arXiv:2605.08156v1 Announce Type: new Abstract: Zero-shot recognition aims to classify an image by selecting the most compatible label description from a set of candidate classes without any task-specific supervision. In fine-grained settings, however, the relevant evidence often lies in localized parts, attributes, or textures rather than in the full image, making whole-image alignment suboptimal. Recent localized visual-text alignment methods address this by comparing class descriptions with multiple image regions, but they typically rely on large sets of random or redundant crops, increasing inference cost and introducing many highly redundant or weakly relevant candidates. Moreover, introducing semantic guidance too early can create an error-amplifying feedback process in which inaccurate intermediate predictions bias later localization and reinforce subsequent mistakes; we refer to this failure mode as the prediction loop. We propose LAGO (LAnguage-Guided adaptive Object-region focus), a framework for efficient and robust zero-shot localized visual-text alignment. LAGO first performs class-agnostic object-centric candidate discovery to obtain a stable visual initialization, and then applies adaptive language-guided refinement with the strength of semantic guidance controlled by intermediate confidence. It further combines object-level, contextual, and full-image evidence through an effective object-context dual-channel aggregation strategy. Extensive experiments show that LAGO consistently achieves state-of-the-art performance on standard zero-shot benchmarks and challenging distribution-shift settings, while requiring substantially fewer candidate regions at inference time.

arXiv:2605.08348v1 Announce Type: new Abstract: The circuits framework in mechanistic interpretability aims to identify causally important sparse subgraphs of model components, typically evaluated by measuring necessity and sufficiency. We measure circuit reuse, the proportion of components shared across per-example circuits within a task, and investigate two less-studied properties of this: consistency, the recurrence of components within a task, and specificity, their uniqueness to a task. Using edge attribution patching across six tasks and seven models, we find that within-task reuse is high and that shared components are necessary for task performance, with ablations causing up to $\sim$100% relative accuracy drops. However, circuits turn out not to be task-specific: ablating one task's circuit damages another task's performance about as much as that task's own circuit does. We discover that this is due to substantial overlap between circuits across tasks, which are causally important for performance. Some circuits do contain a smaller set of task-specific components, but these account for only a modest portion of circuit performance. Overall, our findings suggest that while circuit discovery at the level of attention heads and MLP layers identifies important components, their lack of task-specificity raises questions about the degree to which circuits can support targeted understanding and intervention on model behavior.

arXiv:2605.08133v1 Announce Type: new Abstract: Vision-Language-Action (VLA) models have emerged as a promising paradigm for end-to-end autonomous driving, yet their reliance on implicit parametric knowledge limits generalization in long-tail scenarios. While Retrieval-Augmented Generation (RAG) offers a solution by accessing external expert priors, standard visual retrieval suffers from high latency and semantic ambiguity. To address these challenges, we propose \textbf{VLADriver-RAG}, a framework that grounds planning in explicit, structure-aware historical knowledge. Specifically, we abstract sensory inputs into spatiotemporal semantic graphs via a \textit{Visual-to-Scenario} mechanism, effectively filtering visual noise. To ensure retrieval relevance, we employ a \textit{Scenario-Aligned Embedding Model} that utilizes Graph-DTW metric alignment to prioritize intrinsic topological consistency over superficial visual similarity. These retrieved priors are then fused within a query-based VLA backbone to synthesize precise, disentangled trajectories. Extensive experiments on the Bench2Drive benchmark establish a new state-of-the-art, achieving a Driving Score of 89.12.

arXiv:2605.08161v1 Announce Type: new Abstract: Tumor segmentation in whole-body PET/CT imaging is crucial for precise disease evaluation and treatment planning. However, it remains challenging due to variability in lesion size, contrast, and anatomical distribution. Relying on manual segmentation makes the process time-consuming and prone to intra- and inter-observer variability. This work presents a whole-body tumor segmentation method developed for the AutoPET III challenge, where the goal is to build models that generalize across tracers and multi-center data. We employ the nnU-Net framework with a ResNet-based encoder as our baseline and systematically investigate the impact of training strategies, including intensity normalization, batch dice optimization, and data augmentation using CraveMix. Our experiments show that these strategies significantly influence model performance, particularly in reducing false positives and improving robustness to lesion variability. The best-performing configuration achieves a Dice score of up to 0.80 on the preliminary test phase, and our method ranked third in the AutoPET III challenge. The code is publicly available here.

arXiv:2605.08146v1 Announce Type: new Abstract: Multi-model learning has attracted great attention in visual-text tasks. However, visual-tabular data, which plays a pivotal role in high-stakes domains like healthcare and industry, remains underexplored. In this paper, we introduce \textit{VT-Bench}, the first unified benchmark for standardizing vision-tabular discriminative prediction and generative reasoning tasks. VT-Bench aggregates 14 datasets across 9 domains (medical-centric, while covering pets, media, and transportation) with over 756K samples. We evaluate 23 representative models, including unimodal experts, specialized visual-tabular models, general-purpose vision-language models (VLMs), and tool-augmented methods, highlighting substantial challenges of visual-tabular learning. We believe VT-Bench will stimulate the community to build more powerful multi-modal vision-tabular foundation models. Benchmark: https://github.com/Ziyi-Jia990/VT-Bench

Discussion | Link

Georgia Wells / Wall Street Journal: AppMagic: Grok downloads fell to ~8.3M in April, from a high of 20M+ in January; Recon Analytics says Grok paid adoption in the US remains nearly flat YoY in Q2  —  Adoption by business and consumer users has slowed as parent SpaceX rents out spare computing capacity to rival Anthropic

The Information: Source: new revenue sharing terms cap OpenAI's payments to Microsoft at $38B; they previously could have totaled $135B through 2030 if long-term goals were met  —  OpenAI Chief Financial Officer Sarah Friar has plenty of things to worry about.  But paying OpenAI's early backer Microsoft …

Discussion | Link

Courtney Rozen / Reuters: The US Commerce Department removed from its website details about its May 5 agreement with Google, xAI, and Microsoft to test their AI models  —  The U.S. Commerce Department removed details from its website about its agreement with Google, xAI and Microsoft to test their artificial intelligence models …

Discussion | Link

Discussion | Link

Canadians appear to be making far fewer work trips, new cell phone data show. Bloomberg/Getty Images Donald Trump's trade war has had a large impact on Canadian tourism. New analysis out of the University of Toronto suggests that other areas of the US economy have taken a major hit as well. The findings underscore how the relationship between Canada and the US is under strain. The economic effects of President Donald Trump's disruptive approach to foreign policy are coming into sharper focus with a year's worth of data. As expected, a range of numbers like border crossings and hotel bookings show how Trump's trade war has killed off a large segment of Canadian tourism. But new analysis from the University of Toronto's School of Cities now shows that other areas of the economy have taken a major hit as well. Using cell phone data, researchers Karen Chapple, Yihoi Jung, and Jeff Allen documented an average year-over-year decline in Canadian visits to US cities of approximately 42%. That's a much larger hit than the 25% drop previously estimated based on border crossing data. "The top 20 cities were a number of big metros that aren't exactly known as, you know, big tourist areas," Chapple told Business Insider. A visualization of the drop-off of Canadian tourism to US cities, based on cell phone data analysis. https://mappingtariffs.org/canada-us-visits While the drop-offs in Las Vegas and Orlando could reasonably be attributed to less tourism, major commercial hubs like New York, Los Angeles, San Francisco, Dallas, and Houston also saw significant declines. The researchers attribute the change to fewer tech and finance employees making work trips. "I fly to San Francisco all the time," Chapple said. "It's a tech flight to San Francisco — that's who's with me on the flight." Chapple said her team was particularly surprised by mid-sized cities like Grand Rapids and Flint, Michigan, which have robust automotive industry relationships that are increasingly complicated by the White House's changing tariff policy. "High-tech and financial centres like San Francisco and Houston appear to be experiencing reductions not only in tourists but also in business-related travel, reflecting changing travel preferences due to broader economic uncertainties on both sides of the border," the researchers wrote. "As another example, Grand Rapids, which has close ties to the automotive industry in Ontario, has experienced the second largest drop in visitation, likely due to the tariffs." The findings underscore how the ties between Canada and the US are under strain: Of the 267 cities included in the analysis, only three saw increased Canadian visits last year compared to the year before. It took about three years for Canadian visits to largely recover the plunge that followed the COVID-19 pandemic, and it's still too early to tell how long this latest dip will last. Read the original article on Business Insider

Jeremy Stern / Colossus: Q&A with Scott Wu, CEO of Cognition, which hit a $445M revenue run rate in its first 18 months, on his math competition roots, the Devin AI coding agent, more  —  An afternoon with Cognition founder and nerd emperor Scott Wu … There is a familiar arc to the type of mathematical genius …

Robinhood files confidentially for its second venture fund, this time targeting growth and early-stage startups.

GitLab's CEO announced plans to lay off an unspecified number of employees. Jaque Silva/NurPhoto via Getty Images GitLab stock is down after hours following its plans to lay off staff as it looks to meet the "agentic era." "The planning is happening openly, including a voluntary separation window," GitLab's CEO wrote in a memo. The CEO acknowledged the limbo period "creates real uncertainty for our team over the next few weeks." GitLab employees are the latest group to be sweating it out for a few weeks as they wait to learn whether they dodged the layoff blade. The software company, which had 2,580 employees as of January, announced on Monday that it was restructuring to meet "the agentic era" and planned to cut an unspecified number of its workforce by June 1. Shares in the company fell 7% after hours following the announcement. GitLab CEO Bill Staples wrote in a memo to employees and investors that the restructuring would happen "openly" and entail flattening the org chart — removing "up to three layers of management so leaders are closer to the work." The company's research and development teams will be reorganized into around 60 "smaller, more empowered teams with end-to-end ownership," Staples wrote. AI agents will also be embedded into GitLab's internal processes, he wrote, "automating the reviews, approvals, and handoffs to speed us up," and that means the company will "right-size" roles accordingly. The CEO said that GitLab, which makes a software platform to help coders manage projects and tasks, is evolving its strategy "for the future state of software engineering." "As the cost of producing software collapses, demand for it will expand. Last year, the developer platform market used to be measured in tens of dollars per user per month, this year it is hundreds/user/month and headed to thousands," Staples wrote. Addressing the state of developer industry itself, GitLab's CEO said that "the supply of deep technical problems is multiplying, and the engineers who can solve them will be among the scarcest and most valuable talent in the market." Read GitLab CEO Bill Staples' memo to employees: GitLab Act 2 A letter to our customers and our investors. We've been working through some significant changes inside GitLab over the past few days, and I want to share them with you directly. The email I sent the team is included below for full context. The agentic era affords GitLab the largest opportunity in our history as a company, and we're making the structural and strategic decisions to meet it. This letter has three parts. First, the operational and structural news, which is hard. Second, the strategic thesis we're betting on. And finally, what this means specifically for you, our customers and investors. The structural news This morning we shared with team members that we're beginning a restructuring process at GitLab, and we're running it differently than most. The planning is happening openly, including a voluntary separation window. That creates real uncertainty for our team over the next few weeks, but we believe the outcome will be better for it. Where we can, we plan to finalize the new shape of the company on or before June 1. Where local requirements apply we will not make any changes until the local process is complete. Four operational changes are part of the workforce reduction. 1. We're reevaluating our operational footprint, and are planning to reduce the number of countries by up to 30% where we have small teams. We'll continue serving customers in those markets through our partner network. 2. We're planning to flatten the organization, removing up to three layers of management in some functions so leaders are closer to the work. 3. We're re-organizing R&D to create roughly 60 smaller, more empowered teams with end-to-end ownership, nearly doubling the number of independent teams. 4. We're rewiring internal processes with AI agents, automating the reviews, approvals, and handoffs to speed us up, and plan to right-size roles across the company to follow suit. Operational changes and the update to our strategy are happening together: they are related but independent. Operationally, we grew into a shape that was right for the last era and isn't right for this one. The strategy below is what we're betting on next, and stands on its own. We are reaffirming our Q1 and full year FY27 guidance today. The final scope and financial impact of the restructuring will be shared on our June 2 earnings call, once we've finished the plan and received approval by our board. Our Core Beliefs Underpinning the changes we're making today, and our go forward strategy are 10 core beliefs that span the world we're building for, the architectural bets we're making and how we'll deliver. The world we're building for We're evolving our strategy to optimize for the future state of software engineering: 1- Software will be built by machines, directed by people. AI is the substrate on which future software gets built. Agents will plan, code, review, deploy, and repair. Humans still own the judgment that matters most: architecture, deep understanding of the customer problem, the tradeoffs that require taste. This is why we built and released the Duo Agent Platform in January. Our first quarter adoption is promising, and we're ready to accelerate. 2- The agentic era multiplies demand for software. Software has been the force multiplier behind nearly every business transformation of the last two decades. The constraint was the cost and time of producing and managing it. That constraint is collapsing. As the cost of producing software collapses, demand for it will expand. Last year, the developer platform market used to be measured in tens of dollars per user per month, this year it is hundreds/user/month and headed to thousands. Not only is the value of software for builders increasing, but we believe there will be more software and builders than ever, and we will serve an increasing volume of both. 3- The consequential work belongs to engineers. Engineering has always been about more than writing code. Great engineers are problem solvers and builders who care about system design, distributed systems, reasoning through failures, safely integrating new capability into critical systems, and making decisions under ambiguity. These are exactly the skills the agentic era needs more of, especially as the volume of software increases. The supply of deep technical problems is multiplying, and the engineers who can solve them will be among the scarcest and most valuable talent in the market. Our core users' roles are evolving, their importance is only increasing. The architectural bets we're making Platforms that weren't built for machine scale are starting to break under it. Winning means investing in the fundamentals that really matter: security, performance, scalability, reliability and user experience. We're making five, fundamental architectural bets. Each one is underway and we plan to deliver without disruption to GitLab customers that depend on us every day. 4- Machine-scale infrastructure. Agents open merge requests in parallel, trigger pipelines around the clock, and push commits at a rate no human team ever did. Git itself wasn't designed for that load, and bolting AI onto platforms not built for agents is the biggest mistake of this era. We're doing a generational rebuild of the underlying infrastructure to handle agent-rate work as the default. Git itself is being reengineered for machine scale. The monolith is giving way to modern, API-first, composable services. And agent-specific APIs are being built so agents can act as first-class users of the platform, not as bolted-on consumers of human-shaped interfaces. The value of this 100x scale infrastructure, and the reliability and performance it provides is much higher than the generation of infrastructure in the market today. 5- Orchestration across the full lifecycle. A single agent that writes code or opens a merge request produces activity. Enterprises don't need agent activity. They need running software that moves the business forward. Orchestration is the layer that gets you there. It coordinates agents across the lifecycle, assigning work, managing state, passing context, resolving conflicts, enforcing policy, and keeping a human in the loop when it matters. CI/CD is one of the components getting reimagined. The GitLab pipeline was designed to take human-rate commits and ship them safely; in the agentic era our orchestration service becomes the runtime that coordinates agents, validates the work and enforces guardrails, and drives change all the way to production at machine rate. 6- Context is our superpower. Every dev tool vendor is converging on similar code generation capabilities. Enterprise AI bills are climbing as fast as adoption. What doesn't commoditize is the unique context the model gets to work with: a data model that connects planning, code, review, security, deployment, and operations across every project and repository, accumulated over years of a team's work. We're investing in that connected data model as a first-class, API-accessible service, and it delivers more value with every human and agent action. Context is what lets agents spend fewer tokens and deliver better results. 7- Governance built into the core. Governance is what lets enterprises move fast in the agentic era. Like a race car, it doesn't matter how fast you can go if you can't maintain control. As agents take on more of the work, enterprises need a platform that can enforce who's allowed to do what, prove what happened and why, and keep sensitive code and data where it belongs. We're building identity, audit, policy, and deployment flexibility as core platform services that every agent, pipeline, and merge request runs through by default, rather than a separate product layered on top. 8- One platform, three modes. Trillions of lines of code run the world's businesses today. Rewriting most of it is too risky and too expensive to justify. The cloud era taught us enterprises run hybrid, and operating across that mix has been painful, expensive, and never fully solved. The agentic era will be the same. Every enterprise will live across a spectrum of human-owned, agent-assisted, and agent-autonomous work. We're building one platform, one data model, one governance system that operates across all three modes, and delivering it cloud and model neutral. How we'll deliver it 9- A flexible business model. As the way software gets built changes, the business model must evolve with it. Agentic AI can augment teams, perform real work and the business model must scale with the cost and value of the work performed. We're keeping what works: the predictability of subscriptions for what customers have today. We've already added consumption pricing for the work agents do, with other major players following over the past few months. Next, we're introducing more flexibility to mix both as the way of work evolves. 10- Culture of excellence. Operational character is a key differentiator. What matters most right now is the ability to move quickly, own outcomes, and deliver real value to our customers. Speed with Quality, Ownership Mindset, and Customer Outcomes are our new operating principles, built on a culture of excellence. To our customers For our customers, the most important thing today is what doesn't change. The support, roadmap commitments, contractual terms — all of it continues without disruption. Your account team is available to walk you through today's news if you'd like a conversation. Where you should expect to see us evolve is in the quality, depth and pace of innovation we ship. We will lead the way in agentic engineering by being customer zero of our platform, demonstrating with our innovation and our results the success you can bet on as our customers. Our vision for the product and business model is clearer than it has ever been and we're accelerating the work. We'll share the next wave of our innovation roadmap at GitLab Transcend on June 10, 2026 and hope you'll join us. To our investors Today's announcement is a deliberate move to lead in a market we believe is in the middle of its largest shift in twenty years. The opportunity here isn't incremental growth on a DevSecOps platform — we're building toward becoming the trusted enterprise platform for software creation in the AI era. We look forward to sharing an update on the business and our Q1 results in our upcoming earnings call on June 2, 2026. We'll also share the final scope and financial impact of the restructuring at that time, although we anticipate reinvesting the majority of savings into accelerating our progress against the specific growth and technological initiatives that we've outlined. This is the most consequential work we've taken on as a company. We'll prove it in the innovation we bring to market, how we serve our customers, and how we create value for our shareholders over the near- and long-term. Thank you, -Bill Staples CEO, GitLab Read the original article on Business Insider

Discussion | Link

Learn how AutoScout24 Group uses Codex and ChatGPT to speed development cycles, improve code quality, and expand AI adoption.

Parameter Golf brought together 1,000+ participants and 2,000+ submissions to explore AI-assisted machine learning research, coding agents, quantization, and novel model design under strict constraints.

Teams use Codex with GPT-5.5 to ship production systems and turn research ideas into runnable experiments.