{
  "title": "2026-06-19 AI Daily Update | AI in Healthcare Moves Towards Clinical Assistance, Enterprise Deployment Enters Cost-Controllable Phase",
  "url": "https://miaok.ong/en/ai-daily/ai-daily-2026-06-19/",
  "date": "2026-06-19T07:00:00+08:00",
  "lastmod": "2026-06-19T07:00:00+08:00",
  "type": "ai-daily",
  "kind": "page",
  "language": "en",
  "description": "Today\u0026rsquo;s main theme is AI moving from general capabilities to verifiable implementation: OpenAI strengthens health intelligence and demonstrates cases of inference models assisting in the diagnosis of rare childhood diseases, with medical AI approaching clinical workflows; meanwhile, Enterprise ChatGPT enhances usage analysis and expenditure control, showing that commercial competition is shifting towards cost, auditing, and ROI. The agent field is also improving retrieval, memory, and collaboration reliability.",
  "keywords": null,
  "tags": [],
  "categories": [],
  "author": "Mark (Miao) Kong",
  "image": "https://miaok.ong/images/avatar.jpg",
  "content": "\u003ch1 id=\"2026-06-19-ai-daily--ai-in-healthcare-moves-toward-clinical-assistance-enterprise-deployment-enters-cost-controllable-stage\"\u003e\n  2026-06-19 AI Daily | AI in Healthcare Moves Toward Clinical Assistance, Enterprise Deployment Enters Cost-Controllable Stage\n  \u003ca class=\"heading-link\" href=\"#2026-06-19-ai-daily--ai-in-healthcare-moves-toward-clinical-assistance-enterprise-deployment-enters-cost-controllable-stage\"\u003e\n    \u003ci class=\"fa-solid fa-link\" aria-hidden=\"true\" title=\"Link to heading\"\u003e\u003c/i\u003e\n    \u003cspan class=\"sr-only\"\u003eLink to heading\u003c/span\u003e\n  \u003c/a\u003e\n\u003c/h1\u003e\n\u003cblockquote\u003e\n\u003cp\u003eToday\u0026rsquo;s main theme is AI\u0026rsquo;s shift from general capabilities to verifiable implementation: OpenAI enhances health intelligence and showcases a case where an inference model assists in diagnosing rare pediatric diseases, bringing medical AI closer to clinical workflows; meanwhile, ChatGPT Enterprise strengthens usage analysis and expenditure control, indicating that commercial competition is shifting towards cost, auditing, and ROI. The field of agents is also shoring up reliability in retrieval, memory, and collaboration.\u003c/p\u003e\n\u003c/blockquote\u003e\n\u003ch2 id=\"-deep-dive-this-issues-watch-list\"\u003e\n  📖 Deep Dive: This Issue\u0026rsquo;s Watch List\n  \u003ca class=\"heading-link\" href=\"#-deep-dive-this-issues-watch-list\"\u003e\n    \u003ci class=\"fa-solid fa-link\" aria-hidden=\"true\" title=\"Link to heading\"\u003e\u003c/i\u003e\n    \u003cspan class=\"sr-only\"\u003eLink to heading\u003c/span\u003e\n  \u003c/a\u003e\n\u003c/h2\u003e\n\u003cp\u003eThe most important read today is how AI in healthcare is evolving from \u0026ldquo;health Q\u0026amp;A\u0026rdquo; to clinical-grade assistance: OpenAI has updated ChatGPT\u0026rsquo;s health intelligence and disclosed a case where an inference model aids in diagnosing rare pediatric genetic diseases; combined with papers on SpeechDx and digital twin clinical decision-making, it\u0026rsquo;s crucial for medical AI teams to closely follow evaluations, safety boundaries, and real-world workflow implementation.\u003c/p\u003e\n\u003cp\u003eThe second theme is the reliability and scalability of agents. Agentic Search, self-evolving rules for legal retrieval, the long-term memory benchmark MemTrace, and distributed general agent networks all address the same question: agents are not just about \u0026ldquo;running more iterations,\u0026rdquo; but about solving retrieval redundancy, memory consistency, collaborative governance, and verifiable reasoning.\u003c/p\u003e\n\u003cp\u003eFinally, enterprise AI is entering a stage of fine-tuned operations. ChatGPT Enterprise\u0026rsquo;s new features for usage analysis and expenditure control echo discussions on ROI in e-commerce, supply chain, and manufacturing scenarios. It is recommended that managers view today\u0026rsquo;s updates together: the competitive focus of AI is shifting from model capabilities to controllable costs, auditable results, and industry-specific deployment.\u003c/p\u003e\n\u003ch2 id=\"-quick-brief-ai-hot-topics-on-x\"\u003e\n  🌐 Quick Brief: AI Hot Topics on X\n  \u003ca class=\"heading-link\" href=\"#-quick-brief-ai-hot-topics-on-x\"\u003e\n    \u003ci class=\"fa-solid fa-link\" aria-hidden=\"true\" title=\"Link to heading\"\u003e\u003c/i\u003e\n    \u003cspan class=\"sr-only\"\u003eLink to heading\u003c/span\u003e\n  \u003c/a\u003e\n\u003c/h2\u003e\n\u003ch3 id=\"topic-1-anthropic-launches-artifacts-for-shareable-claude-code-sessions\"\u003e\n  Topic 1: Anthropic Launches Artifacts for Shareable Claude Code Sessions\n  \u003ca class=\"heading-link\" href=\"#topic-1-anthropic-launches-artifacts-for-shareable-claude-code-sessions\"\u003e\n    \u003ci class=\"fa-solid fa-link\" aria-hidden=\"true\" title=\"Link to heading\"\u003e\u003c/i\u003e\n    \u003cspan class=\"sr-only\"\u003eLink to heading\u003c/span\u003e\n  \u003c/a\u003e\n\u003c/h3\u003e\n\u003cul\u003e\n\u003cli\u003eCategory: AI · News\u003c/li\u003e\n\u003cli\u003eOverview: Trending Time: 5 hours ago, Related Posts: 2500\u003c/li\u003e\n\u003cli\u003eWhat it is: Anthropic has launched Artifacts, a feature that allows users to share coding sessions and outputs from Claude Code.\u003c/li\u003e\n\u003cli\u003eWhy it\u0026rsquo;s important: This shows that AI programming assistants are evolving from single-query tools into reusable, presentable, and collaborative workflow platforms, which helps lower the barrier for teams to share their AI development processes.\u003c/li\u003e\n\u003cli\u003eDiscussion summary: Discussions on X are mainly focused on whether this feature will improve collaboration efficiency, code reproducibility, and educational value. Some are also concerned about potential privacy risks, code leaks, and corporate compliance issues associated with sharing sessions.\u003c/li\u003e\n\u003c/ul\u003e\n\u003ch3 id=\"topic-2-aether-ai-raises-20m-for-causal-world-models\"\u003e\n  Topic 2: Aether AI Raises $20M for Causal World Models\n  \u003ca class=\"heading-link\" href=\"#topic-2-aether-ai-raises-20m-for-causal-world-models\"\u003e\n    \u003ci class=\"fa-solid fa-link\" aria-hidden=\"true\" title=\"Link to heading\"\u003e\u003c/i\u003e\n    \u003cspan class=\"sr-only\"\u003eLink to heading\u003c/span\u003e\n  \u003c/a\u003e\n\u003c/h3\u003e\n\u003cul\u003e\n\u003cli\u003eCategory: AI · News\u003c/li\u003e\n\u003cli\u003eOverview: Trending Time: 21 hours ago, Related Posts: 7400\u003c/li\u003e\n\u003cli\u003eWhat it is: Aether AI announced it has raised $20 million in funding to develop world models for causal reasoning.\u003c/li\u003e\n\u003cli\u003eWhy it\u0026rsquo;s important: Causal world models are seen as a crucial direction for enhancing AI\u0026rsquo;s understanding of environmental changes, action consequences, and complex decision-making. This could advance robotics, autonomous driving, and agent systems from pattern recognition to more reliable reasoning and planning.\u003c/li\u003e\n\u003cli\u003eDiscussion summary: Discussions on X are focused on whether causal modeling can overcome the limitations of current large models, Aether AI\u0026rsquo;s technical roadmap and commercial prospects, and whether $20 million is sufficient to support the high cost of world model R\u0026amp;D. Some also question if the concept is being overhyped.\u003c/li\u003e\n\u003c/ul\u003e\n\u003ch3 id=\"topic-3-google-ai-pioneer-noam-shazeer-joins-openai-after-leading-gemini\"\u003e\n  Topic 3: Google AI Pioneer Noam Shazeer Joins OpenAI After Leading Gemini\n  \u003ca class=\"heading-link\" href=\"#topic-3-google-ai-pioneer-noam-shazeer-joins-openai-after-leading-gemini\"\u003e\n    \u003ci class=\"fa-solid fa-link\" aria-hidden=\"true\" title=\"Link to heading\"\u003e\u003c/i\u003e\n    \u003cspan class=\"sr-only\"\u003eLink to heading\u003c/span\u003e\n  \u003c/a\u003e\n\u003c/h3\u003e\n\u003cul\u003e\n\u003cli\u003eCategory: AI · News\u003c/li\u003e\n\u003cli\u003eOverview: Trending Time: 23 hours ago, Related Posts: 12000\u003c/li\u003e\n\u003cli\u003eWhat it is: Noam Shazeer, co-author of the Transformer paper and former head of Gemini, has reportedly left Google to join OpenAI.\u003c/li\u003e\n\u003cli\u003eWhy it\u0026rsquo;s important: Shazeer is one of the key figures behind modern large model architecture. His move highlights the intensifying competition among leading AI companies for top research talent, model architecture capabilities, and product leadership.\u003c/li\u003e\n\u003cli\u003eDiscussion summary: The discussion on X centers on whether this signifies a further lead for OpenAI in the talent war, why Google has lost another key figure, and how Shazeer\u0026rsquo;s arrival will impact the next generation of models and the competitive landscape between Gemini and OpenAI.\u003c/li\u003e\n\u003c/ul\u003e\n\u003ch3 id=\"topic-4-wispr-flow-ceo-collects-800-user-feedback-replies-for-app-fixes\"\u003e\n  Topic 4: Wispr Flow CEO Collects 800 User Feedback Replies for App Fixes\n  \u003ca class=\"heading-link\" href=\"#topic-4-wispr-flow-ceo-collects-800-user-feedback-replies-for-app-fixes\"\u003e\n    \u003ci class=\"fa-solid fa-link\" aria-hidden=\"true\" title=\"Link to heading\"\u003e\u003c/i\u003e\n    \u003cspan class=\"sr-only\"\u003eLink to heading\u003c/span\u003e\n  \u003c/a\u003e\n\u003c/h3\u003e\n\u003cul\u003e\n\u003cli\u003eCategory: AI · News\u003c/li\u003e\n\u003cli\u003eOverview: Trending Time: , Related Posts: 20\u003c/li\u003e\n\u003cli\u003eWhat it is: The CEO of the voice input app Wispr Flow solicited and collected around 800 pieces of user feedback on X to drive app fixes and improvements.\u003c/li\u003e\n\u003cli\u003eWhy it\u0026rsquo;s important: This reflects that AI voice interaction products are entering a phase of rapid iteration. User experience, accuracy, and workflow integration will directly affect the adoption of such tools.\u003c/li\u003e\n\u003cli\u003eDiscussion Overview: Discussions on X are primarily focused on Wispr Flow\u0026rsquo;s recognition quality, stability, privacy handling, and product responsiveness. Supporters believe that the CEO directly collecting feedback contributes to rapid improvement, while critics worry about the large volume of feedback and the opaque prioritization of its implementation.\u003c/li\u003e\n\u003c/ul\u003e\n\u003ch3 id=\"topic-5-trump-says-anthropic-ai-negotiations-progress-amid-export-ban\"\u003e\n  Topic 5: Trump Says Anthropic AI Negotiations Progress Amid Export Ban\n  \u003ca class=\"heading-link\" href=\"#topic-5-trump-says-anthropic-ai-negotiations-progress-amid-export-ban\"\u003e\n    \u003ci class=\"fa-solid fa-link\" aria-hidden=\"true\" title=\"Link to heading\"\u003e\u003c/i\u003e\n    \u003cspan class=\"sr-only\"\u003eLink to heading\u003c/span\u003e\n  \u003c/a\u003e\n\u003c/h3\u003e\n\u003cul\u003e\n\u003cli\u003eCategory: AI · News\u003c/li\u003e\n\u003cli\u003eOverview: Trending 2 days ago, 30,000 related posts\u003c/li\u003e\n\u003cli\u003eWhat it is: Trump stated that negotiations with Anthropic on AI-related matters are progressing, against the backdrop of ongoing U.S. export restrictions on certain AI technologies and chips.\u003c/li\u003e\n\u003cli\u003eWhy it\u0026rsquo;s important: This event highlights how the relationship between leading AI companies, government policies, and export controls is impacting the international expansion, compute acquisition, and geopolitical landscape of the large model industry.\u003c/li\u003e\n\u003cli\u003eDiscussion Overview: Discussions on X are mainly focused on whether the negotiations will lead to a relaxation of export policies, whether the U.S. should prioritize AI safety and national security, and how companies like Anthropic should balance business expansion with regulatory compliance.\u003c/li\u003e\n\u003c/ul\u003e\n\u003ch4 id=\"summary-of-ai-public-opinion-on-x-today\"\u003e\n  Summary of AI Public Opinion on X Today\n  \u003ca class=\"heading-link\" href=\"#summary-of-ai-public-opinion-on-x-today\"\u003e\n    \u003ci class=\"fa-solid fa-link\" aria-hidden=\"true\" title=\"Link to heading\"\u003e\u003c/i\u003e\n    \u003cspan class=\"sr-only\"\u003eLink to heading\u003c/span\u003e\n  \u003c/a\u003e\n\u003c/h4\u003e\n\u003cp\u003eToday\u0026rsquo;s main narrative focuses on AI\u0026rsquo;s transition from technological breakthroughs to industrialization, platformization, and geopolitical competition: programming assistants, voice input, and world models are all being discussed in terms of how to truly embed them into workflows and boost productivity. A clear consensus is that competition in AI products is no longer just about model capabilities, but also about collaborative experience, iteration speed, talent density, and the ability to support complex reasoning. The main points of contention lie in whether the market narrative is outpacing actual capabilities—for instance, whether causal world models are being overhyped, whether shared coding sessions can genuinely enhance collaboration, and to what extent the movement of top talent will alter the corporate landscape. Potential risks are concentrated in privacy and code leakage, corporate compliance, opaque governance of user feedback, and uncertainties brought by export controls and government negotiations. Overall, today\u0026rsquo;s discussions reflect both optimism about the accelerating maturity of AI tools and continued vigilance regarding security, regulation, and commercialization bubbles.\u003c/p\u003e\n\u003ch2 id=\"-influencer-insights\"\u003e\n  💡 Influencer Insights\n  \u003ca class=\"heading-link\" href=\"#-influencer-insights\"\u003e\n    \u003ci class=\"fa-solid fa-link\" aria-hidden=\"true\" title=\"Link to heading\"\u003e\u003c/i\u003e\n    \u003cspan class=\"sr-only\"\u003eLink to heading\u003c/span\u003e\n  \u003c/a\u003e\n\u003c/h2\u003e\n\u003cp\u003eBased on the latest discussions from several top AI influencers on the X platform over the past 24 hours, here are the key industry insights:\u003c/p\u003e\n\u003chr\u003e\n\u003ch1 id=\"ai-industry-daily-insights-report\"\u003e\n  AI Industry Daily Insights Report\n  \u003ca class=\"heading-link\" href=\"#ai-industry-daily-insights-report\"\u003e\n    \u003ci class=\"fa-solid fa-link\" aria-hidden=\"true\" title=\"Link to heading\"\u003e\u003c/i\u003e\n    \u003cspan class=\"sr-only\"\u003eLink to heading\u003c/span\u003e\n  \u003c/a\u003e\n\u003c/h1\u003e\n\u003cp\u003e\u003cstrong\u003eDate: June 18, 2026\u003c/strong\u003e\u003c/p\u003e\n\u003ch2 id=\"1-todays-focus-the-rise-of-on-device-models-and-the-evolution-of-ai-programming-tools\"\u003e\n  1. Today\u0026rsquo;s Focus: The Rise of On-Device Models and the Evolution of AI Programming Tools\n  \u003ca class=\"heading-link\" href=\"#1-todays-focus-the-rise-of-on-device-models-and-the-evolution-of-ai-programming-tools\"\u003e\n    \u003ci class=\"fa-solid fa-link\" aria-hidden=\"true\" title=\"Link to heading\"\u003e\u003c/i\u003e\n    \u003cspan class=\"sr-only\"\u003eLink to heading\u003c/span\u003e\n  \u003c/a\u003e\n\u003c/h2\u003e\n\u003cp\u003eDiscussions among top influencers today were highly concentrated on two main directions: \u003cstrong\u003ethe practical progress of local/on-device models\u003c/strong\u003e and \u003cstrong\u003eAI programming tools moving towards deeper workflow integration\u003c/strong\u003e.\u003c/p\u003e\n\u003ch3 id=\"on-device-models-the-tipping-point-from-usable-to-great\"\u003e\n  On-Device Models: The Tipping Point from \u0026ldquo;Usable\u0026rdquo; to \u0026ldquo;Great\u0026rdquo;\n  \u003ca class=\"heading-link\" href=\"#on-device-models-the-tipping-point-from-usable-to-great\"\u003e\n    \u003ci class=\"fa-solid fa-link\" aria-hidden=\"true\" title=\"Link to heading\"\u003e\u003c/i\u003e\n    \u003cspan class=\"sr-only\"\u003eLink to heading\u003c/span\u003e\n  \u003c/a\u003e\n\u003c/h3\u003e\n\u003cp\u003e\u003ccode\u003e@zhixianio\u003c/code\u003e was at the center of today\u0026rsquo;s on-device model discussions, intensively sharing hands-on experiences with several local models and drawing quite optimistic conclusions:\u003c/p\u003e\n\u003cul\u003e\n\u003cli\u003e\u003cstrong\u003eMiniCPM-o 4.5\u003c/strong\u003e: He tested its full-duplex audio-visual capabilities and, despite occasional audio loss and stability issues during prolonged operation, found the quality \u0026ldquo;very satisfactory.\u0026rdquo; He marveled, \u0026ldquo;It\u0026rsquo;s hard to imagine a 9B model can achieve this.\u0026rdquo; This suggests that the real-time multimodal interaction capabilities of small-parameter models are approaching the threshold of practical use.\u003c/li\u003e\n\u003cli\u003e\u003cstrong\u003eGemma 4 12B\u003c/strong\u003e: His testing of Google\u0026rsquo;s new model revealed its multilingual capabilities (excellent in English and Japanese, poor in Chinese) and pointed out its limited audio parsing ability (\u0026ldquo;knows it\u0026rsquo;s a man\u0026rsquo;s voice, but can\u0026rsquo;t describe the music\u0026rdquo;).\u003c/li\u003e\n\u003cli\u003e\u003cstrong\u003eQwen3.6-35B-A3B\u003c/strong\u003e: \u003ccode\u003e@zhixianio\u003c/code\u003e hailed this MoE model as the \u0026ldquo;sweet spot 🍮,\u0026rdquo; claiming that for personal assistant and programming tasks on a Mac, its speed and intelligence surpass remote LLMs, offering an even better experience than DSV4 Pro. This indicates that \u003cstrong\u003erunning MoE models on high-end consumer hardware can now provide an experience comparable to or even surpassing cloud services.\u003c/strong\u003e\u003c/li\u003e\n\u003c/ul\u003e\n\u003ch3 id=\"ai-programming-tools-from-writing-code-to-collaboration-and-automation\"\u003e\n  AI Programming Tools: From \u0026ldquo;Writing Code\u0026rdquo; to \u0026ldquo;Collaboration and Automation\u0026rdquo;\n  \u003ca class=\"heading-link\" href=\"#ai-programming-tools-from-writing-code-to-collaboration-and-automation\"\u003e\n    \u003ci class=\"fa-solid fa-link\" aria-hidden=\"true\" title=\"Link to heading\"\u003e\u003c/i\u003e\n    \u003cspan class=\"sr-only\"\u003eLink to heading\u003c/span\u003e\n  \u003c/a\u003e\n\u003c/h3\u003e\n\u003cul\u003e\n\u003cli\u003e\u003cstrong\u003eClaude Code Introduces Artifacts\u003c/strong\u003e: \u003ccode\u003e@dotey\u003c/code\u003e offered a detailed interpretation of this feature. He argues that it solves a core pain point for AI Agents: work products are no longer trapped in a terminal \u0026ldquo;black box\u0026rdquo; but become visible and shareable. It transforms debugging timelines, architectural documents, and more into real-time updated web links, allowing team members to \u003cstrong\u003ecollaborate asynchronously based on a single view\u003c/strong\u003e, eliminating the \u0026ldquo;manual translation\u0026rdquo; step.\u003c/li\u003e\n\u003cli\u003e\u003cstrong\u003eCodex\u0026rsquo;s Record \u0026amp; Replay\u003c/strong\u003e: \u003ccode\u003e@dotey\u003c/code\u003e also introduced this newly launched feature. A user simply demonstrates a repetitive task once on a Mac, and Codex can observe and generate a reusable Skill file. This marks an evolution in \u003cstrong\u003ehow AI Agents are instructed, moving from \u0026ldquo;Prompt Engineering\u0026rdquo; to \u0026ldquo;Behavioral Demonstration,\u0026rdquo;\u003c/strong\u003e which greatly lowers the barrier to automating daily office workflows.\u003c/li\u003e\n\u003c/ul\u003e\n\u003ch2 id=\"2-notable-unique-perspectives-and-industry-foresight\"\u003e\n  2. Notable Unique Perspectives and Industry Foresight\n  \u003ca class=\"heading-link\" href=\"#2-notable-unique-perspectives-and-industry-foresight\"\u003e\n    \u003ci class=\"fa-solid fa-link\" aria-hidden=\"true\" title=\"Link to heading\"\u003e\u003c/i\u003e\n    \u003cspan class=\"sr-only\"\u003eLink to heading\u003c/span\u003e\n  \u003c/a\u003e\n\u003c/h2\u003e\n\u003ch3 id=\"-re-evaluating-the-next-token-prediction-path\"\u003e\n  ① Re-evaluating the \u0026ldquo;Next Token Prediction\u0026rdquo; Path\n  \u003ca class=\"heading-link\" href=\"#-re-evaluating-the-next-token-prediction-path\"\u003e\n    \u003ci class=\"fa-solid fa-link\" aria-hidden=\"true\" title=\"Link to heading\"\u003e\u003c/i\u003e\n    \u003cspan class=\"sr-only\"\u003eLink to heading\u003c/span\u003e\n  \u003c/a\u003e\n\u003c/h3\u003e\n\u003cp\u003e\u003ccode\u003e@Pluvio9yte\u003c/code\u003e, connecting the news of the Ministry of Industry and Information Technology\u0026rsquo;s push for a \u0026ldquo;task mode\u0026rdquo; for humanoid robots and Unitree Robotics\u0026rsquo; IPO, retweeted a fundamental question: \u003cstrong\u003e\u0026ldquo;The essence of current models is to predict the next token. Was this the wrong path from the very beginning?\u0026rdquo;\u003c/strong\u003e\nNext, he introduced \u003cstrong\u003eAether AI\u003c/strong\u003e, founded by Professor Huang Biwei, a company dedicated to building \u0026ldquo;Causal World Models.\u0026rdquo; This initiative aims to shift AI\u0026rsquo;s understanding from \u0026ldquo;data correlation\u0026rdquo; to \u0026ldquo;how the physical world operates.\u0026rdquo;\n\u003cstrong\u003eInsight:\u003c/strong\u003e This represents deep industry thinking about the limitations of LLMs. When tasks like embodied intelligence and scientific discovery require stable and rigorous reasoning about the physical world, purely probability-based token prediction models may fall short. \u003cstrong\u003eIntroducing causal reasoning could be the key variable in the next technological evolution.\u003c/strong\u003e\u003c/p\u003e\n\u003ch3 id=\"-ai-efficiency-paradox-efficiency-improved-now-what\"\u003e\n  ② \u0026ldquo;AI Efficiency Paradox\u0026rdquo;: Efficiency improved, now what?\n  \u003ca class=\"heading-link\" href=\"#-ai-efficiency-paradox-efficiency-improved-now-what\"\u003e\n    \u003ci class=\"fa-solid fa-link\" aria-hidden=\"true\" title=\"Link to heading\"\u003e\u003c/i\u003e\n    \u003cspan class=\"sr-only\"\u003eLink to heading\u003c/span\u003e\n  \u003c/a\u003e\n\u003c/h3\u003e\n\u003cp\u003e\u003ccode\u003e@ruanyf\u003c/code\u003e shared a thought-provoking discussion: AI allows work that previously took a week to be completed in a few hours. \u003cstrong\u003e\u0026ldquo;Can we take a day off because of this?\u0026rdquo;\u003c/strong\u003e This question directly addresses the huge gap between increased AI productivity and personal well-being. He pointed out that if there are no days off and no salary increases, what is the meaning of AI for employees? A possible long-term answer is: \u003cstrong\u003eAI will force an increase in average social wages or welfare levels across society.\u003c/strong\u003e This is not just a technical problem, but a redesign of the social contract.\u003c/p\u003e\n\u003ch3 id=\"-the-path-from-vibe-coder-to-software-engineer\"\u003e\n  ③ The Path from \u0026ldquo;Vibe Coder\u0026rdquo; to \u0026ldquo;Software Engineer\u0026rdquo;\n  \u003ca class=\"heading-link\" href=\"#-the-path-from-vibe-coder-to-software-engineer\"\u003e\n    \u003ci class=\"fa-solid fa-link\" aria-hidden=\"true\" title=\"Link to heading\"\u003e\u003c/i\u003e\n    \u003cspan class=\"sr-only\"\u003eLink to heading\u003c/span\u003e\n  \u003c/a\u003e\n\u003c/h3\u003e\n\u003cp\u003e\u003ccode\u003e@Pluvio9yte\u003c/code\u003e shared his journey of catching up on full-stack development experience and distilled a core insight: \u003cstrong\u003eThe best practice for Vibe Coding is neither \u0026ldquo;requirements-first\u0026rdquo; nor \u0026ldquo;code-first,\u0026rdquo; but \u0026ldquo;Contract First.\u0026rdquo;\u003c/strong\u003e Based on this, he developed the OpenSpec framework, which aims to provide a stable, non-driftable collaboration foundation for humans and AI by defining API contracts and data models in advance. This is a profound practical summary of how to use AI for serious engineering development in a standardized way.\u003c/p\u003e\n\u003ch3 id=\"-other-forward-looking-perspectives\"\u003e\n  ③ Other Forward-Looking Perspectives\n  \u003ca class=\"heading-link\" href=\"#-other-forward-looking-perspectives\"\u003e\n    \u003ci class=\"fa-solid fa-link\" aria-hidden=\"true\" title=\"Link to heading\"\u003e\u003c/i\u003e\n    \u003cspan class=\"sr-only\"\u003eLink to heading\u003c/span\u003e\n  \u003c/a\u003e\n\u003c/h3\u003e\n\u003cul\u003e\n\u003cli\u003e\u003cstrong\u003eAI programming costs more than human engineers\u003c/strong\u003e: \u003ccode\u003e@ruanyf\u003c/code\u003e pointed out that OpenAI employees consume $1.3 million worth of Tokens in a month. Even with cheaper domestic models, it would still cost 2-3 million RMB per year. This reminds enterprises that \u003cstrong\u003eunlimited use of top-tier AI for programming might cost far more than hiring experienced human engineers.\u003c/strong\u003e\u003c/li\u003e\n\u003cli\u003e\u003cstrong\u003eTesting is the new moat\u003c/strong\u003e: \u003ccode\u003e@ruanyf\u003c/code\u003e observed that Cloudflare engineers spent $1100 in Tokens to replicate Next.js with AI. He believes that the moat of code has disappeared, and in the future, \u003cstrong\u003ethe key to preventing software from being easily replicated lies in its vast test cases and engineering accumulation.\u003c/strong\u003e\u003c/li\u003e\n\u003c/ul\u003e\n\u003ch2 id=\"3-recommended-tools-and-resources\"\u003e\n  3. Recommended Tools and Resources\n  \u003ca class=\"heading-link\" href=\"#3-recommended-tools-and-resources\"\u003e\n    \u003ci class=\"fa-solid fa-link\" aria-hidden=\"true\" title=\"Link to heading\"\u003e\u003c/i\u003e\n    \u003cspan class=\"sr-only\"\u003eLink to heading\u003c/span\u003e\n  \u003c/a\u003e\n\u003c/h2\u003e\n\u003ch3 id=\"-productivity-tools\"\u003e\n  🛠️ Productivity Tools\n  \u003ca class=\"heading-link\" href=\"#-productivity-tools\"\u003e\n    \u003ci class=\"fa-solid fa-link\" aria-hidden=\"true\" title=\"Link to heading\"\u003e\u003c/i\u003e\n    \u003cspan class=\"sr-only\"\u003eLink to heading\u003c/span\u003e\n  \u003c/a\u003e\n\u003c/h3\u003e\n\u003cul\u003e\n\u003cli\u003e\u003cstrong\u003eYoumind\u003c/strong\u003e: A content creation tool highly recommended by several bloggers, including \u003ccode\u003e@AI_Jasonyu\u003c/code\u003e and \u003ccode\u003e@gefei55\u003c/code\u003e, especially good at generating high-quality long-form articles and cross-platform (public accounts, X) typesetting.\u003c/li\u003e\n\u003cli\u003e\u003cstrong\u003eNotebookLM\u003c/strong\u003e: \u003ccode\u003e@vista8\u003c/code\u003e shared a practice of a multinational small team using it to generate podcasts to synchronize key information with members of different languages, which is considered a novel way of communication alignment.\u003c/li\u003e\n\u003cli\u003e\u003cstrong\u003ePapr\u003c/strong\u003e: \u003ccode\u003e@vista8\u003c/code\u003e recommended a lightweight, fast RSS client that supports AI summarization and Q\u0026amp;A with one\u0026rsquo;s own API Key, improving information acquisition efficiency.\u003c/li\u003e\n\u003cli\u003e\u003cstrong\u003eFigma Chrome Plugin\u003c/strong\u003e: A new tool discovered by \u003ccode\u003e@vista8\u003c/code\u003e that can turn any web element into editable layers and import them into Figma, which is like \u0026ldquo;dimension reduction attack\u0026rdquo; for designers and website replicators.\u003c/li\u003e\n\u003c/ul\u003e\n\u003ch3 id=\"-ai-models-and-skills\"\u003e\n  🧠 AI Models and Skills\n  \u003ca class=\"heading-link\" href=\"#-ai-models-and-skills\"\u003e\n    \u003ci class=\"fa-solid fa-link\" aria-hidden=\"true\" title=\"Link to heading\"\u003e\u003c/i\u003e\n    \u003cspan class=\"sr-only\"\u003eLink to heading\u003c/span\u003e\n  \u003c/a\u003e\n\u003c/h3\u003e\n\u003cul\u003e\n\u003cli\u003e\u003cstrong\u003ePP-OCRv6 (Baidu)\u003c/strong\u003e: Strongly recommended by \u003ccode\u003e@AI_Jasonyu\u003c/code\u003e, an ultra-lightweight OCR model of only 1.5MB that runs extremely fast in the browser, with accuracy even surpassing GPT-5.5 and other large models in some scenarios. This proves that \u003cstrong\u003efor tasks with clear boundaries, clever small models have an advantage over large models.\u003c/strong\u003e\u003c/li\u003e\n\u003cli\u003e\u003cstrong\u003eyao-meta-skill\u003c/strong\u003e: Strongly recommended by \u003ccode\u003e@vista8\u003c/code\u003e for creating Skills, derived from the analysis of official Claude Code source code, which can help users write Skills that score above 90 points.\u003c/li\u003e\n\u003cli\u003e\u003cstrong\u003ebaoyu-design skill\u003c/strong\u003e: A local tool developed by \u003ccode\u003e@dotey\u003c/code\u003e that can not only generate PPTs and animations locally but also export animations directly as mp4 videos. Its underlying declarative animation engine principle allows arbitrary jumps in the timeline without dropping frames, demonstrating the huge potential of AI combined with creative tools.\u003c/li\u003e\n\u003c/ul\u003e\n\u003ch3 id=\"-development-and-deployment\"\u003e\n  💻 Development and Deployment\n  \u003ca class=\"heading-link\" href=\"#-development-and-deployment\"\u003e\n    \u003ci class=\"fa-solid fa-link\" aria-hidden=\"true\" title=\"Link to heading\"\u003e\u003c/i\u003e\n    \u003cspan class=\"sr-only\"\u003eLink to heading\u003c/span\u003e\n  \u003c/a\u003e\n\u003c/h3\u003e\n\u003cul\u003e\n\u003cli\u003e\u003cstrong\u003eVercel Drop\u003c/strong\u003e: Shared by \u003ccode\u003e@vista8\u003c/code\u003e, uploading files/folders directly through the \u003ccode\u003edrop.new\u003c/code\u003e website can generate a temporary website, greatly simplifying sharing and deployment processes.\u003c/li\u003e\n\u003cli\u003e\u003cstrong\u003eDevSpace\u003c/strong\u003e: An open-source project introduced by \u003ccode\u003e@gefei55\u003c/code\u003e, which connects through MCP, allowing ChatGPT web to directly read and write local project code, essentially turning ChatGPT into Codex.\u003c/li\u003e\n\u003c/ul\u003e\n\u003ch2 id=\"-appendix-todays-watch-list-update-sources\"\u003e\n  📚 Appendix: Today\u0026rsquo;s Watch List Update Sources\n  \u003ca class=\"heading-link\" href=\"#-appendix-todays-watch-list-update-sources\"\u003e\n    \u003ci class=\"fa-solid fa-link\" aria-hidden=\"true\" title=\"Link to heading\"\u003e\u003c/i\u003e\n    \u003cspan class=\"sr-only\"\u003eLink to heading\u003c/span\u003e\n  \u003c/a\u003e\n\u003c/h2\u003e\n\u003cblockquote\u003e\n\u003cp\u003eTime window: Last 3 days; 22 sources covered; 34 updates in total.\u003c/p\u003e\n\u003c/blockquote\u003e\n\u003ch3 id=\"stratechery-by-ben-thompson-a_full\"\u003e\n  Stratechery by Ben Thompson (A_full)\n  \u003ca class=\"heading-link\" href=\"#stratechery-by-ben-thompson-a_full\"\u003e\n    \u003ci class=\"fa-solid fa-link\" aria-hidden=\"true\" title=\"Link to heading\"\u003e\u003c/i\u003e\n    \u003cspan class=\"sr-only\"\u003eLink to heading\u003c/span\u003e\n  \u003c/a\u003e\n\u003c/h3\u003e\n\u003cul\u003e\n\u003cli\u003e\u003cstrong\u003e\u003ca href=\"https://stratechery.com/2026/an-interview-with-michael-morton-about-e-commerce-in-the-age-of-ai/\"  class=\"external-link\" target=\"_blank\" rel=\"noopener\"\u003eAn Interview with Michael Morton About E-Commerce in the Age of AI\u003c/a\u003e\u003c/strong\u003e\n\u003cul\u003e\n\u003cli\u003ePublished: 2026-06-18 18:00 Beijing Time\u003c/li\u003e\n\u003cli\u003eSummary: - An interview with Michael Morton about e-commerce and AI, including unfalsifiable bear cases, distribution versus referral models, and the challenges of groceries and self-driving cars.\n\u003cul\u003e\n\u003cli\u003e\u003cstrong\u003e$15\u003c/strong\u003e/month* or *\u003cstrong\u003e$150\u003c/strong\u003e/year.\u003c/li\u003e\n\u003cli\u003eSubstantive analysis of the day\u0026rsquo;s news via three weekly emails or a podcast.\u003c/li\u003e\n\u003cli\u003e\u003cstrong\u003eStrategy Interviews\u003c/strong\u003e.\u003c/li\u003e\n\u003cli\u003eInterviews with leading public company CEOs, private company founders, and discussions with peer analysts.\u003c/li\u003e\n\u003c/ul\u003e\n\u003c/li\u003e\n\u003cli\u003eEN Highlights:\n\u003cul\u003e\n\u003cli\u003eAn interview with Michael Morton about e-commerce and AI, including the challenges of unfalsifiable bear cases, distribution versus referal models, grocery, and…\u003c/li\u003e\n\u003c/ul\u003e\n\u003c/li\u003e\n\u003c/ul\u003e\n\u003c/li\u003e\n\u003c/ul\u003e\n\u003ch3 id=\"openai-blog-a_full\"\u003e\n  OpenAI Blog (A_full)\n  \u003ca class=\"heading-link\" href=\"#openai-blog-a_full\"\u003e\n    \u003ci class=\"fa-solid fa-link\" aria-hidden=\"true\" title=\"Link to heading\"\u003e\u003c/i\u003e\n    \u003cspan class=\"sr-only\"\u003eLink to heading\u003c/span\u003e\n  \u003c/a\u003e\n\u003c/h3\u003e\n\u003cul\u003e\n\u003cli\u003e\n\u003cp\u003e\u003cstrong\u003e\u003ca href=\"https://openai.com/index/chatgpt-enterprise-spend-controls\"  class=\"external-link\" target=\"_blank\" rel=\"noopener\"\u003eNew usage analytics and updated spend controls for enterprises\u003c/a\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cul\u003e\n\u003cli\u003ePublished: 2026-06-19 01:00 Beijing Time\u003c/li\u003e\n\u003cli\u003eSummary: - As AI becomes an integral part of daily work, organizations need to manage it with the same rigor as any critical business investment.\n\u003cul\u003e\n\u003cli\u003eCompanies need clear visibility into usage, adoption, and spending to scale with confidence and understand where AI is creating value.\u003c/li\u003e\n\u003cli\u003eToday, we are introducing credit usage analytics and updated spending controls for ChatGPT Enterprise.\u003c/li\u003e\n\u003cli\u003eThese features help companies track credit usage, understand adoption patterns, and make more informed decisions about how to deploy AI in their organizations.\u003c/li\u003e\n\u003cli\u003eWith clearer visibility and more flexible controls, organizations can proactively manage costs, provide teams with the access they need, and focus AI investments on the most important work.\u003c/li\u003e\n\u003c/ul\u003e\n\u003c/li\u003e\n\u003cli\u003eEN Highlights:\n\u003cul\u003e\n\u003cli\u003eOpenAI introduces new spend controls and usage analytics for ChatGPT Enterprise, helping organizations manage costs and scale AI with confidence.\u003c/li\u003e\n\u003c/ul\u003e\n\u003c/li\u003e\n\u003c/ul\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003e\u003cstrong\u003e\u003ca href=\"https://openai.com/index/improving-health-intelligence-in-chatgpt\"  class=\"external-link\" target=\"_blank\" rel=\"noopener\"\u003eImproving health intelligence in ChatGPT\u003c/a\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cul\u003e\n\u003cli\u003ePublished: 2026-06-18 19:00 Beijing Time\u003c/li\u003e\n\u003cli\u003eSummary: - Health is one of the most meaningful ways people use ChatGPT.\n\u003cul\u003e\n\u003cli\u003eEvery week, over 230 million people turn to ChatGPT for help with health and wellness issues: understanding medical information, interpreting lab results, preparing for appointments, learning about insurance, building healthier habits, and figuring out what to ask next.\u003c/li\u003e\n\u003cli\u003eWith GPT-5.5 Instant, we are seeing a substantial step forward in health capabilities, with improvements in identifying when emergency care might be needed, asking for relevant context, explaining uncertainty, and making complex information more understandable.\u003c/li\u003e\n\u003cli\u003eIn our most challenging health assessments, GPT-5.5 Instant now performs at a level comparable to our frontier models.\u003c/li\u003e\n\u003cli\u003eAs it is available to all free users in ChatGPT, more people can benefit from these improvements.\u003c/li\u003e\n\u003c/ul\u003e\n\u003c/li\u003e\n\u003cli\u003eEN Highlights:\n\u003cul\u003e\n\u003cli\u003eLearn how GPT-5.5 Instant improves ChatGPT’s health and wellness responses with stronger reasoning, better context, clearer communication, and physician-informe…\u003c/li\u003e\n\u003c/ul\u003e\n\u003c/li\u003e\n\u003c/ul\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003e\u003cstrong\u003e\u003ca href=\"https://openai.com/index/diagnose-rare-childhood-diseases\"  class=\"external-link\" target=\"_blank\" rel=\"noopener\"\u003eUsing AI to help physicians diagnose rare genetic diseases affecting children\u003c/a\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cul\u003e\n\u003cli\u003ePublished: 2026-06-18 16:00 Beijing Time\u003c/li\u003e\n\u003c/ul\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003eSummary: Researchers used an OpenAI reasoning model to help diagnose rare diseases, identifying 18 new diagnoses in previously unsolved cases.\u003c/p\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003eThis article from the OpenAI blog explains how using AI to help doctors diagnose rare genetic diseases affecting children is shaping the broader AI and infrastructure landscape.\u003c/p\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003eFollowing the use of AI to assist doctors in diagnosing rare genetic diseases in children, it also presents practical implications for founders, operators, and investors.\u003c/p\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003eEN Highlights:\u003c/p\u003e\n\u003cul\u003e\n\u003cli\u003eResearchers used an OpenAI reasoning model to help diagnose rare diseases, identifying 18 new diagnoses in previously unsolved cases.\u003c/li\u003e\n\u003c/ul\u003e\n\u003c/li\u003e\n\u003c/ul\u003e\n\u003ch3 id=\"arxiv-csai-b_introsearch\"\u003e\n  ArXiv cs.AI (B_intro+search)\n  \u003ca class=\"heading-link\" href=\"#arxiv-csai-b_introsearch\"\u003e\n    \u003ci class=\"fa-solid fa-link\" aria-hidden=\"true\" title=\"Link to heading\"\u003e\u003c/i\u003e\n    \u003cspan class=\"sr-only\"\u003eLink to heading\u003c/span\u003e\n  \u003c/a\u003e\n\u003c/h3\u003e\n\u003cul\u003e\n\u003cli\u003e\n\u003cp\u003e\u003cstrong\u003e\u003ca href=\"https://arxiv.org/abs/2606.17209\"  class=\"external-link\" target=\"_blank\" rel=\"noopener\"\u003eBeyond Parallel Sampling: Diverse Query Initialization for Agentic Search\u003c/a\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cul\u003e\n\u003cli\u003eRelease Time: 2026-06-18 12:00 Beijing Time\u003c/li\u003e\n\u003cli\u003eSummary: - arXiv:2606.17209v1 Announce Type: new.\n\u003cul\u003e\n\u003cli\u003eAbstract: Test-time scaling for agentic search typically increases depth (i.e., more turns and tokens per trajectory) or breadth (i.e., more parallel rollouts).\u003c/li\u003e\n\u003cli\u003eHere, we focus on breadth scaling, showing that standard parallel sampling yields diminishing returns, tracing this to query redundancy in the first turn.\u003c/li\u003e\n\u003cli\u003eWhen models issue similar first queries across rollouts, the threads retrieve overlapping evidence, and subsequent turns are conditioned on this shared retrieval.\u003c/li\u003e\n\u003c/ul\u003e\n\u003c/li\u003e\n\u003cli\u003eEN Highlights:\n\u003cul\u003e\n\u003cli\u003earXiv:2606.17209v1 Announce Type: new\u003c/li\u003e\n\u003cli\u003eAbstract: Test-time scaling for agentic search typically increases depth (i.e., more turns and tokens per trajectory) or breadth (i.e., more parallel rollouts)\u003c/li\u003e\n\u003cli\u003eHere we focus on breadth scaling, showing that standard parallel sampling yields diminishing returns, tracing this to query redundancy at the first turn\u003c/li\u003e\n\u003cli\u003eWhen models issue similar first queries across rollouts, the threads retrieve overlapping evidence, and subsequent turns are conditioned on this shared retrieva…\u003c/li\u003e\n\u003c/ul\u003e\n\u003c/li\u003e\n\u003c/ul\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003e\u003cstrong\u003e\u003ca href=\"https://arxiv.org/abs/2606.17220\"  class=\"external-link\" target=\"_blank\" rel=\"noopener\"\u003eWhen Rules Learn: A Self-Evolving Agent for Legal Case Retrieval\u003c/a\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cul\u003e\n\u003cli\u003eRelease Time: 2026-06-18 12:00 Beijing Time\u003c/li\u003e\n\u003cli\u003eSummary: - arXiv:2606.17220v1 Announce Type: new.\n\u003cul\u003e\n\u003cli\u003eAbstract: Legal case retrieval remains challenging due to the complexity of legal language and the need for precise lexical alignment between queries and relevant cases.\u003c/li\u003e\n\u003cli\u003eAlthough dense retrieval models have achieved notable progress, empirical studies show that BM25 continues to serve as a strong baseline in this domain.\u003c/li\u003e\n\u003cli\u003eThis motivated us to propose a self-evolving framework for rule-driven query rewriting that enhances BM25 without any parametric training.\u003c/li\u003e\n\u003c/ul\u003e\n\u003c/li\u003e\n\u003cli\u003eEN Highlights:\n\u003cul\u003e\n\u003cli\u003earXiv:2606.17220v1 Announce Type: new\u003c/li\u003e\n\u003cli\u003eAbstract: Legal case retrieval remains challenging due to the complexity of legal language and the need for precise lexical alignment between queries and releva…\u003c/li\u003e\n\u003cli\u003eAlthough dense retrieval models have achieved notable progress, empirical studies show that BM25 continues to serve as a strong baseline in this domain\u003c/li\u003e\n\u003c/ul\u003e\n\u003c/li\u003e\n\u003c/ul\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003eIt motivates us to propose a self-evolving framework for rule-driven query rewriting that enhances BM25 without any parameter training\u003c/p\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003e\u003cstrong\u003e\u003ca href=\"https://arxiv.org/abs/2606.17266\"  class=\"external-link\" target=\"_blank\" rel=\"noopener\"\u003eSkillChain-Gym: A Benchmark for Reskilling-Aware Production-Inventory Control under Disruptions\u003c/a\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cul\u003e\n\u003cli\u003ePublish Time: 2026-06-18 12:00 Beijing Time\u003c/li\u003e\n\u003cli\u003eAbstract: - arXiv:2606.17266v1 Announce Type: new.\n\u003cul\u003e\n\u003cli\u003eAbstract: Production planning increasingly has to treat workforce capability as a decision variable: certifications lapse if skills are not maintained, new products require skills not possessed by the current workforce, and retraining competes for the same worker time needed for production.\u003c/li\u003e\n\u003cli\u003eExisting operations benchmarks usually treat labor as an exogenous factor, while workforce planning models with skills and learning capabilities are rarely released as reusable test platforms.\u003c/li\u003e\n\u003cli\u003eWe introduce SkillChain-Gym, a benchmark specification for reskilling-aware production-inventory control: a single-site environment with stylized worker skill-state dynamics, hard-threshold certifications, forgetting, and capacity-consuming training operations, constrained by the same per-worker time budget as production.\u003c/li\u003e\n\u003c/ul\u003e\n\u003c/li\u003e\n\u003cli\u003eEN Key Points:\n\u003cul\u003e\n\u003cli\u003earXiv:2606.17266v1 Announce Type: new\u003c/li\u003e\n\u003cli\u003eAbstract: Production planning increasingly has to treat workforce capability as a decision variable: certifications lapse when skills are not maintained, new pr…\u003c/li\u003e\n\u003cli\u003eExisting operations benchmarks usually treat labor as exogenous, while workforce-planning models with skills and learning are rarely released as reusable testbe…\u003c/li\u003e\n\u003cli\u003eWe introduce SkillChain-Gym, a benchmark specification for reskilling-aware production-inventory control: a single-site environment with stylized worker skill-s…\u003c/li\u003e\n\u003c/ul\u003e\n\u003c/li\u003e\n\u003c/ul\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003e\u003cstrong\u003e\u003ca href=\"https://arxiv.org/abs/2606.17269\"  class=\"external-link\" target=\"_blank\" rel=\"noopener\"\u003eSkill-Constrained Model Predictive Control for Resilient Manufacturing Supply Chains\u003c/a\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cul\u003e\n\u003cli\u003ePublish Time: 2026-06-18 12:00 Beijing Time\u003c/li\u003e\n\u003cli\u003eAbstract: - arXiv:2606.17269v1 Announce Type: new.\n\u003cul\u003e\n\u003cli\u003eAbstract: In skill-constrained production-inventory systems, the qualified human capacity available tomorrow depends on training decisions made today: production requires certified workers, certifications decay if not maintained, and training consumes the same scarce worker time needed for production now.\u003c/li\u003e\n\u003cli\u003eWe study a closed-loop skill-constrained model predictive controller that, each shift, solves a finite-horizon mixed-integer program over production, inventory, backlog, and training, with binary forecast certifications, hard production qualification, and an interpretable terminal value that prices certification capacity deficits at the horizon boundary; only first-stage operations are applied before re-planning.\u003c/li\u003e\n\u003cli\u003eIn synthetic, seed-controlled SkillChain-Gym scenarios—announced and surprising new skill shocks, demand shocks, absences, forecast and availability quality patterns, capacity boundary and training rate scans, and negative controls—we evaluate the controller against production-only and maintenance-only ablations, static cross-training insurance schemes, and strong reactive heuristics under ex-ante locked-in configurations and paired statistics.\u003c/li\u003e\n\u003c/ul\u003e\n\u003c/li\u003e\n\u003cli\u003eEN Key Points:\n\u003cul\u003e\n\u003cli\u003earXiv:2606.17269v1 Announce Type: new\u003c/li\u003e\n\u003cli\u003eAbstract: In skill-constrained production-inventory systems, the qualified human capacity available tomorrow depends on training decisions made today: productio…\u003c/li\u003e\n\u003c/ul\u003e\n\u003c/li\u003e\n\u003c/ul\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003eWe study a closed-loop skill-constrained model predictive controller that, at every shift, solves a finite-horizon mixed-integer program over production, invent…\u003c/p\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003eOn synthetic, seed-controlled SkillChain-Gym scenarios - announced and surprise new-skill shocks, demand shocks, absenteeism, forecast- and availability-quality…\u003c/p\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003e\u003cstrong\u003e\u003ca href=\"https://arxiv.org/abs/2606.17289\"  class=\"external-link\" target=\"_blank\" rel=\"noopener\"\u003eNothing from Something: Can a Language Model Discover 0?\u003c/a\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cul\u003e\n\u003cli\u003ePublication Time: 2026-06-18 12:00 Beijing Time\u003c/li\u003e\n\u003cli\u003eAbstract: - arXiv:2606.17289v1 Announcement Type: New.\n\u003cul\u003e\n\u003cli\u003eAbstract: AI systems based on artificial neural networks are being developed with aspirations of pushing the boundary of human mathematical knowledge.\u003c/li\u003e\n\u003cli\u003eA key question for these systems is how much they can reach beyond their training data.\u003c/li\u003e\n\u003cli\u003eMathematical discovery requires a strong form of out-of-distribution generalization; the ability to hypothesize genuinely new—and potentially logically more powerful—mathematical structures.\u003c/li\u003e\n\u003c/ul\u003e\n\u003c/li\u003e\n\u003cli\u003eEN Highlights:\n\u003cul\u003e\n\u003cli\u003earXiv:2606.17289v1 Announce Type: new\u003c/li\u003e\n\u003cli\u003eAbstract: AI systems based on artificial neural networks are being developed with aspirations of pushing the boundary of human mathematical knowledge\u003c/li\u003e\n\u003cli\u003eA key question for these systems is how much they can reach beyond their training data\u003c/li\u003e\n\u003cli\u003eMathematical discovery requires a strong form of out of distribution generalization; the ability to hypothesize genuinely new - and potentially logically more p…\u003c/li\u003e\n\u003c/ul\u003e\n\u003c/li\u003e\n\u003c/ul\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003e\u003cstrong\u003e\u003ca href=\"https://arxiv.org/abs/2606.17312\"  class=\"external-link\" target=\"_blank\" rel=\"noopener\"\u003eQuantifying Consistency in LLM Logical Reasoning via Structural Uncertainty\u003c/a\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cul\u003e\n\u003cli\u003ePublication Time: 2026-06-18 12:00 Beijing Time\u003c/li\u003e\n\u003cli\u003eAbstract: - arXiv:2606.17312v1 Announcement Type: New.\n\u003cul\u003e\n\u003cli\u003eAbstract: Large language models can arrive at the same answer through reasoning paths that are unstable, contradictory, or difficult to rank consistently—this failure mode is particularly prevalent in multi-step deductive reasoning.\u003c/li\u003e\n\u003cli\u003eExisting methods primarily assess reliability through output dispersion (measuring how much sampled answers differ), but this discards a complementary signal: whether the model can consistently rank competing reasoning candidates.\u003c/li\u003e\n\u003cli\u003eWe propose structural uncertainty, a consistency-aware framework derived from the stability of rankings over sampled reasoning solutions induced by self-preferences.\u003c/li\u003e\n\u003c/ul\u003e\n\u003c/li\u003e\n\u003cli\u003eEN Highlights:\n\u003cul\u003e\n\u003cli\u003earXiv:2606.17312v1 Announce Type: new\u003c/li\u003e\n\u003cli\u003eAbstract: Large language models can arrive at the same answer through reasoning paths that are unstable, contradictory, or difficult to rank consistently \u0026ndash; a f…\u003c/li\u003e\n\u003cli\u003eExisting methods assess reliability primarily through output dispersion \u0026ndash; measuring how much sampled answers differ \u0026ndash; but this discards a complementary signal…\u003c/li\u003e\n\u003c/ul\u003e\n\u003c/li\u003e\n\u003c/ul\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003eWe propose structural uncertainty, a consistency-aware framework derived from the stability of self-preference-induced rankings over sampled reasoning solutions\u003c/p\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003e\u003cstrong\u003e\u003ca href=\"https://arxiv.org/abs/2606.17328\"  class=\"external-link\" target=\"_blank\" rel=\"noopener\"\u003eMemTrace: Probing What Final Accuracy Misses in Long-Term Memory\u003c/a\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cul\u003e\n\u003cli\u003ePublish Time: 2026-06-18 12:00 Beijing Time\u003c/li\u003e\n\u003cli\u003eAbstract: - arXiv:2606.17328v1 Announcement Type: new.\n\u003cul\u003e\n\u003cli\u003eAbstract: LLM agents increasingly maintain long-term memory of user facts across sessions.\u003c/li\u003e\n\u003cli\u003eHowever, such memory is usually evaluated by aggregating accuracy over question rows or episodes.\u003c/li\u003e\n\u003cli\u003eBecause this approach scores question rows independently, even when multiple questions probe the same fact, it cannot show how that fact behaves as conditions change.\u003c/li\u003e\n\u003c/ul\u003e\n\u003c/li\u003e\n\u003cli\u003eEN Key Points:\n\u003cul\u003e\n\u003cli\u003earXiv:2606.17328v1 Announce Type: new\u003c/li\u003e\n\u003cli\u003eAbstract: LLM agents increasingly maintain long-term memory of user facts across sessions\u003c/li\u003e\n\u003cli\u003eYet such memory is usually evaluated by aggregating accuracy over question rows or episodes\u003c/li\u003e\n\u003cli\u003eBecause this approach scores question rows independently, even when several questions probe the same fact, it cannot show how that fact behaves as conditions ch…\u003c/li\u003e\n\u003c/ul\u003e\n\u003c/li\u003e\n\u003c/ul\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003e\u003cstrong\u003e\u003ca href=\"https://arxiv.org/abs/2606.17339\"  class=\"external-link\" target=\"_blank\" rel=\"noopener\"\u003eSpeechDx: A Multi-Task Benchmark for Clinical Speech AI\u003c/a\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cul\u003e\n\u003cli\u003ePublish Time: 2026-06-18 12:00 Beijing Time\u003c/li\u003e\n\u003cli\u003eAbstract: - arXiv:2606.17339v1 Announcement Type: new.\n\u003cul\u003e\n\u003cli\u003eAbstract: Speech offers a uniquely informative window into health by simultaneously engaging neurological, motor, respiratory, and vocal systems.\u003c/li\u003e\n\u003cli\u003eCurrent clinical speech AI methods have largely progressed through isolated condition-specific studies, which makes results difficult to compare and generalizability difficult to assess.\u003c/li\u003e\n\u003cli\u003eWe introduce SpeechDx, a large-scale benchmark for clinical speech AI spanning 12 datasets and 27 tasks across diverse health conditions.\u003c/li\u003e\n\u003c/ul\u003e\n\u003c/li\u003e\n\u003cli\u003eEN Key Points:\n\u003cul\u003e\n\u003cli\u003earXiv:2606.17339v1 Announce Type: new\u003c/li\u003e\n\u003cli\u003eAbstract: Speech offers a uniquely informative window into health by simultaneously engaging neurological, motor, respiratory, and vocal systems\u003c/li\u003e\n\u003cli\u003eCurrent clinical speech AI methods have largely progressed through isolated condition-specific studies, making results difficult to compare and generalization d…\u003c/li\u003e\n\u003cli\u003eWe introduce SpeechDx, a large-scale benchmark for clinical speech AI spanning 12 datasets and 27 tasks across diverse health conditions\u003c/li\u003e\n\u003c/ul\u003e\n\u003c/li\u003e\n\u003c/ul\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003e\u003cstrong\u003e\u003ca href=\"https://arxiv.org/abs/2606.17368\"  class=\"external-link\" target=\"_blank\" rel=\"noopener\"\u003eDistributed General-Purpose Agent Networks: Architecture, Key Mechanisms, and Prototypes\u003c/a\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cul\u003e\n\u003cli\u003ePublish Time: 2026-06-18 12:00 Beijing Time\u003c/li\u003e\n\u003cli\u003eAbstract: - arXiv:2606.17368v1 Announcement Type: new.\n\u003cul\u003e\n\u003cli\u003eAbstract: Large Language Models have accelerated the shift from passive conversational assistants to autonomous agents capable of understanding goals, planning actions, invoking tools, and executing multi-step tasks.\u003c/li\u003e\n\u003c/ul\u003e\n\u003c/li\u003e\n\u003c/ul\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003eHowever, the capability of a single agent remains constrained by its local data, tool permissions, runtime environment, and governance boundary.\u003c/p\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003eThis paper studies distributed general-purpose agent networks: open peer-to-peer networks in which heterogeneous agents deployed on personal devices, edge nodes, or autonomous computing environments can discover each other, establish trust, negotiate rules of cooperation, and execute open-ended tasks.\u003c/p\u003e\n\u003cul\u003e\n\u003cli\u003eEN Highlights:\n\u003cul\u003e\n\u003cli\u003earXiv:2606.17368v1 Announce Type: new\u003c/li\u003e\n\u003cli\u003eAbstract: Large language models have accelerated the transition from passive conversational assistants to autonomous agents that can understand goals, plan acti…\u003c/li\u003e\n\u003cli\u003eYet the capability of a single agent remains constrained by its local data, tool permissions, runtime environment, and governance boundary\u003c/li\u003e\n\u003cli\u003eThis paper studies distributed general-purpose agent networks: open peer-to-peer networks in which heterogeneous agents deployed on personal devices, edge nodes…\u003c/li\u003e\n\u003c/ul\u003e\n\u003c/li\u003e\n\u003c/ul\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003e\u003cstrong\u003e\u003ca href=\"https://arxiv.org/abs/2606.17405\"  class=\"external-link\" target=\"_blank\" rel=\"noopener\"\u003eTreatment Response Optimized Clinical Decision Support AI System via Digital Twin Simulation\u003c/a\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cul\u003e\n\u003cli\u003ePublication Time: 2026-06-18 12:00 Beijing Time\u003c/li\u003e\n\u003cli\u003eAbstract: - arXiv:2606.17405v1 Announce Type: new.\u003c/li\u003e\n\u003cli\u003eAbstract: Clinical decision support AI systems (CDSAS) must adapt to evolving patient conditions in real-time while adhering to strict safety constraints.\u003c/li\u003e\n\u003cli\u003eWe propose an online adaptive framework that integrates Treatment Effect (TE) estimation to quantify clinical benefits, a patient Digital Twin (DT) to simulate treatment trajectories, and Reinforcement Learning (RL) for sequential decision-making.\u003c/li\u003e\n\u003cli\u003eThe AI system is initially trained on historical medical records and operates in a continuous learning loop.\u003c/li\u003e\n\u003cli\u003eEN Highlights:\n\u003cul\u003e\n\u003cli\u003earXiv:2606.17405v1 Announce Type: new\u003c/li\u003e\n\u003cli\u003eAbstract: Clinical decision support AI systems (CDSASs) must adapt to evolving patient conditions in real-time while adhering to strict safety constraints\u003c/li\u003e\n\u003cli\u003eWe present an online adaptive framework that integrates Treatment Effect (TE) estimation to quantify clinical benefits, a patient Digital Twin (DT) to simulate…\u003c/li\u003e\n\u003cli\u003eThe AI system is initially trained on historical medical records and operates in a continuous learning loop\u003c/li\u003e\n\u003c/ul\u003e\n\u003c/li\u003e\n\u003c/ul\u003e\n\u003c/li\u003e\n\u003c/ul\u003e\n\u003ch3 id=\"arxiv-cscl-b_introsearch\"\u003e\n  ArXiv cs.CL (B_intro+search)\n  \u003ca class=\"heading-link\" href=\"#arxiv-cscl-b_introsearch\"\u003e\n    \u003ci class=\"fa-solid fa-link\" aria-hidden=\"true\" title=\"Link to heading\"\u003e\u003c/i\u003e\n    \u003cspan class=\"sr-only\"\u003eLink to heading\u003c/span\u003e\n  \u003c/a\u003e\n\u003c/h3\u003e\n\u003cul\u003e\n\u003cli\u003e\n\u003cp\u003e\u003cstrong\u003e\u003ca href=\"https://arxiv.org/abs/2606.18273\"  class=\"external-link\" target=\"_blank\" rel=\"noopener\"\u003eContinuous Audio Thinking for Large Audio Language Models\u003c/a\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cul\u003e\n\u003cli\u003ePublication Time: 2026-06-18 12:00 Beijing Time\u003c/li\u003e\n\u003cli\u003eAbstract: - arXiv:2606.18273v1 Announce Type: new.\u003c/li\u003e\n\u003cli\u003eAbstract: Large Audio Language Models (LALMs) have demonstrated impressive capabilities in various audio understanding tasks, from speech transcription to music analysis.\u003c/li\u003e\n\u003cli\u003eHowever, as LALMs are typically trained to produce text-aligned responses, their hidden states are gradually shaped for text generation rather than preserving acoustic information.\u003c/li\u003e\n\u003cli\u003eConsequently, the diverse acoustic content carried by the audio (e.g., speech details, prosody, sound events, emotions, and tones) is lost along the way and becomes difficult to leverage in the response.\u003c/li\u003e\n\u003cli\u003eEN Highlights:\n\u003cul\u003e\n\u003cli\u003earXiv:2606.18273v1 Announce Type: new\u003c/li\u003e\n\u003c/ul\u003e\n\u003c/li\u003e\n\u003c/ul\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003eAbstract: Large audio language models (LALMs) have shown impressive capabilities on diverse audio understanding tasks, ranging from speech transcription to musi…\u003c/p\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003eHowever, because LALMs are typically trained to produce text-aligned responses, their hidden states are progressively shaped for text generation rather than for…\u003c/p\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003eAs a result, the diverse acoustic content that audio carries, such as phonetic detail, prosody, sound events, affect, and pitch, is lost along the way and diffi…\u003c/p\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003e\u003cstrong\u003e\u003ca href=\"https://arxiv.org/abs/2606.18372\"  class=\"external-link\" target=\"_blank\" rel=\"noopener\"\u003eRedact or Keep? A Fully Local AI Cascade for Educational Dialogue De-Identification\u003c/a\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cul\u003e\n\u003cli\u003ePublished: 2026-06-18 12:00 Beijing Time\u003c/li\u003e\n\u003cli\u003eAbstract: - arXiv:2606.18372v1 Announcement Type: New.\n\u003cul\u003e\n\u003cli\u003eAbstract: Educational dialogue is a valuable but sensitive resource for research: the same transcripts that capture authentic learning often capture personally identifiable information (PII) entangled with course content, where \u0026ldquo;Riemann\u0026rdquo; could refer to a real student or a mathematical concept.\u003c/li\u003e\n\u003cli\u003eExisting approaches force a trade-off between governance and accuracy.\u003c/li\u003e\n\u003cli\u003eCommercial Large Language Models (LLMs) can handle this ambiguity but require sending student data to third parties, while local named entity recognition (NER) systems preserve governance but over-redact course terminology.\u003c/li\u003e\n\u003c/ul\u003e\n\u003c/li\u003e\n\u003cli\u003eEN Key Points:\n\u003cul\u003e\n\u003cli\u003earXiv:2606.18372v1 Announce Type: new\u003c/li\u003e\n\u003cli\u003eAbstract: Educational dialogue is a valuable but sensitive resource for research: the same transcripts that capture authentic learning often capture personally…\u003c/li\u003e\n\u003cli\u003eExisting approaches force a tradeoff between governance and accuracy\u003c/li\u003e\n\u003cli\u003eCommercial Large Language Models (LLMs) can handle this ambiguity but require sending student data to third parties, while local named entity recognition (NER)…\u003c/li\u003e\n\u003c/ul\u003e\n\u003c/li\u003e\n\u003c/ul\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003e\u003cstrong\u003e\u003ca href=\"https://arxiv.org/abs/2606.18381\"  class=\"external-link\" target=\"_blank\" rel=\"noopener\"\u003eSproutRAG: Attention-Guided Tree Search with Progressive Embeddings for Long-Document RAG\u003c/a\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cul\u003e\n\u003cli\u003ePublished: 2026-06-18 12:00 Beijing Time\u003c/li\u003e\n\u003cli\u003eAbstract: - arXiv:2606.18381v1 Announcement Type: New.\n\u003cul\u003e\n\u003cli\u003eAbstract: Retrieval-Augmented Generation (RAG) systems must balance retrieval granularity and contextual coherence, a challenge that existing methods address through LLM-guided chunking, single-level context expansion, or hierarchical summarization.\u003c/li\u003e\n\u003cli\u003eThese methods variously rely on costly LLM calls during indexing or retrieval, limit context aggregation to a single level of granularity, or introduce information loss through summarization.\u003c/li\u003e\n\u003cli\u003eWe propose SproutRAG, an attention-guided hierarchical RAG framework that resolves this trade-off by building a binary chunking tree using learned inter-sentence attention to organize sentence-level chunks into progressively larger yet semantically coherent units.\u003c/li\u003e\n\u003c/ul\u003e\n\u003c/li\u003e\n\u003cli\u003eEN Key Points:\n\u003cul\u003e\n\u003cli\u003earXiv:2606.18381v1 Announce Type: new\u003c/li\u003e\n\u003cli\u003eAbstract: Retrieval-augmented generation (RAG) systems must balance retrieval granularity with contextual coherence, a challenge that existing methods address t…\u003c/li\u003e\n\u003c/ul\u003e\n\u003c/li\u003e\n\u003c/ul\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003eThese approaches variously depend on costly LLM calls during indexing or retrieval, limit context aggregation to a single granularity level, or introduce inform…\u003c/p\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003eWe present SproutRAG, an attention-guided hierarchical RAG framework that addresses this trade-off by organizing sentence-level chunks into progressively larger…\u003c/p\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003e\u003cstrong\u003e\u003ca href=\"https://arxiv.org/abs/2606.18389\"  class=\"external-link\" target=\"_blank\" rel=\"noopener\"\u003eWant Better Synthetic Data? Steer It: Activation Steering for Low-Resource Language Generation\u003c/a\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cul\u003e\n\u003cli\u003ePublication Time: 2026-06-18 12:00 Beijing Time\u003c/li\u003e\n\u003cli\u003eAbstract: - arXiv:2606.18389v1 Announcement Type: new.\n\u003cul\u003e\n\u003cli\u003eAbstract: Large Language Models (LLMs) have become an effective tool for synthetic data generation, including for low-resource languages, where generated data can improve downstream task performance.\u003c/li\u003e\n\u003cli\u003eCurrent best-performing approaches typically rely on few-shot prompting with target-language examples, which increases inference costs and may reduce diversity due to lexical anchoring.\u003c/li\u003e\n\u003cli\u003eIn this work, we investigate activation steering as an alternative for low-resource synthetic data generation.\u003c/li\u003e\n\u003c/ul\u003e\n\u003c/li\u003e\n\u003cli\u003eEN Key Points:\n\u003cul\u003e\n\u003cli\u003earXiv:2606.18389v1 Announce Type: new\u003c/li\u003e\n\u003cli\u003eAbstract: Large language models (LLMs) have become an effective tool for synthetic data generation, including for low-resource languages, where generated data c…\u003c/li\u003e\n\u003cli\u003eCurrent best-performing approaches typically rely on few-shot prompting with target-language examples, which increases inference costs and may reduce diversity…\u003c/li\u003e\n\u003cli\u003eIn this work, we investigate activation steering as an alternative for low-resource synthetic data generation\u003c/li\u003e\n\u003c/ul\u003e\n\u003c/li\u003e\n\u003c/ul\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003e\u003cstrong\u003e\u003ca href=\"https://arxiv.org/abs/2606.18394\"  class=\"external-link\" target=\"_blank\" rel=\"noopener\"\u003eJetFlow: Breaking the Scaling Ceiling of Speculative Decoding with Parallel Tree Drafting\u003c/a\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cul\u003e\n\u003cli\u003ePublication Time: 2026-06-18 12:00 Beijing Time\u003c/li\u003e\n\u003cli\u003eAbstract: - arXiv:2606.18394v1 Announcement Type: new.\n\u003cul\u003e\n\u003cli\u003eAbstract: Speculative Decoding (SD) accelerates autoregressive Large Language Models (LLMs) by drafting multiple tokens and verifying them in parallel, but it faces scaling limitations: increasing the draft budget only improves speed when acceptance remains high and drafting overhead remains low.\u003c/li\u003e\n\u003cli\u003eThis ceiling has been difficult to break because prior head-based SD methods face a causality-efficiency dilemma.\u003c/li\u003e\n\u003cli\u003eAutoregressive drafters generate path-conditioned candidates, which are effective for tree speculative decoding with higher acceptance lengths, but their drafting cost grows with tree depth.\u003c/li\u003e\n\u003c/ul\u003e\n\u003c/li\u003e\n\u003cli\u003eEN Key Points:\n\u003cul\u003e\n\u003cli\u003earXiv:2606.18394v1 Announce Type: new\u003c/li\u003e\n\u003cli\u003eAbstract: Speculative decoding (SD) accelerates autoregressive Large Language Models (LLMs) by drafting multiple tokens and verifying them in parallel, but it f…\u003c/li\u003e\n\u003cli\u003eThis ceiling has been difficult to break because prior head-based SD methods face a causality-efficiency dilemma\u003c/li\u003e\n\u003c/ul\u003e\n\u003c/li\u003e\n\u003c/ul\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003eAutoregressive drafters produce path-conditioned candidates that are effective for tree speculative decoding with higher acceptance length, but their drafting c…\u003c/p\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003e\u003cstrong\u003e\u003ca href=\"https://arxiv.org/abs/2606.18406\"  class=\"external-link\" target=\"_blank\" rel=\"noopener\"\u003eCoreMem: Riemannian Retrieval and Fisher-Guided Distillation for Long-Term Memory in Dialogue Agents\u003c/a\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cul\u003e\n\u003cli\u003ePublished: 2026-06-18 12:00 Beijing Time\u003c/li\u003e\n\u003cli\u003eAbstract: - arXiv:2606.18406v1 Announcement Type: new.\n\u003cul\u003e\n\u003cli\u003eAbstract: Personalized dialogue agents require continuous long-term memory to maintain coherent interactions across multiple sessions.\u003c/li\u003e\n\u003cli\u003eHowever, deploying these capabilities on consumer-grade hardware (e.g., 8 GB VRAM edge devices) introduces severe memory and compute bottlenecks.\u003c/li\u003e\n\u003cli\u003eExisting systems typically rely on isotropic cosine similarity for retrieval and heuristic rules for context compression.\u003c/li\u003e\n\u003c/ul\u003e\n\u003c/li\u003e\n\u003cli\u003eEN Highlights:\n\u003cul\u003e\n\u003cli\u003earXiv:2606.18406v1 Announce Type: new\u003c/li\u003e\n\u003cli\u003eAbstract: Personalized dialogue agents require continuous long-term memory to maintain coherent interactions across multiple sessions\u003c/li\u003e\n\u003cli\u003eHowever, deploying these capabilities on consumer-grade hardware (e.g., 8 GB VRAM edge devices) introduces severe memory and compute bottlenecks\u003c/li\u003e\n\u003cli\u003eExisting systems typically rely on isotropic cosine similarity for retrieval and heuristic rules for context compression\u003c/li\u003e\n\u003c/ul\u003e\n\u003c/li\u003e\n\u003c/ul\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003e\u003cstrong\u003e\u003ca href=\"https://arxiv.org/abs/2606.18448\"  class=\"external-link\" target=\"_blank\" rel=\"noopener\"\u003eVISUALSKILL: Multimodal Skills for Computer-Use Agents\u003c/a\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cul\u003e\n\u003cli\u003ePublished: 2026-06-18 12:00 Beijing Time\u003c/li\u003e\n\u003cli\u003eAbstract: - arXiv:2606.18448v1 Announcement Type: new.\n\u003cul\u003e\n\u003cli\u003eAbstract: Computer-use agents (CUAs) approach human-level performance on standardised benchmarks but still struggle on long-horizon tasks and unseen software.\u003c/li\u003e\n\u003cli\u003eExisting skill libraries address this with reusable skills, but represent the skill artifact as text only, despite the visual nature of GUI interaction.\u003c/li\u003e\n\u003cli\u003eWe propose VISUALSKILL: a hierarchical multimodal skill, tailored to each target application and organised as a central index over per-topic files, which the agent uses via the load_topic MCP tool to retrieve text and graphics for relevant topics on demand.\u003c/li\u003e\n\u003c/ul\u003e\n\u003c/li\u003e\n\u003cli\u003eEN Highlights:\n\u003cul\u003e\n\u003cli\u003earXiv:2606.18448v1 Announce Type: new\u003c/li\u003e\n\u003cli\u003eAbstract: Computer-use agents (CUAs) approach human-level performance on standardised benchmarks but still struggle on long-horizon tasks and unseen software\u003c/li\u003e\n\u003cli\u003eExisting skill libraries address this with reusable skills, but represent the skill artifact as text only, despite the visual nature of GUI interaction\u003c/li\u003e\n\u003cli\u003eWe propose VISUALSKILL: a hierarchical multimodal skill, tailored to each target application and organised as a central index over per-topic files, which the ag…\u003c/li\u003e\n\u003c/ul\u003e\n\u003c/li\u003e\n\u003c/ul\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003e\u003cstrong\u003e\u003ca href=\"https://arxiv.org/abs/2606.18453\"  class=\"external-link\" target=\"_blank\" rel=\"noopener\"\u003eLLM Parameters for Math Across Languages: Shared or Separate?\u003c/a\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003ePublication Time: 2026-06-18 12:00 Beijing Time\u003c/p\u003e\n\u003cul\u003e\n\u003cli\u003eAbstract: - arXiv:2606.18453v1 Announce Type: new.\n\u003cul\u003e\n\u003cli\u003eAbstract: Large language models (LLMs) exhibit substantial cross-lingual variation in mathematical reasoning performance, but it remains unclear whether these differences reflect language-specific parameters or shared mechanisms that manifest differently across languages.\u003c/li\u003e\n\u003cli\u003eWe present a cross-lingual mechanistic analysis of mathematical reasoning in LLMs, enabling us to localize and compare the model parameters that support cross-lingual mathematical reasoning.\u003c/li\u003e\n\u003cli\u003eWe find that the extracted math-associated parameters exhibit partial cross-lingual overlap, with the strongest overlap concentrated in the intermediate model layers.\u003c/li\u003e\n\u003c/ul\u003e\n\u003c/li\u003e\n\u003cli\u003eEN Highlights:\n\u003cul\u003e\n\u003cli\u003earXiv:2606.18453v1 Announce Type: new\u003c/li\u003e\n\u003cli\u003eAbstract: Large language models (LLMs) exhibit substantial cross-lingual variation in mathematical reasoning performance, but it remains unclear whether these d…\u003c/li\u003e\n\u003cli\u003eWe present a cross-lingual mechanistic analysis of mathematical reasoning in LLMs, enabling us to localize and compare model parameters that support mathematica…\u003c/li\u003e\n\u003cli\u003eWe find that the extracted math-associated parameters exhibit partial cross-lingual overlap, with the strongest overlap concentrated in intermediate model layer…\u003c/li\u003e\n\u003c/ul\u003e\n\u003c/li\u003e\n\u003c/ul\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003e\u003cstrong\u003e\u003ca href=\"https://arxiv.org/abs/2606.18466\"  class=\"external-link\" target=\"_blank\" rel=\"noopener\"\u003eMontreal Forced Aligner and the state of speech-to-text alignment in 2026\u003c/a\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cul\u003e\n\u003cli\u003ePublication Time: 2026-06-18 12:00 Beijing Time\u003c/li\u003e\n\u003cli\u003eAbstract: - arXiv:2606.18466v1 Announce Type: new.\n\u003cul\u003e\n\u003cli\u003eAbstract: The Montreal Forced Aligner (MFA), released in 2016, has since become the most widely used tool for forced alignment in both research and industry.\u003c/li\u003e\n\u003cli\u003eIn the decade since, MFA has undergone substantial development, including expanded coverage across more languages and dialects using larger open-source datasets, unified IPA dictionaries, model adaptation, cross-lingual phone remapping, and supporting utilities.\u003c/li\u003e\n\u003cli\u003eThis paper documents the developments of MFA 3.0 since version 1.0 and evaluates MFA\u0026rsquo;s performance across English, Japanese, and Korean, benchmarked against classic and neural forced aligners.\u003c/li\u003e\n\u003c/ul\u003e\n\u003c/li\u003e\n\u003cli\u003eEN Highlights:\n\u003cul\u003e\n\u003cli\u003earXiv:2606.18466v1 Announce Type: new\u003c/li\u003e\n\u003cli\u003eAbstract: The Montreal Forced Aligner (MFA) was released in 2016 and has since become the most widely used tool for forced alignment in research and industry\u003c/li\u003e\n\u003cli\u003eIn the decade since, MFA has undergone substantial development, including expanded coverage across more languages and dialects using larger open-source datasets…\u003c/li\u003e\n\u003cli\u003eThis paper documents MFA 3.0\u0026rsquo;s developments since version 1.0 and evaluates MFA\u0026rsquo;s performance across English, Japanese, and Korean, benchmarked against classic…\u003c/li\u003e\n\u003c/ul\u003e\n\u003c/li\u003e\n\u003c/ul\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003e\u003cstrong\u003e\u003ca href=\"https://arxiv.org/abs/2606.18471\"  class=\"external-link\" target=\"_blank\" rel=\"noopener\"\u003ePossible or Definite? A Benchmark for Evaluating Diagnostic Uncertainty Preservation in Clinical Text\u003c/a\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cul\u003e\n\u003cli\u003ePublication Time: 2026-06-18 12:00 Beijing Time\u003c/li\u003e\n\u003cli\u003eAbstract: - arXiv:2606.18471v1 Announce Type: new.\n\u003cul\u003e\n\u003cli\u003eAbstract: Large language models (LLMs) are increasingly used for clinical text tasks, such as summarization and revision.\u003c/li\u003e\n\u003c/ul\u003e\n\u003c/li\u003e\n\u003c/ul\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003eWhile most studies evaluate the fluency and coherence of LLM-generated text, whether LLMs correctly preserve diagnostic uncertainty remains underexplored.\u003c/p\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003eIn clinical practice, phrases such as \u0026ldquo;possible pneumonia\u0026rdquo; communicate the strength of available evidence and directly guide decisions about follow-up testing and treatment.\u003c/p\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003eEN Bullet Points:\u003c/p\u003e\n\u003cul\u003e\n\u003cli\u003earXiv:2606.18471v1 Announce Type: new\u003c/li\u003e\n\u003cli\u003eAbstract: Large language models (LLMs) are increasingly used for clinical text tasks such as summarization and revision\u003c/li\u003e\n\u003cli\u003eWhile most studies evaluate the fluency and coherence of LLM-generated text, whether LLMs correctly preserve diagnostic uncertainty remains underexplored\u003c/li\u003e\n\u003cli\u003eIn clinical practice, phrases such as ``possible pneumonia\u0026rsquo;\u0026rsquo; communicate the strength of available evidence and directly guide decisions about follow-up testing…\u003c/li\u003e\n\u003c/ul\u003e\n\u003c/li\u003e\n\u003c/ul\u003e\n\u003ch3 id=\"arxiv-cslg-b_introsearch\"\u003e\n  ArXiv cs.LG (B_intro+search)\n  \u003ca class=\"heading-link\" href=\"#arxiv-cslg-b_introsearch\"\u003e\n    \u003ci class=\"fa-solid fa-link\" aria-hidden=\"true\" title=\"Link to heading\"\u003e\u003c/i\u003e\n    \u003cspan class=\"sr-only\"\u003eLink to heading\u003c/span\u003e\n  \u003c/a\u003e\n\u003c/h3\u003e\n\u003cul\u003e\n\u003cli\u003e\n\u003cp\u003e\u003cstrong\u003e\u003ca href=\"https://arxiv.org/abs/2606.18283\"  class=\"external-link\" target=\"_blank\" rel=\"noopener\"\u003eGaussian Mixture Attention: Linear-Time Sequence Mixing via Probabilistic Latent Routing\u003c/a\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cul\u003e\n\u003cli\u003ePublication Time: 2026-06-18 12:00 Beijing Time\u003c/li\u003e\n\u003cli\u003eAbstract: - arXiv:2606.18283v1 Announce Type: new.\n\u003cul\u003e\n\u003cli\u003eAbstract: The dense token-to-token interaction pattern of standard dot-product attention remains a central bottleneck in scaling Transformer architectures to long contexts.\u003c/li\u003e\n\u003cli\u003eWe introduce \\textbf{Gaussian Mixture Attention (GMA)}, a probabilistic attention-style sequence mixer that replaces explicit pairwise query\u0026ndash;key comparison with routing through $K$ learned Gaussian mixture components.\u003c/li\u003e\n\u003cli\u003eQueries and keys are mapped to posterior \\textit{responsibility} vectors over a shared latent routing space; their overlap defines an implicit responsibility-space association, while values are written to and read from a $K$-slot latent memory.\u003c/li\u003e\n\u003c/ul\u003e\n\u003c/li\u003e\n\u003cli\u003eEN Bullet Points:\n\u003cul\u003e\n\u003cli\u003earXiv:2606.18283v1 Announce Type: new\u003c/li\u003e\n\u003cli\u003eAbstract: The dense token-to-token interaction pattern of standard dot-product attention remains a central bottleneck in scaling Transformer architectures to lo…\u003c/li\u003e\n\u003cli\u003eWe introduce \\textbf{Gaussian Mixture Attention (GMA)}, a probabilistic attention-style sequence mixer that replaces explicit pairwise query\u0026ndash;key comparison wit…\u003c/li\u003e\n\u003cli\u003eQueries and keys are mapped to posterior \\textit{responsibility} vectors over a shared latent routing space; their overlap defines an implicit responsibility-sp…\u003c/li\u003e\n\u003c/ul\u003e\n\u003c/li\u003e\n\u003c/ul\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003e\u003cstrong\u003e\u003ca href=\"https://arxiv.org/abs/2606.18284\"  class=\"external-link\" target=\"_blank\" rel=\"noopener\"\u003eBreaking the Solver Bottleneck: Training Task Generators at the Learnable Frontier\u003c/a\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cul\u003e\n\u003cli\u003ePublication Time: 2026-06-18 12:00 Beijing Time\u003c/li\u003e\n\u003cli\u003eAbstract: - arXiv:2606.18284v1 Announce Type: new.\n\u003cul\u003e\n\u003cli\u003eAbstract: Training agents via reinforcement learning (RL) is increasingly resource-limited by the supply of frontier tasks: effective, solvable tasks sufficient to train current models.\u003c/li\u003e\n\u003cli\u003eAs reasoning and agent models improve, fixed task distributions saturate, while naive synthetic generation produces tasks that are trivial, impossible, or ill-posed.\u003c/li\u003e\n\u003c/ul\u003e\n\u003c/li\u003e\n\u003c/ul\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003eTraining a task generator with RL to optimize for validity and learnability can address this bottleneck, but direct optimization requires repeated solver rollouts per candidate.\u003c/p\u003e\n\u003cul\u003e\n\u003cli\u003eEN Highlights:\n\u003cul\u003e\n\u003cli\u003earXiv:2606.18284v1 Announce Type: new\u003c/li\u003e\n\u003cli\u003eAbstract: The limiting resource for training agents via reinforcement learning (RL) is increasingly frontier task supply: valid, solvable tasks just difficult e…\u003c/li\u003e\n\u003cli\u003eAs reasoning and agentic models improve, fixed task distributions saturate, while naive synthetic generation yields tasks that are trivial, impossible, or ill-p…\u003c/li\u003e\n\u003cli\u003eTraining a task generator with RL to optimize validity and learnability can address this bottleneck, but direct optimization requires repeated solver rollouts p…\u003c/li\u003e\n\u003c/ul\u003e\n\u003c/li\u003e\n\u003c/ul\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003e\u003cstrong\u003e\u003ca href=\"https://arxiv.org/abs/2606.18286\"  class=\"external-link\" target=\"_blank\" rel=\"noopener\"\u003eCODEBLOCK: Learning to Supervise Code at the Right Granularity\u003c/a\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cul\u003e\n\u003cli\u003ePublication Time: 2026-06-18 12:00 Beijing Time\u003c/li\u003e\n\u003cli\u003eAbstract: - arXiv:2606.18286v1 Announce Type: new.\n\u003cul\u003e\n\u003cli\u003eAbstract: Supervised fine-tuning of code LLMs typically applies a uniform cross-entropy loss to all response tokens, implicitly assuming that every token provides an equally useful learning signal.\u003c/li\u003e\n\u003cli\u003eRecent token-level selection methods challenge this assumption in natural-language SFT by supervising only high-value tokens.\u003c/li\u003e\n\u003cli\u003eHowever, directly transferring token-level masking to code can break syntactically and semantically coherent program units, because code depends on structural integrity and define-use relationships.\u003c/li\u003e\n\u003c/ul\u003e\n\u003c/li\u003e\n\u003cli\u003eEN Highlights:\n\u003cul\u003e\n\u003cli\u003earXiv:2606.18286v1 Announce Type: new\u003c/li\u003e\n\u003cli\u003eAbstract: Supervised fine-tuning of code LLMs typically applies uniform cross-entropy loss to all response tokens, implicitly assuming that every token provides…\u003c/li\u003e\n\u003cli\u003eRecent token-level selection methods challenge this assumption in natural-language SFT by supervising only high-value tokens\u003c/li\u003e\n\u003cli\u003eHowever, directly transferring token-level masking to code can break syntactically and semantically coherent program units, because code depends on structural c…\u003c/li\u003e\n\u003c/ul\u003e\n\u003c/li\u003e\n\u003c/ul\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003e\u003cstrong\u003e\u003ca href=\"https://arxiv.org/abs/2606.18287\"  class=\"external-link\" target=\"_blank\" rel=\"noopener\"\u003eArtemis: Anatomy-Resolved inTervention for Eliminating Multimodal NeuroImage confounderS\u003c/a\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cul\u003e\n\u003cli\u003ePublication Time: 2026-06-18 12:00 Beijing Time\u003c/li\u003e\n\u003cli\u003eAbstract: - arXiv:2606.18287v1 Announce Type: new.\n\u003cul\u003e\n\u003cli\u003eAbstract: Multimodal neuroimaging integrates functional connectivity from fMRI and structural connectivity from DTI, enabling non-invasive analysis of brain networks using graph neural networks.\u003c/li\u003e\n\u003cli\u003eHowever, demographic factors like age and sex systemically confound the relationship between brain connectivity and clinical outcomes, causing GNNs to exploit spurious shortcuts instead of learning causally-invariant representations.\u003c/li\u003e\n\u003cli\u003eWhile recent causal GNN methods introduce causality at the graph modeling level, their causal mechanisms remain domain-agnostic, failing to account for the real-world confounders inherent in clinical neuroimaging data.\u003c/li\u003e\n\u003c/ul\u003e\n\u003c/li\u003e\n\u003cli\u003eEN Highlights:\n\u003cul\u003e\n\u003cli\u003earXiv:2606.18287v1 Announce Type: new\u003c/li\u003e\n\u003c/ul\u003e\n\u003c/li\u003e\n\u003c/ul\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003eAbstract: Multimodal neuroimaging, integrating functional connectivity from fMRI and structural connectivity from DTI, enables non-invasive analysis of brain ne…\u003c/p\u003e\n\u003cul\u003e\n\u003cli\u003eHowever, demographic factors such as age and sex systematically confound the relationship between brain connectivity and clinical outcomes, causing GNNs to expl…\u003c/li\u003e\n\u003cli\u003eWhile recent causal GNN methods introduce causality at the graph-modeling level, their causal mechanisms remain domain-agnostic without accounting for the real-…\u003c/li\u003e\n\u003c/ul\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003e\u003cstrong\u003e\u003ca href=\"https://arxiv.org/abs/2606.18303\"  class=\"external-link\" target=\"_blank\" rel=\"noopener\"\u003eA Link between Shock-wave Theory and Symmetry-reduced Stochastic Gradient Descent for Artificial Neural Networks\u003c/a\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cul\u003e\n\u003cli\u003ePublished: 2026-06-18 12:00 Beijing Time\u003c/li\u003e\n\u003cli\u003eAbstract: - arXiv:2606.18303v1 Announcement Type: new.\n\u003cul\u003e\n\u003cli\u003eAbstract: We leverage differential geometry, Lie group theory, and fluid mechanics to establish a mathematically explicit link between shock-wave theory and the symmetry-quotiented learning dynamics of stochastic gradient descent.\u003c/li\u003e\n\u003cli\u003eSpecifically, after quotienting parameter symmetries and applying local-entropy coarse-graining, the effective dynamics satisfy a viscous Hamilton-Jacobi equation on the quotient manifold.\u003c/li\u003e\n\u003cli\u003eFurthermore, assuming the raw parameter dynamics can be summarized by a gradient field on the quotiented space, the gradient of the coarse-grained loss function obeys a Burgers-type equation, and shock formation can be rigorously established.\u003c/li\u003e\n\u003c/ul\u003e\n\u003c/li\u003e\n\u003cli\u003eEN Highlights:\n\u003cul\u003e\n\u003cli\u003earXiv:2606.18303v1 Announce Type: new\u003c/li\u003e\n\u003cli\u003eAbstract: We develop a mathematically explicit link between shock-wave theory and the symmetry-quotiented learning dynamics of stochastic gradient descent, draw…\u003c/li\u003e\n\u003cli\u003eSpecifically, after quotienting parameter symmetries and applying local-entropy coarse-graining, the effective dynamics satisfy a viscous Hamilton\u0026ndash;Jacobi equat…\u003c/li\u003e\n\u003cli\u003eMoreover, under the assumption that the raw parameter dynamics can be summarized by a gradient field on the quotiented space, the gradient of the coarse-grained…\u003c/li\u003e\n\u003c/ul\u003e\n\u003c/li\u003e\n\u003c/ul\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003e\u003cstrong\u003e\u003ca href=\"https://arxiv.org/abs/2606.18304\"  class=\"external-link\" target=\"_blank\" rel=\"noopener\"\u003eAttribution-Guided and Coverage-Maximized Pruning for Structural MoE Compression\u003c/a\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cul\u003e\n\u003cli\u003ePublished: 2026-06-18 12:00 Beijing Time\u003c/li\u003e\n\u003cli\u003eAbstract: - arXiv:2606.18304v1 Announcement Type: new.\n\u003cul\u003e\n\u003cli\u003eAbstract: Mixture-of-Experts (MoE) models can effectively scale computation, but their deployment costs remain high due to their large memory footprint and inference overhead.\u003c/li\u003e\n\u003cli\u003ePrevious compression methods have primarily operated at the expert level, either removing entire experts or ranking them through coarse-grained importance scores.\u003c/li\u003e\n\u003cli\u003eHowever, such expert-wise decisions are often too coarse to capture fine-grained redundancy, leading to improper pruning budget allocation and limited compression.\u003c/li\u003e\n\u003c/ul\u003e\n\u003c/li\u003e\n\u003cli\u003eEN Highlights:\n\u003cul\u003e\n\u003cli\u003earXiv:2606.18304v1 Announce Type: new\u003c/li\u003e\n\u003c/ul\u003e\n\u003c/li\u003e\n\u003c/ul\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003eAbstract: Mixture-of-Experts (MoE) models scale compute efficiently, yet remain expensive to deploy due to their substantial memory footprint and inference over…\u003c/p\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003ePrior compression methods mainly operate at the expert level, either removing entire experts or ranking experts by coarse-grained importance scores\u003c/p\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003eHowever, such expert-wise decisions are often too coarse to capture fine-grained redundancy, leading to misallocated pruning budgets and limited compression\u003c/p\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003e\u003cstrong\u003e\u003ca href=\"https://arxiv.org/abs/2606.18306\"  class=\"external-link\" target=\"_blank\" rel=\"noopener\"\u003eFisher Width: A Geometric Measure of Complexity on Statistical Manifolds\u003c/a\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cul\u003e\n\u003cli\u003ePublished: 2026-06-18 12:00 Beijing Time\u003c/li\u003e\n\u003cli\u003eAbstract: - arXiv:2606.18306v1 Announce Type: new.\n\u003cul\u003e\n\u003cli\u003eAbstract: Gaussian width is a central geometric complexity measure in high-dimensional probability, compressed sensing, convex optimization, and learning theory.\u003c/li\u003e\n\u003cli\u003eIt quantifies the average extent of a set along random directions, thereby capturing the effective dimension of constraint sets, hypothesis classes, and descent cones.\u003c/li\u003e\n\u003cli\u003eHowever, this notion is intrinsically Euclidean.\u003c/li\u003e\n\u003c/ul\u003e\n\u003c/li\u003e\n\u003cli\u003eEN Highlights:\n\u003cul\u003e\n\u003cli\u003earXiv:2606.18306v1 Announce Type: new\u003c/li\u003e\n\u003cli\u003eAbstract: Gaussian width is a central geometric complexity measure in high-dimensional probability, compressed sensing, convex optimization, and learning theory\u003c/li\u003e\n\u003cli\u003eIt quantifies the average extent of a set along random directions, thereby capturing the effective dimension of constraint sets, hypothesis classes, and descent…\u003c/li\u003e\n\u003cli\u003eHowever, this notion is intrinsically Euclidean\u003c/li\u003e\n\u003c/ul\u003e\n\u003c/li\u003e\n\u003c/ul\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003e\u003cstrong\u003e\u003ca href=\"https://arxiv.org/abs/2606.18307\"  class=\"external-link\" target=\"_blank\" rel=\"noopener\"\u003eDRIFT: Refining Instruction Data via On-Policy Data Attribution\u003c/a\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cul\u003e\n\u003cli\u003ePublished: 2026-06-18 12:00 Beijing Time\u003c/li\u003e\n\u003cli\u003eAbstract: - arXiv:2606.18307v1 Announce Type: new.\n\u003cul\u003e\n\u003cli\u003eAbstract: Optimizing the training data distribution for Supervised Fine-Tuning (SFT) dictates the capability of Large Language Models (LLMs).\u003c/li\u003e\n\u003cli\u003eWhile existing data curation methods excel at accelerating training under constrained budgets, they are less suited to elevating the capability upper bound.\u003c/li\u003e\n\u003cli\u003eThe challenge here is no longer to identify a smaller subset that maintains performance, but to refine the data distribution into instances most capable of improving the final model.\u003c/li\u003e\n\u003c/ul\u003e\n\u003c/li\u003e\n\u003cli\u003eEN Highlights:\n\u003cul\u003e\n\u003cli\u003earXiv:2606.18307v1 Announce Type: new\u003c/li\u003e\n\u003cli\u003eAbstract: Optimizing the training data distribution for Supervised Fine-Tuning (SFT) dictates the capability of Large Language Models (LLMs)\u003c/li\u003e\n\u003cli\u003eWhile existing data curation methods excel at accelerating training under constrained budgets, they are less suited to elevating the capability upper bound\u003c/li\u003e\n\u003c/ul\u003e\n\u003c/li\u003e\n\u003c/ul\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003eThe challenge here is no longer to identify a smaller subset that preserves performance, but to refine the data distribution toward instances most capable of im…\u003c/p\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003e\u003cstrong\u003e\u003ca href=\"https://arxiv.org/abs/2606.18308\"  class=\"external-link\" target=\"_blank\" rel=\"noopener\"\u003eTRIDENT: Breaking the Hybrid-Safety-Physics Coupling for Provably Safe Multi-Agent Reinforcement Learning\u003c/a\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cul\u003e\n\u003cli\u003eRelease Time: 2026-06-18 12:00 Beijing Time\u003c/li\u003e\n\u003cli\u003eAbstract: - arXiv:2606.18308v1 Announcement Type: New.\n\u003cul\u003e\n\u003cli\u003eAbstract: Safe coordination in networked cyber-physical systems forces learning algorithms to simultaneously handle hybrid discrete-continuous actions, hard training-time safety constraints, and physical control dynamics.\u003c/li\u003e\n\u003cli\u003eWe show that these three features form a directed cycle of biases that defeats any naive composition of off-the-shelf modules, and formalize this as a three-way coupling lemma.\u003c/li\u003e\n\u003cli\u003eWe then introduce TRIDENT, the first MARL framework whose three components are co-designed to cancel each leak: a Richardson-Romberg gradient correction reducing Gumbel-Softmax bias from O(tau) to O(tau^2), Lyapunov-constrained sequential trust-region updates that enforce per-iteration feasibility, and a physics-informed residual critic that factorizes value instead of rewards.\u003c/li\u003e\n\u003c/ul\u003e\n\u003c/li\u003e\n\u003cli\u003eEN Key Points:\n\u003cul\u003e\n\u003cli\u003earXiv:2606.18308v1 Announce Type: new\u003c/li\u003e\n\u003cli\u003eAbstract: Safe coordination in networked cyber-physical systems forces learning algorithms to simultaneously handle hybrid discrete-continuous actions, hard tra…\u003c/li\u003e\n\u003cli\u003eWe show that these three features form a directed cycle of biases that defeats any naive composition of off-the-shelf modules, and formalize this as a three-way…\u003c/li\u003e\n\u003cli\u003eWe then introduce TRIDENT, the first MARL framework whose three components are co-designed to cancel each leak: a Richardson-Romberg gradient correction reducin…\u003c/li\u003e\n\u003c/ul\u003e\n\u003c/li\u003e\n\u003c/ul\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003e\u003cstrong\u003e\u003ca href=\"https://arxiv.org/abs/2606.18309\"  class=\"external-link\" target=\"_blank\" rel=\"noopener\"\u003eSAGE: Retain-Aware Post-Hoc Sanitization of Final Unlearning Vector\u003c/a\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cul\u003e\n\u003cli\u003eRelease Time: 2026-06-18 12:00 Beijing Time\u003c/li\u003e\n\u003cli\u003eAbstract: - arXiv:2606.18309v1 Announcement Type: New.\n\u003cul\u003e\n\u003cli\u003eAbstract: Large Language Model (LLM) unlearning aims to remove undesirable knowledge or behaviors while preserving retained capabilities.\u003c/li\u003e\n\u003cli\u003eCurrent unlearning methods all involve a trade-off between unlearning and retention.\u003c/li\u003e\n\u003cli\u003eWe find that the retention activation bias can also be used to quantify the damage an unlearning method inflicts on retention, without considering the specific implementation of the unlearning process.\u003c/li\u003e\n\u003c/ul\u003e\n\u003c/li\u003e\n\u003cli\u003eEN Key Points:\n\u003cul\u003e\n\u003cli\u003earXiv:2606.18309v1 Announce Type: new\u003c/li\u003e\n\u003cli\u003eAbstract: Large Language Model (LLM) unlearning aims to remove undesirable knowledge or behaviors while preserving retained capabilities\u003c/li\u003e\n\u003cli\u003eCurrent unlearning methods all involve a trade-off between unlearning and retention\u003c/li\u003e\n\u003cli\u003eWe have found that the retention activation bias can also be used to quantify the damage an unlearning method inflicts on retention, without considering the spe…\u003c/li\u003e\n\u003c/ul\u003e\n\u003c/li\u003e\n\u003c/ul\u003e\n\u003c/li\u003e\n\u003c/ul\u003e\n",
  "wordCount": 7761,
  "readingTime": 37,
  "tableOfContents": "\u003cnav id=\"TableOfContents\"\u003e\n  \u003cul\u003e\n    \u003cli\u003e\u003ca href=\"#-deep-dive-this-issues-watch-list\"\u003e📖 Deep Dive: This Issue\u0026rsquo;s Watch List\u003c/a\u003e\u003c/li\u003e\n    \u003cli\u003e\u003ca href=\"#-quick-brief-ai-hot-topics-on-x\"\u003e🌐 Quick Brief: AI Hot Topics on X\u003c/a\u003e\n      \u003cul\u003e\n        \u003cli\u003e\u003ca href=\"#topic-1-anthropic-launches-artifacts-for-shareable-claude-code-sessions\"\u003eTopic 1: Anthropic Launches Artifacts for Shareable Claude Code Sessions\u003c/a\u003e\u003c/li\u003e\n        \u003cli\u003e\u003ca href=\"#topic-2-aether-ai-raises-20m-for-causal-world-models\"\u003eTopic 2: Aether AI Raises $20M for Causal World Models\u003c/a\u003e\u003c/li\u003e\n        \u003cli\u003e\u003ca href=\"#topic-3-google-ai-pioneer-noam-shazeer-joins-openai-after-leading-gemini\"\u003eTopic 3: Google AI Pioneer Noam Shazeer Joins OpenAI After Leading Gemini\u003c/a\u003e\u003c/li\u003e\n        \u003cli\u003e\u003ca href=\"#topic-4-wispr-flow-ceo-collects-800-user-feedback-replies-for-app-fixes\"\u003eTopic 4: Wispr Flow CEO Collects 800 User Feedback Replies for App Fixes\u003c/a\u003e\u003c/li\u003e\n        \u003cli\u003e\u003ca href=\"#topic-5-trump-says-anthropic-ai-negotiations-progress-amid-export-ban\"\u003eTopic 5: Trump Says Anthropic AI Negotiations Progress Amid Export Ban\u003c/a\u003e\u003c/li\u003e\n      \u003c/ul\u003e\n    \u003c/li\u003e\n    \u003cli\u003e\u003ca href=\"#-influencer-insights\"\u003e💡 Influencer Insights\u003c/a\u003e\u003c/li\u003e\n  \u003c/ul\u003e\n\n  \u003cul\u003e\n    \u003cli\u003e\u003ca href=\"#1-todays-focus-the-rise-of-on-device-models-and-the-evolution-of-ai-programming-tools\"\u003e1. Today\u0026rsquo;s Focus: The Rise of On-Device Models and the Evolution of AI Programming Tools\u003c/a\u003e\n      \u003cul\u003e\n        \u003cli\u003e\u003ca href=\"#on-device-models-the-tipping-point-from-usable-to-great\"\u003eOn-Device Models: The Tipping Point from \u0026ldquo;Usable\u0026rdquo; to \u0026ldquo;Great\u0026rdquo;\u003c/a\u003e\u003c/li\u003e\n        \u003cli\u003e\u003ca href=\"#ai-programming-tools-from-writing-code-to-collaboration-and-automation\"\u003eAI Programming Tools: From \u0026ldquo;Writing Code\u0026rdquo; to \u0026ldquo;Collaboration and Automation\u0026rdquo;\u003c/a\u003e\u003c/li\u003e\n      \u003c/ul\u003e\n    \u003c/li\u003e\n    \u003cli\u003e\u003ca href=\"#2-notable-unique-perspectives-and-industry-foresight\"\u003e2. Notable Unique Perspectives and Industry Foresight\u003c/a\u003e\n      \u003cul\u003e\n        \u003cli\u003e\u003ca href=\"#-re-evaluating-the-next-token-prediction-path\"\u003e① Re-evaluating the \u0026ldquo;Next Token Prediction\u0026rdquo; Path\u003c/a\u003e\u003c/li\u003e\n        \u003cli\u003e\u003ca href=\"#-ai-efficiency-paradox-efficiency-improved-now-what\"\u003e② \u0026ldquo;AI Efficiency Paradox\u0026rdquo;: Efficiency improved, now what?\u003c/a\u003e\u003c/li\u003e\n        \u003cli\u003e\u003ca href=\"#-the-path-from-vibe-coder-to-software-engineer\"\u003e③ The Path from \u0026ldquo;Vibe Coder\u0026rdquo; to \u0026ldquo;Software Engineer\u0026rdquo;\u003c/a\u003e\u003c/li\u003e\n        \u003cli\u003e\u003ca href=\"#-other-forward-looking-perspectives\"\u003e③ Other Forward-Looking Perspectives\u003c/a\u003e\u003c/li\u003e\n      \u003c/ul\u003e\n    \u003c/li\u003e\n    \u003cli\u003e\u003ca href=\"#3-recommended-tools-and-resources\"\u003e3. Recommended Tools and Resources\u003c/a\u003e\n      \u003cul\u003e\n        \u003cli\u003e\u003ca href=\"#-productivity-tools\"\u003e🛠️ Productivity Tools\u003c/a\u003e\u003c/li\u003e\n        \u003cli\u003e\u003ca href=\"#-ai-models-and-skills\"\u003e🧠 AI Models and Skills\u003c/a\u003e\u003c/li\u003e\n        \u003cli\u003e\u003ca href=\"#-development-and-deployment\"\u003e💻 Development and Deployment\u003c/a\u003e\u003c/li\u003e\n      \u003c/ul\u003e\n    \u003c/li\u003e\n    \u003cli\u003e\u003ca href=\"#-appendix-todays-watch-list-update-sources\"\u003e📚 Appendix: Today\u0026rsquo;s Watch List Update Sources\u003c/a\u003e\n      \u003cul\u003e\n        \u003cli\u003e\u003ca href=\"#stratechery-by-ben-thompson-a_full\"\u003eStratechery by Ben Thompson (A_full)\u003c/a\u003e\u003c/li\u003e\n        \u003cli\u003e\u003ca href=\"#openai-blog-a_full\"\u003eOpenAI Blog (A_full)\u003c/a\u003e\u003c/li\u003e\n        \u003cli\u003e\u003ca href=\"#arxiv-csai-b_introsearch\"\u003eArXiv cs.AI (B_intro+search)\u003c/a\u003e\u003c/li\u003e\n        \u003cli\u003e\u003ca href=\"#arxiv-cscl-b_introsearch\"\u003eArXiv cs.CL (B_intro+search)\u003c/a\u003e\u003c/li\u003e\n        \u003cli\u003e\u003ca href=\"#arxiv-cslg-b_introsearch\"\u003eArXiv cs.LG (B_intro+search)\u003c/a\u003e\u003c/li\u003e\n      \u003c/ul\u003e\n    \u003c/li\u003e\n  \u003c/ul\u003e\n\u003c/nav\u003e",
  "isDraft": false
}