agrees, with its “This is an Xbox” pretty much labelling everything except your toaster and the kitchen sink as a console of sorts. The Xbox as we know it is long gone, it seems, but in its place, Xbox Game Pass feels like it’s grown in momentum month after month .‌Is it enough to turn the tide? It’s hard to say.Halo Infinite marked a sort of 'false dawn' for Game Pass back in 2021(Image: ESA PR Supplied.)Cast your mind back to late 2021. Microsoft released Forza Horizon 5 and Halo Infinite in consecutive months, and it felt as though Xbox Game Pass was about to kick into overdrive.Article continues belowSadly, things went a little quieter after that (and Halo’s momentum was curtailed for other reasons), but now, years later, it feels Xbox Game Pass is poised to finally hit some kind of stride.It’d be fair to say that Microsoft’s acquisition of Bethesda and Activision Blizzard has left it with a studio portfolio that would make many envious, but it’s not really managed to line up its releases. Even Avowed, one of this year’s best games (and one we’ll touch on shortly), was slated for last year before Microsoft nudged it to 2025.Still, as it turns out, having great games on your platform will get players through the door, and if you’ve got a PC then you’re in an even better position.‌Stalker 2 has been heavily patched since launch(Image: GSC Game World)September saw Frostpunk 2 arrive on Xbox Game Pass for PC, while October added Call of Duty Black Ops 6. November brought with it STALKER 2 on day one, and December saw the year out with Indiana Jones and the Great Circle.Unlike 2021, though, Microsoft has kept things moving between calendar years. January kicked off with Ninja Gaiden 2 Black as a stealth release, along with GOTY sequel Citizen Sleeper 2: Starward Vector. February added Avowed, March brought with it Atomfall, and April added South of Midnight, critical darling Blue Prince, and the day one release of Elder Scrolls IV: Oblivion Remastered.‌All of these games are available on PS5, sure, but the cost would be astronomical to play all of them, and Microsoft still isn’t done.DOOM: The Dark Ages is on PC and Xbox, along with Monster Train 2, this month, while Tony Hawk’s Pro Skater 3+4 remakes , Wuchang: Fallen Feathers, Gears of War: Reloaded, The Outer Worlds 2 and much more are still coming in 2025.It’s a frankly ridiculous list of games with more to come in 2026, and that’s saying nothing of freebies like in-game items for League of Legends, Overwatch 2, and more, as well as this week’s announcement of a huge retro game collection .‌ESO is still growing, even after a decade(Image: Bethesda Softworks/ZeniMax Media )One of the biggest knocks against Game Pass in recent years has been the increase in price, and while there’s an argument to be made about Microsoft’s tiers making it harder to know if you’re getting a game on day one, Xbox Game Pass Core continues to grow.Core is cheaper than the Ultimate option that includes console and PC, but it does have a sizeable collection of games for budget-conscious players.‌The recently added Retro Classics are included, for example, as well as Bethesda hits like DOOM Eternal, Dishonored 2, and Fallout 4. For casual gamers, that might be enough.In fact, there’s a whole host of games included that could consume your week if you let them, like State of Decay 2, The Elder Scrolls Online, and Stardew Valley.That means there’s likely to be something for everyone while you wait for more recent releases to trickle down.‌Game Pass has really come into its own this year(Image: Microsoft)Most of my friends have a PlayStation 5, so if a game doesn’t have cross-play, that’s where I play my games—Game Pass or no.I do wonder how many feel the same, and see the appeal of Game Pass but simply don’t have the inclination (or money) to pick up a second console to play games cheaper in the long run, especially with so many coming to PlayStation in the future.‌Given Microsoft now has the keys to Warcraft, Overwatch, Elder Scrolls, Fallout, Gears of War, Halo, Forza and so much more, I wonder if there’s ever a scenario where they start to rebuild walls between their franchises.Sure, Call of Duty will keep making money hand over fist on as many platforms as possible, but would the next Dishonored be worth putting on all platforms when you could keep it to the Xbox platform as an added incentive?Microsoft can’t be happy with being in second place for another console generation, and while I do wonder if Sony will eventually reckon with rising budgets by starting to move its own games elsewhere (and not just to PC), I’d wager we’ll get one more Xbox generation at most before the company goes the way of SEGA.Article continues belowWhat do you think? Does Game Pass interest you? Are you always likely to play where your friends are? And can Microsoft keep up this momentum?For the latest breaking news and stories from across the globe from the Daily Star, sign up for our newsletters.‌‌‌

May 23, 2025The Hacker NewsWeb Security / Threat Detection From zero-day exploits to large-scale bot attacks — the demand for a powerful, self-hosted, and user-friendly web application security solution has never been greater. SafeLine is currently the most starred open-source Web Application Firewall (WAF) on GitHub, with over 16.4K stars and a rapidly growing global user base. This walkthrough covers what SafeLine is, how it works, and why it's becoming the go-to solution over cloud-based WAFs. What is SafeLine WAF? SafeLine is a self-hosted web application firewall that acts as a reverse proxy, filtering and monitoring HTTP/HTTPS traffic to block malicious requests before they reach your backend web applications. Unlike cloud-based WAFs, SafeLine runs entirely on your own servers—giving you unmatched visibility and data sovereignty. Key Features of SafeLine WAFComprehensive Attack Prevention SafeLine effectively blocks a wide range of common and advanced web attacks, including SQL injection(SQLi), cross-site scripting (XSS), OS command injection, CRLF injection, XML External Entity (XXE) attacks, Server Side Request Forgery (SSRF), and directory traversal, etc. Zero-Day Detection via Semantic Analysis Unlike traditional signature-based WAFs, SafeLine uses a patented semantic analysis engine that deeply parses HTTP traffic semantics. This approach enables it to detect complex and zero-day attacks with high accuracy, resulting in an industry-leading detection rate of 99.45% and an ultra-low false positive rate of 0.07%. (The chart below compares SafeLine with the two versions of a globally recognized open-source WAF.) Robust Bot Protection SafeLine delivers comprehensive, multi-layered defenses against automated bot attacks, a growing threat vector responsible for credential stuffing, malicious scraping, inventory hoarding, and vulnerability scanning. It combines several out-of-box powerful mechanisms: CAPTCHA Challenges: Dynamically issued to distinguish human users from automated clients, especially in suspicious or high-risk traffic scenarios. Dynamic Protection: Randomly encrypts and obfuscates frontend code, such as HTML and JavaScript, before delivering it to the client. This prevents bots from reliably parsing page structures or interacting with DOM elements, rendering automated scripts ineffective. Anti-Replay Mechanisms: Detect and block reuse of tokens, headers, or payloads often leveraged in scripted attacks or credential stuffing campaigns. HTTP Flood DDoS Mitigation HTTP flood DDoS attacks attempt to overwhelm servers by sending massive volumes of HTTP requests in a short period of time. These attacks can exhaust server resources, degrade performance, or take applications offline entirely. To counter this, SafeLine implements rate limiting to cap request frequency and mitigate abuse. These measures are highly configurable, allowing defenders to tailor thresholds based on real-world traffic patterns. For sudden traffic spikes—whether legitimate or malicious—SafeLine provides a virtual waiting room mechanism. This ensures service availability by queuing excess users and releasing them gradually, preventing backend overload while maintaining a fair and orderly access experience. Authentication Challenges SafeLine is also designed with Zero Trust principles in mind—never trust, always verify. It offers configurable visitor authentication to secure access to protected applications, enhancing security through enforced identity checks. As a built-in identity gateway, it supports modern authentication protocols such as OIDC and integrates seamlessly with identity providers like GitHub and others. SafeLine also supports Single Sign-On (SSO) to streamline user authentication and simplify login experience in the meantime. Best of all, these enterprise-grade identity features are included for free. Simple Deployment in Minutes SafeLine is designed for quick setup and easy management. It requires the following environment to be installed and run: Operating System: Linux (x86_64 or arm64) Dependencies: Docker (version 20.10.14 or higher) and Docker Compose (version 2.0.0 or higher) Minimum System Requirements: 1 CPU core, 1 GB of RAM, and 5 GB of available disk space Once the environment is ready, installation takes just a few minutes with a single command. bash -c "$(curl -fsSLk https://waf.chaitin.com/release/latest/manager.sh)" -- --en A user-friendly, wizard-based interface guides you through configuration. Full documentation is available here. Why Choose SafeLine Over Cloud-Based WAFs? Unlike traditional cloud-based WAFs that route your traffic through third-party infrastructure, SafeLine offers complete deployment autonomy. Here are the advantages: Full Data Control: Sensitive traffic and logs remain on-premises, reducing exposure to third-party cloud risks. Cost Efficiency: Avoids recurring subscription fees common with cloud WAFs, especially beneficial for high-traffic environments. Free and Out-of-Box Enterprise Features: Advanced threat detection, bot protection, identity authentication, and more—typically gated behind "premium" tiers elsewhere—are out-of-box and included for free. Get SafeLine — free forever for personal use, with optional 7-day Pro trial. Use Cases Ideal for SafeLine SafeLine is a versatile solution built for a wide range of web application security needs. It's particularly well-suited for: Organizations with strict data privacy or regulatory compliance requirements Teams Targeted by Sophisticated Bots and Automated Threats Small and medium-sized businesses seeking affordable, enterprise-grade protection DevOps and Security Teams Requiring Full Deployment Control and Customization Projects requiring rapid deployment and easy maintenance Final Words SafeLine stands out as a powerful, open-source alternative to traditional cloud-based WAFs. With cutting-edge zero-day detection, robust bot mitigation, and zero trust–aligned identity features—all bundled into a self-hosted, easy-to-deploy package—SafeLine empowers developers, security teams, and organizations of all sizes to take control of their web security. Get SafeLine — free forever for personal use, with optional 7-day Pro trial. Found this article interesting? This article is a contributed piece from one of our valued partners. Follow us on Twitter  and LinkedIn to read more exclusive content we post. SHARE    

The U.S. Department of Justice (DoJ) on Thursday announced the disruption of the online infrastructure associated with DanaBot (aka DanaTools) and unsealed charges against 16 individuals for their alleged involvement in the development and deployment of the malware, which it said was controlled by a Russia-based cybercrime organization. The malware, the DoJ said, infected more than 300,000 victim computers around the world, facilitated fraud and ransomware, and caused at least $50 million in damages. Two of the defendants, Aleksandr Stepanov (aka JimmBee), 39, and Artem Aleksandrovich Kalinkin (aka Onix), 34, both from Novosibirsk, Russia, are currently at large. Stepanov has been charged with conspiracy, conspiracy to commit wire fraud and bank fraud, aggravated identity theft, unauthorized access to a protected computer to obtain information, unauthorized impairment of a protected computer, wiretapping, and use of an intercepted communication. Kalinkin has been charged with conspiracy to gain unauthorized access to a computer to obtain information, to gain unauthorized access to a computer to defraud, and to commit unauthorized impairment of a protected computer. The unsealed criminal complaint and indictment show that many of the defendants, counting Kalinkin, exposed their real-life identities after accidentally infecting their own systems with the malware. "In some cases, such self-infections appeared to be deliberately done in order to test, analyze, or improve the malware," the complaint [PDF] read. "In other cases, the infections seemed to be inadvertent – one of the hazards of committing cybercrime is that criminals will sometimes infect themselves with their own malware by mistake." "The inadvertent infections often resulted in sensitive and compromising data being stolen from the actor's computer by the malware and stored on the DanaBot servers, including data that helped identify members of the DanaBot organization." If convicted, Kalinkin is expected to face a statutory maximum sentence of 72 years in federal prison. Stepanov would face a jail term of five years. Concurrent with the action, the law enforcement effort, carried out as part of Operation Endgame, saw DanaBot's command-and-control (C2) servers seized, including dozens of virtual servers hosted in the United States. "DanaBot malware used a variety of methods to infect victim computers, including spam email messages containing malicious attachments or hyperlinks," the DoJ said. "Victim computers infected with DanaBot malware became part of a botnet (a network of compromised computers), enabling the operators and users of the botnet to remotely control the infected computers in a coordinated manner." DanaBot, like the recently dismantled Lumma Stealer malware, operates under a malware-as-a-service (MaaS) scheme, with the administrators leasing out access starting from $500 to "several thousand dollars" a month. Tracked under the monikers Scully Spider and Storm-1044, is a multi-functional tool along the lines of Emotet, TrickBot, QakBot, and IcedID that's capable of acting as a stealer and a delivery vector for next-stage payloads, such as ransomware. The Delphi-based modular malware is equipped to siphon data from victim computers, hijack banking sessions, and steal device information, user browsing histories, stored account credentials, and virtual currency wallet information. It can also provide full remote access, log keystrokes, and capture videos. It's been active in the wild since its debut in May 2018, when it started off as a banking trojan. Example of typical Danabot infrastructure "DanaBot initially targeted victims in Ukraine, Poland, Italy, Germany, Austria, and Australia prior to expanding its targeting posture to include U.S.- and Canada-based financial institutions in October 2018," CrowdStrike said. "The malware's popularity grew due to its early modular development supporting Zeus-based web injects, information stealer capabilities, keystroke logging, screen recording, and hidden virtual network computing (HVNC) functionality." According to Black Lotus Labs and Team Cymru, DanaBot employs a layered communications infrastructure between a victim and the botnet controllers, wherein the C2 traffic is proxied through two or three server tiers before it reaches the final level. At least five to six tier-2 servers were active at any given time. A majority of DanaBot victims are concentrated around Brazil, Mexico, and the United States. "The operators have shown their commitment to their craft, adapted to detection and changes in enterprise defense, and with later iterations, insulating the C2s in tiers to obfuscate tracking," the companies said. "Throughout this time, they have made the bot more user-friendly with structured pricing and customer support." High-level diagram of multi-tiered C2 architecture The DoJ said DanaBot administrators operated a second version of the botnet that was specially designed to target victim computers in military, diplomatic, government, and related entities in North America and Europe. This variant, emerging in January 2021, came fitted with capabilities to record all interactions happening on a victim device and send the data to a different server. "Pervasive malware like DanaBot harms hundreds of thousands of victims around the world, including sensitive military, diplomatic, and government entities, and causes many millions of dollars in losses," said United States Attorney Bill Essayli for the Central District of California. The DoJ further credited several private sector firms, Amazon, CrowdStrike, ESET, Flashpoint, Google, Intel 471, Lumen, PayPal, Proofpoint, Spycloud, Team Cymru, and Zscaler, for providing "valuable assistance." Some of the noteworthy aspects of DanaBot, compiled from various reports, are below - DanaBot's sub-botnet 5 received commands to download a Delphi-based executable leveraged to conduct HTTP-based distributed denial-of-service (DDoS) attacks against the Ukrainian Ministry of Defence (MOD) webmail server and the National Security and Defense Council (NSDC) of Ukraine in March 2022, shortly after Russia's invasion of the country Two DanaBot sub-botnets, 24 and 25, were specifically used for espionage purposes likely with an aim to further intelligence-gathering activities on behalf of Russian government interests DanaBot operators have periodically restructured their offering since 2022 to focus on defense evasion, with at least 85 distinct build numbers identified to date (The most recent version is 4006, which was compiled in March 2025) The malware's infrastructure consists of multiple components: A "bot" that infects target systems and performs data collection, an "OnlineServer" that manages the RAT functionalities, a "client" for processing collected logs and bot management, and a "server" that handles bot generation, packing, and C2 communication DanaBot has been used in targeted espionage attacks against government officials in the Middle East and Eastern Europe The authors of DanaBot operate as a single group, offering the malware for rent to potential affiliates, who subsequently use it for their own malicious purposes by establishing and managing their own botnets using private servers DanaBot's developers have partnered with the authors of several malware cryptors and loaders, such as Matanbuchus, and offered special pricing for distribution bundles DanaBot maintained an average of 150 active tier-1 C2 servers per day, with approximately 1,000 daily victims across more than 40 countries, making it one of the largest MaaS platforms active in 2025 Proofpoint, which first identified and named DanaBot in May 2018, said the disruption of the MaaS operation is a win for defenders and that it will have an impact on the cybercriminal threat landscape. "Cybercriminal disruptions and law enforcement actions not only impair malware functionality and use but also impose a cost to threat actors by forcing them to change their tactics, cause mistrust in the criminal ecosystem, and potentially make criminals think about finding a different career," Selena Larson, a staff threat researcher at Proofpoint, said. "These successes against cyber criminals only come about when business IT teams and security service providers share much-needed insight into the biggest threats to society, affecting the greatest number of people around the world, which law enforcement can use to track down the servers, infrastructure, and criminal organizations behind the attacks. Private and public sector collaboration is crucial to knowing how actors operate and taking action against them." DanaBot's features as promoted on its support site DoJ Unseals Charges Against QakBot Leader The development comes as the DoJ unsealed charges against a 48-year-old Moscow resident, Rustam Rafailevich Gallyamo, for leading efforts to develop and maintain the QakBot malware, which was disrupted in a multinational operation in August 2023. The agency also filed a civil forfeiture complaint against over $24 million in cryptocurrency seized from Gallyamov over the course of the investigation. "Gallyamov developed, deployed, and controlled the Qakbot malware beginning in 2008," the DoJ said. "From 2019 onward, Gallyamov allegedly used the Qakbot malware to infect thousands of victim computers around the world in order to establish a network, or 'botnet,' of infected computers." The DoJ revealed that, following the takedown, Gallyamov and his co-conspirators continued their criminal activities by switching to other tactics like "spam bomb" attacks in order to gain unauthorized access to victim networks and deploy ransomware families like Black Basta and CACTUS. Court documents accuse the e-crime group of engaging in these methods as recently as January 2025. "Mr. Gallyamov's bot network was crippled by the talented men and women of the FBI and our international partners in 2023, but he brazenly continued to deploy alternative methods to make his malware available to criminal cyber gangs conducting ransomware attacks against innocent victims globally," said Assistant Director in Charge Akil Davis of the FBI's Los Angeles Field Office. Found this article interesting? Follow us on Twitter  and LinkedIn to read more exclusive content we post.

The effectiveness of language models relies on their ability to simulate human-like step-by-step deduction. However, these reasoning sequences are resource-intensive and can be wasteful for simple questions that do not require elaborate computation. This lack of awareness regarding the complexity of the task is one of the core challenges in these models. They often default to detailed reasoning even for queries that could be answered directly. Such an approach increases token usage, extends response time, and increases system latency and memory usage. As a result, there’s a pressing need to equip language models with a mechanism that allows them to make autonomous decisions about whether to think deeply or respond succinctly. Current tools attempting to solve this issue either rely on manually set heuristics or prompt engineering to switch between short and long responses. Some methods use separate models and route questions based on complexity estimates. Still, these external routing systems often lack insight into the target model’s strengths and fail to make optimal decisions. Other techniques fine-tune models with prompt-based cues like “reasoning on/off,” but these rely on static rules rather than dynamic understanding. Despite some improvements, these approaches fail to enable fully autonomous and context-sensitive control within a single model. Researchers from the National University of Singapore introduced a new framework called Thinkless, which equips a language model with the ability to dynamically decide between using short or long-form reasoning. The framework is built on reinforcement learning and introduces two special control tokens—<short> for concise answers and <think> for detailed responses. By incorporating a novel algorithm called Decoupled Group Relative Policy Optimization (DeGRPO), Thinkless separates the training focus between selecting the reasoning mode and improving the accuracy of the generated response. This design prevents the model from falling into one-dimensional behavior and enables adaptive reasoning tailored to each query. The methodology involves two stages: warm-up distillation and reinforcement learning. In the distillation phase, Thinkless is trained using outputs from two expert models—one specializing in short responses and the other in detailed reasoning. This stage helps the model establish a firm link between the control token and the desired reasoning format. The reinforcement learning stage then fine-tunes the model’s ability to decide which reasoning mode to use. DeGRPO decomposes the learning into two separate objectives: one for training the control token and another for refining the response tokens. This approach avoids the gradient imbalances in earlier models, where longer responses would overpower the learning signal, leading to a collapse in reasoning diversity. Thinkless ensures that both <short> and <think> tokens receive balanced updates, promoting stable learning across response types. When evaluated, Thinkless significantly reduced long-form reasoning while preserving high accuracy. On the Minerva Algebra benchmark, the model used the <think> token in only 25.88% of cases while achieving 94.59% accuracy. In contrast, conventional reasoning models had to use extended chains of thought much more frequently. On the AIME 2024 dataset, Thinkless reached a 27.33% accuracy rate with 100% usage of the reasoning mode, showing that it could maintain performance when full reasoning was necessary. On the GSM8K dataset, it utilized <think> only 13.31% of the time, yet still achieved 84.18% accuracy. These results reflect the model’s ability to handle simple and complex queries with appropriate reasoning depth, cutting down on unnecessary token generation by as much as 90% in some tasks. Overall, this study from the National University of Singapore researchers presents a compelling solution to the inefficiencies of uniform reasoning in large language models. By introducing a mechanism that enables models to judge task complexity and adjust their inference strategy accordingly, Thinkless optimizes both accuracy and efficiency. The method balances depth of reasoning and response precision without relying on fixed rules, offering a data-driven approach to more intelligent language model behavior. Check out the Paper and GitHub Page. All credit for this research goes to the researchers of this project. Also, feel free to follow us on Twitter and don’t forget to join our 95k+ ML SubReddit and Subscribe to our Newsletter. NikhilNikhil is an intern consultant at Marktechpost. He is pursuing an integrated dual degree in Materials at the Indian Institute of Technology, Kharagpur. Nikhil is an AI/ML enthusiast who is always researching applications in fields like biomaterials and biomedical science. With a strong background in Material Science, he is exploring new advancements and creating opportunities to contribute.Nikhilhttps://www.marktechpost.com/author/nikhil0980/This AI Paper Introduces MathCoder-VL and FigCodifier: Advancing Multimodal Mathematical Reasoning with Vision-to-Code AlignmentNikhilhttps://www.marktechpost.com/author/nikhil0980/This AI Paper Introduces PARSCALE (Parallel Scaling): A Parallel Computation Method for Efficient and Scalable Language Model DeploymentNikhilhttps://www.marktechpost.com/author/nikhil0980/Google AI Releases Standalone NotebookLM Mobile App with Offline Audio and Seamless Source IntegrationNikhilhttps://www.marktechpost.com/author/nikhil0980/This AI Paper from Microsoft Introduces a DiskANN-Integrated System: A Cost-Effective and Low-Latency Vector Search Using Azure Cosmos DB

Remember the Vector? A tiny tractor-shaped anthropomorphized robot that sat on your desk, held up snacks, and kept you company? If you do, you also remember its unceremonious demise. Anki folded due to bankruptcy, selling all its assets to another company that pretty much killed the Vector’s functionality. It’s sort of like the Humane AI pin, which ended up getting pretty much shuttered after the company declared it was shutting down. There are now thousands of Anki Vector toys just gathering dust, sitting in the shadow of what could have been. However, a similar product may be fulfilling the Vector’s role, with the exact same demeanor and cuteness. Meet the Focus Buddy, a tiny robot designed to look like an anthropomorphic computer, sitting on your desk. Like the Vector, the Focus Buddy interacts with you, through input and expression – however, it has a more distinct purpose than simply entertainment – its purpose is to boost human focus. Designer: Neurofocus.ai Focus Buddy is a tiny touch-enabled device that uses a combination of sensors to check your attentiveness. Think a pomodoro timer but with AI chops. It’s playful, but plays a pretty fundamental role in your productivity, checking to see if you’re attentive, distracted, or just dissociating (mea culpa). The tiny device, no larger than a chonky webcam, sits on your tabletop, with a 2-ish inch screen that usually shows a friendly face with blinking eyes similar to the Anki Vector. Tap the screen, and you can access a variety of settings as well as games that help boost your focus. The way the Focus Buddy works is simple yet advanced. A webcam on top captures your face, reading your gaze, how many times you blink, and whether you look attentive, sleepy, or just listless. However, a millimeter-wave sensor inside the robot also manages to read your heart rate, combining the two metrics to judge your focus. If it detects a drop, it alerts you, and encourages you to either get back to work, or boost your focus using one of its on-board focus-enhancing exercises. The device is currently in a very rudimentary phase (the company is purely China-based and hasn’t even built its website yet), with its makers looking to slowly reach critical mass. For now, however, the Focus Buddy is a perfect symbol of what the Vector could have been – a buddy that guides you through the day and enables you to be your best self. Also, it’s the perfect size for cat-hats, so you can customize your Focus Buddy with a rodeo had, a tiara, probably even a modded 3D printed Batman or Iron Man mask that works without blocking the screen or impeding the webcam! The post The Anki Vector Died And Gave Birth To The Focus Buddy: Hands-on at BEYOND Expo 2025 first appeared on Yanko Design.

Transcript [MUSIC] GRETCHEN HUIZINGA: Welcome to Abstracts, a Microsoft Research Podcast that puts the spotlight on world-class research in brief. I’m Gretchen Huizinga. In this series, members of the research community at Microsoft give us a quick snapshot – or a podcast abstract – of their new and noteworthy papers.  [MUSIC FADES] On today’s episode, I’m talking to Alex Lu, a senior researcher at Microsoft Research and co-author of a paper called Assessing the Limits of Zero Shot Foundation Models in Single-cell Biology. Alex Lu, wonderful to have you on the podcast. Welcome to Abstracts!  ALEX LU: Yeah, I’m really excited to be joining you today.  HUIZINGA: So let’s start with a little background of your work. In just a few sentences, tell us about your study and more importantly, why it matters.  LU: Absolutely. And before I dive in, I want to give a shout out to the MSR research intern who actually did this work. This was led by Kasia Kedzierska, who interned with us two summers ago in 2023, and she’s the lead author on the study. But basically, in this research, we study single-cell foundation models, which have really recently rocked the world of biology, because they basically claim to be able to use AI to unlock understanding about single-cell biology. Biologists for a myriad of applications, everything from understanding how single cells differentiate into different kinds of cells, to discovering new drugs for cancer, will conduct experiments where they measure how much of every gene is expressed inside of just one single cell. So these experiments give us a powerful view into the cell’s internal state. But measurements from these experiments are incredibly complex. There are about 20,000 different human genes. So you get this really long chain of numbers that measure how much there is of 20,000 different genes. So deriving meaning from this really long chain of numbers is really difficult. And single-cell foundation models claim to be capable of unraveling deeper insights than ever before. So that’s the claim that these works have made. And in our recent paper, we showed that these models may actually not live up to these claims. Basically, we showed that single-cell foundation models perform worse in settings that are fundamental to biological discovery than much simpler machine learning and statistical methods that were used in the field before single-cell foundation models emerged and are the go-to standard for unpacking meaning from these complicated experiments. So in a nutshell, we should care about these results because it has implications on the toolkits that biologists use to understand their experiments. Our work suggests that single-cell foundation models may not be appropriate for practical use just yet, at least in the discovery applications that we cover.  HUIZINGA: Well, let’s go a little deeper there. Generative pre-trained transformer models, GPTs, are relatively new on the research scene in terms of how they’re being used in novel applications, which is what you’re interested in, like single-cell biology. So I’m curious, just sort of as a foundation, what other research has already been done in this area, and how does this study illuminate or build on it?  LU: Absolutely. Okay, so we were the first to notice and document this issue in single-cell foundation models, specifically. And this is because that we have proposed evaluation methods that, while are common in other areas of AI, have yet to be commonly used to evaluate single-cell foundation models. We performed something called zero-shot evaluation on these models. Prior to our work, most works evaluated single-cell foundation models with fine tuning. And the way to understand this is because single-cell foundation models are trained in a way that tries to expose these models to millions of single-cells. But because you’re exposing them to a large amount of data, you can’t really rely upon this data being annotated or like labeled in any particular fashion then. So in order for them to actually do the specialized tasks that are useful for biologists, you typically have to add on a second training phase. We call this the fine-tuning phase, where you have a smaller number of single cells, but now they are actually labeled with the specialized tasks that you want the model to perform. So most people, they typically evaluate the performance of single-cell models after they fine-tune these models. However, what we noticed is that this evaluating these fine-tuned models has several problems. First, it might not actually align with how these models are actually going to be used by biologists then. A critical distinction in biology is that we’re not just trying to interact with an agent that has access to knowledge through its pre-training, we’re trying to extend these models to discover new biology beyond the sphere of influence then. And so in many cases, the point of using these models, the point of analysis, is to explore the data with the goal of potentially discovering something new about the single cell that the biologists worked with that they weren’t aware of before. So in these kinds of cases, it is really tough to fine-tune a model. There’s a bit of a chicken and egg problem going on. If you don’t know, for example, there’s a new kind of cell in the data, you can’t really instruct the model to help us identify these kinds of new cells. So in other words, fine-tuning these models for those tasks essentially becomes impossible then. So the second issue is that evaluations on fine-tuned models can sometimes mislead us in our ability to understand how these models are working. So for example, the claim behind single-cell foundation model papers is that these models learn a foundation of biological knowledge by being exposed to millions of single cells in its first training phase, right? But it’s possible when you fine-tune a model, it may just be that any performance increases that you see using the model is simply because that you’re using a massive model that is really sophisticated, really large. And even if there’s any exposure to any cells at all then, that model is going to do perfectly fine then. So going back to our paper, what’s really different about this paper is that we propose zero-shot evaluation for these models. What that means is that we do not fine-tune the model at all, and instead we keep the model frozen during the analysis step. So how we specialize it to be a downstream task instead is that we extract the model’s internal embedding of single-cell data, which is essentially a numerical vector that contains information that the model is extracting and organizing from input data. So it’s essentially how the model perceives single-cell data and how it’s organizing in its own internal state. So basically, this is the better way for us to test the claim that single-cell foundation models are learning foundational biological insights. Because if they actually are learning these insights, they should be present in the models embedding space even before we fine-tune the model.  HUIZINGA: Well, let’s talk about methodology on this particular study. You focused on assessing existing models in zero-shot learning for single-cell biology. How did you go about evaluating these models?  LU: Yes, so let’s dive deeper into how zero-shot evaluations are conducted, okay? So the premise here is that we’re relying upon the fact that if these models are fully learning foundational biological insights, if we take the model’s internal representation of cells, then cells that are biologically similar should be close in that internal representation, where cells that are biologically distinct should be further apart. And that is exactly what we tested in our study. We compared two popular single-cell foundation models and importantly, we compared these models against older and reliable tools that biologists have used for exploratory analyses. So these include simpler machine learning methods like scVI, statistical algorithms like Harmony, and even basic data pre-processing steps, just like filtering your data down to a more robust subset of genes, then. So basically, we tested embeddings from our two single-cell foundation models against this baseline in a variety of settings. And we tested the hypothesis that biologically similar cells should be similar across these distinct methods across these datasets.  HUIZINGA: Well, and as you as you did the testing, you obviously were aiming towards research findings, which is my favorite part of a research paper, so tell us what you did find and what you feel the most important takeaways of this paper are.  LU: Absolutely. So in a nutshell, we found that these two newly proposed single-cell foundation models substantially underperformed compared to older methods then. So to contextualize why that is such a surprising result, there is a lot of hype around these methods. So basically, I think that,yeah, it’s a very surprising result, given how hyped these models are and how people were already adopting them. But our results basically caution that these shouldn’t really be adopted for these use purposes.  HUIZINGA: Yeah, so this is serious real-world impact here in terms of if models are being adopted and adapted in these applications, how reliable are they, et cetera? So given that, who would you say benefits most from what you’ve discovered in this paper and why?  LU: Okay, so two ways, right? So I think this has at least immediate implications on the way that we do discovery in biology. And as I’ve discussed, these experiments are used for cases that have practical impact, drug discovery applications, investigations into basic biology, then. But let’s also talk about the impact for methodologists, people who are trying to improve these single-cell foundation models, right? I think at the base, they’re really excited proposals. Because if you look at what some of the prior and less sophisticated methods couldn’t do, they tended to be more bespoke. So the excitement of single-cell foundation models is that you have this general-purpose model that can be used for everything and while they’re not living up to that purpose just now, just currently, I think that it’s important that we continue to bank onto that vision, right? So if you look at our contributions in that area, where single-cell foundation models are a really new proposal, so it makes sense that we may not know how to fully evaluate them just yet then. So you can view our work as basically being a step towards more rigorous evaluation of these models. Now that we did this experiment, I think the methodologists know to use this as a signal on how to improve the models and if they’re going in the right direction. And in fact, you are seeing more and more papers adopt zero-shot evaluations since we put out our paper then. And so this essentially helps future computer scientists that are working on single-cell foundation models know how to train better models.  HUIZINGA: That said, Alex, finally, what are the outstanding challenges that you identified for zero-shot learning research in biology, and what foundation might this paper lay for future research agendas in the field?  LU: Yeah, absolutely. So now that we’ve shown single-cell foundation models don’t necessarily perform well, I think the natural question on everyone’s mind is how do we actually train single-cell foundation models that live up to that vision, that can perform in helping us discover new biology then? So I think in the short term, yeah, we’re actively investigating many hypotheses in this area. So for example, my colleagues, Lorin Crawford and Ava Amini, who were co-authors in the paper, recently put out a pre-print understanding how training data composition impacts model performance. And so one of the surprising findings that they had was that many of the training data sets that people used to train single-cell foundation models are highly redundant, to the point that you can even sample just a tiny fraction of the data and get basically the same performance then. But you can also look forward to many other explorations in this area as we continue to develop this research at the end of the day. But also zooming out into the bigger picture, I think one major takeaway from this paper is that developing AI methods for biology requires thought about the context of use, right? I mean, this is obvious for any AI method then, but I think people have gotten just too used to taking methods that work out there for natural vision or natural language maybe in the consumer domain and then extrapolating these methods to biology and expecting that they will work in the same way then, right? So for example, one reason why zero-shot evaluation was not routine practice for single-cell foundation models prior to our work, I mean, we were the first to fully establish that as a practice for the field, was because I think people who have been working in AI for biology have been looking to these more mainstream AI domains to shape their work then. And so with single-cell foundation models, many of these models are adopted from large language models with natural language processing, recycling the exact same architecture, the exact same code, basically just recycling practices in that field then. So when you look at like practices in like more mainstream domains, zero-shot evaluation is definitely explored in those domains, but it’s more of like a niche instead of being considered central to model understanding. So again, because biology is different from mainstream language processing, it’s a scientific discipline, zero-shot evaluation becomes much more important, and you have no choice but to use these models, zero-shot then. So in other words, I think that we need to be thinking carefully about what it is that makes training a model for biology different from training a model, for example, for consumer purposes.  [MUSIC] HUIZINGA: Alex Lu, thanks for joining us today, and to our listeners, thanks for tuning in. If you want to read this paper, you can find a link at aka.ms/Abstracts, or you can read it on the Genome Biology website. See you next time on Abstracts!  [MUSIC FADES] 

May 22, 2025Ravie LakshmananMalware / Cybercrime A sprawling operation undertaken by global law enforcement agencies and a consortium of private sector firms has disrupted the online infrastructure associated with a commodity information stealer known as Lumma (aka LummaC or LummaC2), seizing 2,300 domains that acted as the command-and-control (C2) backbone to commandeer infected Windows systems. "Malware like LummaC2 is deployed to steal sensitive information such as user login credentials from millions of victims in order to facilitate a host of crimes, including fraudulent bank transfers and cryptocurrency theft," the U.S. Department of Justice (DoJ) said in a statement. The confiscated infrastructure has been used to target millions across the world through affiliates and other cyber criminals. Lumma Stealer, active since late 2022, is estimated to have been used in at least 1.7 million instances to steal information, such as browser data, autofill information, login credentials, and cryptocurrency seed phrases. The U.S. Federal Bureau of Investigation (FBI) has attributed around 10 million infections to Lumma. The seizure impacts five domains that serve as login panels for Lumma Stealer's administrators and paying customers to deploy the malware, thereby preventing them from compromising the computers and stealing victim information. "Between March 16 and May 16, 2025, Microsoft identified over 394,000 Windows computers globally infected by the Lumma malware," Europol said, adding the operation cuts off communications between the malicious tool and victims. The agency described Lumma as the "world's most significant infostealer threat." Microsoft's Digital Crimes Unit (DCU), in partnership with other cybersecurity companies ESET, BitSight, Lumen, Cloudflare, CleanDNS, and GMO Registry, said it took down approximately 2,300 malicious domains that formed the backbone of Lumma's infrastructure. Spread of Lumma Stealer malware infections across Windows devices "The primary developer of Lumma is based in Russia and goes by the internet alias 'Shamel,'" Steven Masada, assistant general counsel at DCU, said. "Shamel markets different tiers of service for Lumma via Telegram and other Russian-language chat forums. Depending on what service a cybercriminal purchases, they can create their own versions of the malware, add tools to conceal and distribute it, and track stolen information through an online portal." The stealer, marketed under a malware-as-a-service (MaaS) model, is available on a subscription basis for anywhere between $250 to $1,000. The developer also offers a $20,000 plan that grants customers access to source code and the right to sell it to other criminal actors. Weekly counts of new C2 domains "Lower tiers include basic filtering and log download options, while higher tiers offer custom data collection, evasion tools, and early access to new features," ESET said. "The most expensive plan emphasizes stealth and adaptability, offering unique build generation and reduced detection." Over the years, Lumma has become something of a notorious threat, being delivered via various distribution vectors, including the increasingly popular ClickFix method. The Windows maker, which is tracking the threat actor behind the stealer under the name Storm-2477, said its distribution infrastructure is both "dynamic and resilient," leveraging a combination of phishing, malvertising, drive-by download schemes, abuse of trusted platforms, and traffic distribution systems like Prometheus. Lumma C2 selection mechanism Cato Networks, in a report published Wednesday, revealed that suspected Russian threat actors are leveraging Tigris Object Storage, Oracle Cloud Infrastructure (OCI) Object Storage, and Scaleway Object Storage to host fake reCAPTCHA pages that make use of ClickFix-style lures to trick users into downloading Lumma Stealer. "The recent campaign leveraging Tigris Object Storage, OCI Object Storage, and Scaleway Object Storage builds upon earlier methods, introducing new delivery mechanisms aimed at evading detection and targeting technically proficient users," researchers Guile Domingo, Guy Waizel, and Tomer Agayev said. Attack flow for ClickFix leading to Lumma Stealer using Prometheus TDS Some of the notable aspects of the malware are below - It employs a multi-tiered C2 infrastructure consisting of a set of nine frequently changing tier-1 domains hard-coded into the malware's configuration and fallback C2s hosted on Steam profiles and Telegram channels that point to tier-1 C2s The payloads are typically spread using pay-per-install (PPI) networks or traffic sellers that deliver installs-as-a-service. The stealer is typically bundled with spoofed software or cracked versions of popular commercial software, targeting users looking to avoid paying for legitimate licenses The operators have created a Telegram marketplace with a rating system for affiliates to sell stolen data without intermediaries The core binary is obfuscated with advanced protection such as low-level virtual machine (LLVM core), Control Flow Flattening (CFF), Control Flow Obfuscation, customized stack decryption, huge stack variables, and dead codes, among others to make static analysis difficult There were more than 21,000 market listings selling Lumma Stealer logs on multiple cybercriminal forums from April through June of 2024, a 71.7% increase from April through June of 2023 "The Lumma Stealer distribution infrastructure is flexible and adaptable," Microsoft said. "Operators continually refine their techniques, rotating malicious domains, exploiting ad networks, and leveraging legitimate cloud services to evade detection and maintain operational continuity. To further hide the real C2 servers, all the C2 servers are hidden behind the Cloudflare proxy." "This dynamic structure enables operators to maximize the success of campaigns while complicating efforts to trace or dismantle their activities. The growth and resilience of Lumma Stealer highlights the broader evolution of cybercrime and underscores the need for layered defenses and industry collaboration to counter threats." In an interview with security researcher g0njxa in January 2025, the developer behind Lumma said they intended to cease operations by next fall. "We have done a lot of work over two years to achieve what we have now," they said. "We are proud of this. It has become a part of our daily life for us, and not just work." Found this article interesting? Follow us on Twitter  and LinkedIn to read more exclusive content we post. SHARE    

From The Heretic’s Gawain to Louise in Enemies, our Demo team continues to create real-time cinematics that push the boundaries of Unity’s capabilities for high-fidelity productions, with a special focus on digital humans.The pursuit to create ever more realistic digital characters is endless. And since the launch of Enemies at GDC 2022, we have continued our research and development into solutions for better and more believable digital human creation, in collaboration with Unity’s Graphics Engineering team and commercially available service providers specializing in that area.At SIGGRAPH 2022, we announced our next step: replacing the heavy 4D data playback of the protagonist’s performance with a lightweight Ziva puppet. This recent iteration sees the integration of Ziva animation technology with the latest in Unity’s graphics advancements, including the High Definition Render Pipeline (HDRP) – all with the aim of further developing an end-to-end pipeline for character asset creation, animation, and authoring.Along with the launch of a new strand-based Hair Solution and updated Digital Human package, the Enemies real-time demo is now available to download. You can run it in real-time and experience it for yourself, just as it was shown at Unite 2022.While the cinematic may not appear too different from the original, its final rendered version shows how the integration of Ziva technology has brought a new dimension to our protagonist.Ziva brings decades of experience and pioneering research from the VFX industry to enable greater animation quality for games, linear content production, and real-time projects. Its machine learning (ML)-based technology helps achieve extraordinary realism in facial animation, and also for body and muscle deformations.To achieve the level of realism in Enemies, Ziva used machine learning and 4D data capture, which goes beyond the traditional process of scanning actors in 3D scans. The static, uneditable 4D captured facial performance has now been transformed into a real-time puppet with a facial rig that can be animated and adjusted at any time – all while maintaining high fidelity.Our team built on that 4D capture data and trained a machine-learned model that could be animated to create any performance. The end result is a 50 MB facial rig that has all the detail of the 4D captured performance, without having to carry its original 3.7 GB of weight.This technology means that you can replicate the results with a fraction of the animation data, creating real-time results in a way that 4D does not typically allow.In order to achieve this, Unity’s Demo team focused on:Creating the puppetTo create this new version of Louise, we worked with the Ziva team. They handled the machine learning workflow using a preexisting 4D data library. Additional 4D data was collected from a new performance by the original Enemies actor (we only needed to collect a few additional expressions). This is one of the unique advantages of our machine learning approach.With this combined dataset, we trained a Ziva puppet to accurately reproduce the original performance. We could alter this performance in any way, ranging from tweaking minute details to changing the entire expression.Using the 4D data capture through machine learning, we could enable any future performance to run on any 3D head by showing a single performance applied to multiple faces of varying proportions. This makes it easier to expand the range of performances to multiple actors and real-time digital humans for any future editions.The puppet’s control schemeOnce the machine learning was completed, we had 200–300 parameters that, when used in combination and at different weights, could recreate everything we had seen in the 4D data with incredible accuracy. We didn’t have to worry about a hand-animated performance looking different when used by a group of different animators. The persona and idiosyncrasies of the original actor would come through no matter how we chose to animate the face.As Ziva is based on deformations and not an underlying facial rig, we could manipulate even the smallest detail because the trained face uses a control scheme that was developed to take advantage of the fidelity of the machine-learned parameters/data.At this point, creating a rig is a relatively flexible process as we can just tap into those machine-learned parameters – this, in turn, deforms the face. There are no joints in a Ziva puppet, besides the basic logical face and neck joints.So what does this all mean?There are many advantages to this new workflow. First and foremost, we now have the ability to dynamically interact with the performance of the digital human in Enemies.This allows us to change the character’s performance after it has already been delivered. Digital Louise can now say the same lines as before, but with very different facial expressions. For example, she can be friendlier or angrier or convey any other emotion that the director envisions.We are also able to manually author new performances with the puppet – facial expressions and reactions that the original actress never performed. If we wanted to develop the story into an interactive experience, it would be important to expand the possibility of what the digital character reacts to, such as a player’s chess moves, with nuances of approval or disapproval.For the highest level of fidelity, the Ziva team can even create a new puppet with its own 4D dataset. Ziva also recently released a beta version of Face Trainer, a product built on a comprehensive library of 4D data and ML algorithms. It can be used to train any face mesh to perform the most complex expressions in real-time without any new 4D capture.Additionally, it is possible to create new lines of dialogue, all at a fraction of the time and cost that the creation of the first line required. We can do this either by getting the original actress to perform additional lines with an HMC and then using the HMC data to drive the puppet, or by getting another performer to deliver the new lines and retargeting their HMC data to the existing puppet.At SIGGRAPH Real-Time Live! we demonstrated how to apply the original performance from Enemies to the puppet of another actress – ultimately replacing the protagonist of the story with a different person, without changing anything else.This performance was then shown at Unite 2022 during the keynote (segment 01:03:00), where Enemies ran on an Xbox Series X, with DX12 and real-time ray tracing.To further enhance the visual quality of Enemies, a number of HDRP systems were leveraged. These include Shader Graph motion vectors, Adaptive Probe Volumes (APV), and of course, hair shading.Enemies also makes use of real-time ray tracing in HDRP and Unity’s native support for NVIDIA DLSS 2.0 (Deep Learning Super Sampling), which enable it to run at 4K image quality, comparable to native resolution. All of these updated Unity features are now available in Unity 2022 LTS.The brand new strand-based Hair Solution, developed during the creation of the Enemies demo, can simulate individual hairs in real-time. This technology is now available as an experimental package via GitHub (requires Unity 2020.2.0f1 or newer), along with a tutorial to get started.By integrating a complete pipeline for authoring, simulation, shading, and rendering hair in Unity, this solution is applicable to digital humans and creatures, in both realistic and stylized projects. The development work continues with a more performant solution for hair rendering enabled by the upcoming Software Rasterizer in HDRP. We are also diversifying the authoring options available by adopting and integrating the Wētā Wig tool for more complex grooms, as showcased in the Lion demo.Expanding on the technological innovations from The Heretic, the updated Digital Human package provides a realistic shading model for the characters rendered in Unity.Such updates include:A better 4D pipelineA more performant Skin Attachment system on the GPU for high-density meshesMore realistic eyes with caustics on the iris (available in HDRP as of Unity 2022.2)A new skin shader, built with the available Editor technologyTension tech for blood flow simulation and wrinkle maps, eliminating the need for a facial rigAnd as always, there is more to come.Discover how Ziva can help bring your next project to life. Register your interest to receive updates or get early access to future Ziva beta programs. If you’d like to learn more, you can contact us here.

Multimodal mathematical reasoning enables machines to solve problems involving textual information and visual components like diagrams and figures. This requires combining language understanding and visual interpretation to make sense of complex mathematical contexts. Such capabilities are vital in education, automated tutoring, and document analysis, where problems are often presented with a blend of text and images. A major obstacle in this area is the lack of high-quality, precise alignment between math images and their textual or symbolic representations. Most datasets used to train large multimodal models are derived from image captions in natural settings, which often miss the detailed elements essential for mathematical accuracy. This creates problems for models that rely on these data sources, making them unreliable when dealing with geometry, figures, or technical diagrams. A model’s performance in mathematical reasoning depends heavily on its ability to correctly interpret and link these visual details with mathematical expressions or instructions. In the past, some approaches tried to address this by either enhancing the visual encoders or using manually crafted datasets. However, these methods tend to produce low image diversity, relying on hand-coded or template-based generation, which limits their applicability. Some efforts, like Math-LLaVA and MAVIS, developed synthetic datasets and used templates or predefined categories. Still, they could not dynamically create a wide variety of math-related visuals. This shortfall restricts the learning scope of models and leaves them struggling with more complex or less structured mathematical problems. Researchers from the Multimedia Laboratory at The Chinese University of Hong Kong and CPII under InnoHK introduced a novel approach called MathCoder-VL. This method combines a vision-to-code model named FigCodifier and a synthetic data engine. They constructed the ImgCode-8.6M dataset using a model-in-the-loop strategy, which allowed them to build the largest image-code dataset to date iteratively. Further, they developed MM-MathInstruct-3M, a multimodal instruction dataset enriched with newly synthesized images. The MathCoder-VL model is trained in two stages: mid-training on ImgCode-8.6M to improve visual-text alignment and fine-tuning on MM-MathInstruct-3M to strengthen reasoning abilities. The FigCodifier model works by translating mathematical figures into code that can recreate those figures exactly. This code-image pairing ensures strict alignment and accuracy, unlike caption-based datasets. The process begins with 119K image-code pairs from DaTikZ and expands through iterative training using images collected from textbooks, K12 datasets, and arXiv papers. The final dataset includes 8.6 million code-image pairs and covers various mathematical topics. FigCodifier also supports Python-based rendering, which adds variety to image generation. The system filters low-quality data by checking code validity and removing redundant or unhelpful visuals, resulting in 4.3M high-quality TikZ and 4.3M Python-based pairs. Performance evaluations show that MathCoder-VL outperforms multiple open-source models. The 8B version achieved 73.6% accuracy on the MathVista Geometry Problem Solving subset, surpassing GPT-4o and Claude 3.5 Sonnet by 8.9% and 9.2%, respectively. It also scored 26.1% on MATH-Vision and 46.5% on MathVerse. In Chinese-language benchmarks, it achieved 51.2% on GAOKAO-MM. On the We-Math benchmark, it solved two-step problems at 58.6%, outperforming GPT-4o’s 58.1%. Its performance on three-step problems reached 52.1%, again exceeding GPT-4o’s 43.6%. Compared to its base model InternVL2-8B, it showed gains of 6.1% on MATH-Vision and 11.6% on MathVista. This work clearly defines the problem of insufficient visual-textual alignment in multimodal math reasoning and provides a scalable and innovative solution. The introduction of FigCodifier and synthetic datasets allows models to learn from accurate, diverse visuals paired with exact code, significantly boosting their reasoning abilities. MathCoder-VL represents a practical advancement in this field, demonstrating how thoughtful model design and high-quality data can overcome longstanding limitations in mathematical AI. Check out the Paper and GitHub Page. All credit for this research goes to the researchers of this project. Also, feel free to follow us on Twitter and don’t forget to join our 95k+ ML SubReddit and Subscribe to our Newsletter. NikhilNikhil is an intern consultant at Marktechpost. He is pursuing an integrated dual degree in Materials at the Indian Institute of Technology, Kharagpur. Nikhil is an AI/ML enthusiast who is always researching applications in fields like biomaterials and biomedical science. With a strong background in Material Science, he is exploring new advancements and creating opportunities to contribute.Nikhilhttps://www.marktechpost.com/author/nikhil0980/This AI Paper Introduces PARSCALE (Parallel Scaling): A Parallel Computation Method for Efficient and Scalable Language Model DeploymentNikhilhttps://www.marktechpost.com/author/nikhil0980/Google AI Releases Standalone NotebookLM Mobile App with Offline Audio and Seamless Source IntegrationNikhilhttps://www.marktechpost.com/author/nikhil0980/This AI Paper from Microsoft Introduces a DiskANN-Integrated System: A Cost-Effective and Low-Latency Vector Search Using Azure Cosmos DBNikhilhttps://www.marktechpost.com/author/nikhil0980/LLMs Struggle to Act on What They Know: Google DeepMind Researchers Use Reinforcement Learning Fine-Tuning to Bridge the Knowing-Doing Gap