arXiv:2605.11212v2 Announce Type: new Abstract: Computer-use agents (CUAs) rely on visual observations of graphical user interfaces, where each screenshot is encoded into a large number of visual tokens. As interaction trajectories grow, the token cost increases rapidly, limiting the amount of history that can be incorporated under fixed context and compute budgets. This has resulted in no or very limited improvement in the performance when using history unlike other domains. We address this inefficiency by introducing ReVision, which is used to train multimodal language models on trajectories where redundant visual patches are removed using a learned patch selector that compares patch representations across consecutive screenshots while preserving spatial structure required by the model. Across three benchmarks, OSWorld, WebTailBench, and AgentNetBench, when processing trajectories with 5 history screenshots using Qwen2.5-VL-7B, ReVision reduces token usage by approximately 46% on average while improving success rate by 3% over the no drop baseline. This establishes a clear efficiency gain, enabling agents to process longer trajectories with fewer tokens. With this improved efficiency, we revisit the role of history in CUAs and find that performance continues to improve as more past observations are incorporated when redundancy is removed. This suggests that the commonly observed saturation in visual history is not due to limited usefulness of past information, but rather a consequence of inefficient token representations.

arXiv:2605.12567v1 Announce Type: new Abstract: The inherent electronic and speckle noise complicates clinical interpretation of ultrasound images. Conventional denoising methods rely on explicit noise assumptions whose validity diminishes under composite noise conditions. Learning-based methods require massive labeled data and model parameters. These pre-defined and pre-trained manners entail an inevitable domain shift in complex in vivo environments, so they are limited to a specific noise type and often blur structural details. In this study, we propose a pure test-time training framework for one-shot ultrasound image denoising and apply it to synthetic aperture ultrasound (SAU), which synthesizes transmit focus from sub-aperture transmissions. Our Aperture-to-Aperture (A2A) framework disentangles anatomical similarity and noise randomness from shuffled sub-apertures through self-contrastive learning in pyramid latent spaces. The clean image is then decoded from the anatomy space, while discarding the noise space. A2A is trained at test time on one noisy sample of SAU signals, so it fundamentally eliminates the domain shift and pretraining costs. Simulation experiments, including electronic noise levels of 0 to 30 dB and different inclusion geometries, demonstrated an improvement of 69.3% SNR and 34.4% CNR by A2A. The in vivo results showed 84.8% SNR and 25.7% CNR gains using only two aperture data of the heart in six echocardiographic views, liver, and kidney. A2A delivers clear images/signals across diverse imaging targets and configurations, paving the way for more reliable anatomical visualization and functional assessment by ultrasound.

arXiv:2605.12550v1 Announce Type: new Abstract: Large vision models (LVMs) have recently proven to be surprisingly effective time series forecasters, simply by rendering temporal data as images. This success, how ever, rests on a largely unexamined premise: the rendered time series images are sufficiently close to natural images for knowledge in pre-trained models to transfer effectively. We argue that two gaps still remain, i.e., spectral and structural gaps, fundamentally limiting the potential of LVMs for time series forecasting. Spectrally, we systematically reveal that rendered time series images exhibit a markedly shallower power spectrum than the natural images LVMs are pre-trained to recognize. Structurally, reshaping 1D temporal sequences into 2D grids fabricates spurious spatial adjacencies while severing genuine temporal continuities, misleading the spatial inductive biases of pre-trained LVMs. To bridge these gaps, we propose SSDA, a dual-branch network that spectrally and structurally adapts to unlock the full potential of LVMs for time series forecasting. At the data level, a Spectral Magnitude Aligner (SMA) applies 2D FFT to selectively enhance the magnitude spectrum toward natural-image statistics while preserving phase. At the model level, a Structural-Guided Low-Rank Adaptation (SG-LoRA) injects position-aware temporal encodings into patch embeddings and adapts at tention via low-rank updates. The two branches are further adaptively fused to produce the final forecast. Extensive experiments on seven real-world benchmarks demonstrate that SSDA consistently outperforms strong LVM- and LLM-based baselines under both full-shot and few-shot settings. Code is publicly available at https://anonymous.4open.science/r/SSDA-8C5B.

arXiv:2605.12648v1 Announce Type: new Abstract: We establish the first population risk bounds for Kolmogorov-Arnold Networks (KANs) trained by mini-batch SGD with gradient clipping, covering non-private SGD as well as differentially private SGD (DP-SGD) with Gaussian perturbations that interpolate between independent and temporally correlated noise. This setting is substantially closer to practice than prior KAN theory along two axes: training is by mini-batch SGD, the standard recipe for modern networks, rather than full-batch gradient descent (GD); and correlated-noise mechanisms have empirically shown a more favorable privacy-utility tradeoff than independent-noise mechanisms. Our results cover the corresponding full-batch GD and independent-noise DP-GD results for KANs by Wang et al. (2026), while yielding sharper fixed-second-layer specializations. The technical core is a new analysis route for correlated-noise DP training in the non-convex regime. Temporal dependence breaks the conditional-centering structure underlying standard one-step SGD arguments, and the projection step obstructs the exact cancellation structure of correlated perturbations. We address these difficulties through an auxiliary unprojected dynamics, a shifted iterate that absorbs the current noise perturbation, and a high-probability bootstrap certifying projection inactivity. Combining this optimization analysis with a stability-based generalization argument yields the stated population risk bounds. To the best of our knowledge, this is the first optimization and population risk analysis of a correlated-noise mechanism for DP training beyond convex learning, in particular for neural networks.

Katherine Bindley / Wall Street Journal: ZipRecruiter: IT and CS job postings are up 14.2% YoY in April; entry-level roles fell from 8.1% a year ago to 7.4%, while senior roles grew from 38.8% to 43.1%  —  Layoffs keep coming but there's still a market for higher-grade talent to harness agents  —  It should be a tough …

Discussion | Link

Discussion | Link

The corporate blockchain space looks very different today. The initial hype faded years ago. Companies ignore … Best 5 Companies Building Blockchain Solutions for BusinessRead more The post Best 5 Companies Building Blockchain Solutions for Business appeared first on .

Andrew Left, right, with his lawyer, arriving at court in Los Angeles. Bloomberg/Getty Images Prosecutors say short-seller Andrew Left manipulated the market and deceived retail investors. Two stock analysts who disagreed with Left's negative reports on stocks have testified. Left's defense has pressed the idea that disagreeing on stocks is normal and doesn't indicate fraud. Amid the complicated financial jargon at play in short-seller Andrew Left's ongoing securities fraud trial, the defense has repeatedly returned to a simple question: Can't people disagree about stocks? Citron Research founder Left, known for his bets on GameStop and frequent appearances on financial shows, has been accused of manipulating the market and deceiving retail investors through misleading statements. Prosecutors allege Left said one thing about over 20 stocks and then traded in the opposite direction, making over $20 million in the process. Three days into the trial, it's becoming clearer how the defense plans to handle the prosecution's witnesses. So far, prosecutors have focused on the stocks of two companies they say Left tweeted about and then traded on: the social media site formerly known as Twitter and Cronos Group, a cannabis company. Prosecutors have highlighted the negative reports Left issued about those companies and called stock analysts who disagreed with his takes to testify. Meanwhile, the defense argued that disagreement over a stock does not necessarily constitute fraud. Left issued a report in December 2018 titled "Twitter has become the Harvey Weinstein of social media" that cited an Amnesty International study documenting issues Twitter had regarding abusive content against women. He set a target stock price of $20, well below its then-trading price of around $30. Douglas Anmuth, an internet analyst at JPMorgan, testified Wednesday that he covered Twitter at the time and disagreed with Left's report. He said issues with hate speech on Twitter were well known, and the company had taken steps to address them. The defense pressed Anmuth on whether disagreements among analysts necessarily meant something nefarious was going on. "Fair to say when you and any other analyst disagree, that might just be a difference of opinion, correct?" Adam Fee, Left's lawyer, asked. Anmuth agreed, and a bit later, Fee raised the point again: "You would agree generally that reasonable people can disagree about how information in the market might impact stock prices?" "People can disagree," Anmuth said. On a third occasion, Fee asked Anmuth, whose Twitter price target was $45 at the time of the Citron report, whether "reasonable people" could disagree about price targets. A similar pattern played out during the earlier testimony on Wednesday of Martin Landry, an analyst who was covering Cronos Group in August 2018, when Left tweeted a negative report about the company. Landry testified that he wrote a rebuttal report at the time that said some of Left's allegations about Cronos appeared to be "unfounded and biased." In cross-examination, Fee noted that Landry did not accuse Left of sharing false information, but that he disagreed with the importance Left placed on certain facts. Fee also asked, "If you said 'buy,' another said 'sell,' the guy who said sell wasn't committing fraud, correct?" Landry agreed. Once more, he asked Landry if he agreed that "people can reasonably disagree about the value of a company." The line of questioning seemed to align with the central argument Left's defense led with during opening statements on Tuesday: "He tells the public what he believes — the truth — and then he trades on the truth to make a profit. It's not fraud. That's trading." Left's lawyers also leaned on his more than two decades of experience in the stock market. In their opening statement, they said he wouldn't have lasted so long in the business and had so much influence if he'd been deceiving people. "You would not have made it 20 years picking stocks if you were influenced by improper considerations?" Fee asked Landry. "Exactly," Landry said. "Because people would've stopped listening, correct?" Fee asked. "Yes," Landry said. Read the original article on Business Insider

Discussion | Link

Discussion | Link

Discussion | Link

Carmen Arroyo / Bloomberg: Sources: Apollo Global, Morgan Stanley, and others are testing Grok internally as part of xAI's finance sector push, but financiers are rarely using it for work  —  Elon Musk's xAI has recruited multiple Wall Street firms with ties to the billionaire's business empire to test its Grok chatbot …

From live speech translation in video calls to auto-dubbing on TikTok, the technology to dissolve language barriers has arrived. Real-time translation powered by artificial intelligence (AI) is now embedded in everyday life. Tools from OpenAI, Meta, Google and many others now offer near-instant translation across dozens of languages, and they keep improving. All this raises a vital question. If machines can do this faster and more accurately than humans, is investing years in learning another language still worth it? The logic is appealing. Humans have always offloaded cognitive work onto tools. Writing reduced demands on our memory. Calculators removed the burden of mental arithmetic. AI sits within this long tradition. Used well, it can support learning and expand access in ways that matter enormously. But there’s a difference between using a tool to extend your capabilities and using it to avoid doing something altogether. That distinction becomes important when you are not just replacing a skill, but a form of cognitive and cultural engagement. The effort is the point Effort plays a central role in how we acquire knowledge. Psychologists use the phrase “desirable difficulties” to describe challenges that may feel inefficient, but produce stronger long-term retention and understanding. Struggling with grammar, searching for the right word, or constructing meaning across multiple languages engages brain networks that support memory, attention and cognitive flexibility. Over time, they consolidate knowledge far more deeply than passive exposure. Sustained mental engagement contributes to what researchers call cognitive resilience – the brain’s capacity to maintain function as we age. Managing multiple languages is one form of this engagement. It requires the brain to resolve competition, monitor context and adapt dynamically. These are not trivial demands. And they’re difficult to achieve if you just use translation tools passively, such as resolving the meaning of a foreign phrase with the click of a button. What multilingualism research actually shows The evidence on multilingualism is often presented as a simple “bilingual advantage”, a shorthand that obscures a more complicated picture. Some studies report benefits for attention or working memory, while others find no differences. The truth appears to be more selective. Our recent study examined cognitive performance in 94 adults aged 18 to 83, using both visuospatial and auditory tasks across working memory, attention and inhibition. Put simply, we looked at how people process and respond to information they see or mentally map out in space (visuospatial) and information they hear (auditory). Examples include remembering sounds, focusing on visual patterns, or ignoring distractions. Our study measured multilingualism as a spectrum, not a category. This allowed us to capture diverse language backgrounds and experiences. Multilingual participants spoke a range of languages with varying levels of proficiency and daily use, reflecting the linguistic diversity common within multicultural communities. Across most tasks, multilinguals and monolinguals performed similarly. However, one pattern was striking. Individuals with richer, more diverse multilingual experience showed markedly better performance in visuospatial working memory. These effects were most pronounced in older people. This suggests that multilingual experience doesn’t broadly enhance cognition, like some headlines claim. Instead, it may help preserve specific functions over time. Separate population-level research has also linked multilingualism to later onset of Alzheimer’s disease and better overall ageing outcomes, though the mechanisms continue to be debated. Overall, however, it appears that sustained use of multiple languages represents a form of mental activity with effects that accumulate across a lifetime. What AI translation can’t replicate AI translation excels at speed and accessibility. For many practical purposes, it works remarkably well. But it operates through pattern recognition, not lived understanding. It can struggle with cultural context, humour, register and emotionally embedded meaning, especially for languages with less representation in training data. At best, AI captures literal dimensions of language while missing social ones. Consider the scene in the 2003 film Love Actually where Jamie, played by Colin Firth, delivers an awkward but sincere proposal to Aurelia in broken Portuguese. It is moving because of the effort, vulnerability and intent his imperfect words carry. Resort to real-time translation software and what remains is information, not expression. This is the deeper distinction: translation is not the same as participation. Learning a language involves understanding how people think, their values, and how meaning is shaped by context and history. This cultural literacy develops through interaction and experience. We can’t fully outsource that to systems that translate on demand. The multilingual participants in our research spoke to this directly: I definitely think in Telugu, but I remember numbers and count using English. Afrikaans is the language of my heart and best used to express intense emotion. English is the language of business and used mostly in everyday life. These are not descriptions of switching between translation modes. They are descriptions of inhabiting different selves. AI will continue to change how we engage with language learning. It can personalise instruction, minimise barriers and provide feedback at scale. What it can’t do is replace the cognitive and cultural work that comes from learning a language. This work leads to a deeper relationship with how other people see the world, and with how you express yourself. And that difference still matters. Olivia Maurice completed her PhD at the MARCS Institute, Western Sydney University. Mark Antoniou receives funding from the Australian Research Council.

Learn how new ChatGPT safety updates improve context awareness in sensitive conversations, helping detect risk over time and respond more safely.

<p><strong><a href="https://datasette.io/blog/2026/new-blog/">Welcome to the Datasette blog</a></strong></p> We have a bunch of neat Datasette announcements in the pipeline so we decided it was time the project grew an official blog.</p> <p>I built this using OpenAI Codex desktop, which turns out to have the Markdown session transcript export feature I've always wanted. Here's <a href="https://gist.github.com/simonw/885b11eee46822622b8031a1f4e5f3a3">the session that built the blog</a>. See also <a href="https://github.com/simonw/datasette.io/issues/179">issue 179</a>. <p>Tags: <a href="https://simonwillison.net/tags/ai">ai</a>, <a href="https://simonwillison.net/tags/datasette">datasette</a>, <a href="https://simonwillison.net/tags/generative-ai">generative-ai</a>, <a href="https://simonwillison.net/tags/llms">llms</a>, <a href="https://simonwillison.net/tags/ai-assisted-programming">ai-assisted-programming</a>, <a href="https://simonwillison.net/tags/codex">codex</a></p>

LinkedIn CMO Jessica Jensen. Arnold Jerocki/Variety via Getty Images LinkedIn's marketing team is embracing AI as it cuts staff, an internal memo shows. Marketing will also reduce its ad spending, according to an email from LinkedIn CMO Jessica Jensen. LinkedIn says it's positioning itself for future success as it implements "organizational changes." The marketing team wasn't spared from LinkedIn's layoffs on Wednesday. LinkedIn chief marketing and strategy officer Jessica Jensen sent an email saying the company is reducing roles "across our team" to cut costs. "Growth is more competitive, infrastructure costs continue to rise and AI is reshaping how work gets done," read Jensen's email, which was viewed by Business Insider. "So we are at a point where our work is important and deeply valued, and we need to reduce costs in Marketing to ensure LinkedIn is positioned for the future we're building." It's not clear exactly how many members of LinkedIn's marketing team got cut. A LinkedIn spokesperson declined to provide specific numbers, telling Business Insider that it is implementing "organizational changes" to best position itself for future success. Jensen's memo says the marketing team is reducing its paid media spending. While Jensen didn't give specific examples, paid media typically means paying for Google Ads or other platforms for promotion. LinkedIn plans to embrace unspecified "AI tools and workflows" to help staff get things done faster. LinkedIn's product organizations have also been affected by the layoffs, as they made changes to their operations. These cuts put LinkedIn squarely in line with the long list of tech companies that are pushing staff to use AI as they reduce head count. Cisco announced an AI-driven restructuring on Wednesday affecting about 4,000 roles. Companies like Coinbase and Block also cited increased productivity from AI as part of the rationale for their own layoffs earlier this year. LinkedIn is investing in AI in other ways, too. It launched an "AI labor marketplace" where people can make money training AI chatbots to get better at everything from coding to nursing to finance, Business Insider previously reported. Read the full memo Jensen sent out on Wednesday below: Team, I want to start by saying that this team is doing extraordinary work. You're pushing our brand into new creative territory, delivering results that drive business growth, and being recognized for that in the industry — and it's making a real impact. Thank you. At the same time, we're operating in a moment that requires more focus and intentionality. Growth is more competitive, infrastructure costs continue to rise and AI is reshaping how work gets done. So we are at a point where our work is important and deeply valued, and we need to reduce costs in Marketing to ensure LinkedIn is positioned for the future we're building. With this in mind, we're making some organizational changes, including reducing roles across our team. If your role is part of the reductions, or proposed reductions in EMEA and APAC, you will receive a calendar invite in the next hour titled 'Attendance Required: Marketing Organizational Updates.' If you do not receive this exact invite your role is not part of the reductions or proposed reductions. To give you more context, our organizational changes are across three areas: First, sharpening our financial focus. In addition to the difficult role reductions we're announcing today, we're also reducing our paid media and program spend, while continuing to invest where we see the highest ROI. This is about being more intentional with our resources, and creating space to fund the work that moves the business the most. Second, focusing our investments on where we can grow and drive high ROI in FY27. That includes accelerating LMS growth, advancing our agentic hiring solutions, and building on strong momentum with Premium and Small Business offerings. We'll prioritize the consumer experiences that most directly drive engagement and revenue, concentrate our marketing investments in the US and UK, and maintain our market leadership in LSS as we continue to strengthen the product. Third, setting teams up for success. We're continuing to bring teams together that do similar work in our Creative CoE, our Scaled Solutions org, and Small Business Marketing. We'll also embrace new AI-enabled tools and workflows that allow our human creativity and judgment to go further, faster. To those leaving LinkedIn, thank you for the care, creativity and commitment you brought to our culture and brand. You've made a lasting impact, and we're deeply grateful. For those staying, please join me in supporting our departing colleagues as they navigate what's next. We'll come together tomorrow for an all Marketing team meeting where I'll provide more clarity and you'll have a chance to ask questions. You can learn more about the changes within each team on the Marketing Hub which will include links to updates that GBO, Product, and Engineering are sharing today as well. Your leaders will provide more context on what's happening on your team in meetings later this week. While I know that this level of change can bring a sense of uncertainty, it can also clarify our purpose. The world needs LinkedIn now more than ever, and while these are hard choices to make, they focus our work for our short and long-term impact. Please take the time you need to process this news and give one another space at this moment. Thank you for everything you do for LinkedIn and for one another. Jessica Have a tip? Contact this reporter via email at crollet@businessinsider.com or on Signal and WhatsApp at 628-282-2811. Use a personal email address, a nonwork WiFi network, and a nonwork device; here's our guide to sharing information securely. Read the original article on Business Insider

Discussion | Link

Carl Franzen / VentureBeat: Anthropic unveils Claude Agent SDK credits for paid plans, which users can allocate for programmatic use of third-party agents like OpenClaw, starting June 15  —  Good news, OpenClaw fans — you can once again use your Claude AI subscription to power the hit, open source, autonomous AI agentic harness!