{
  "title": "2026-09-10 AI Daily Update | When AI Starts Long-Term Work: Memory Costs, Process Feedback, and Safety Governance Emerge",
  "url": "https://miaok.ong/en/ai-daily/ai-daily-2026-09-10/",
  "date": "2026-09-10T07:00:00+08:00",
  "lastmod": "2026-09-10T07:00:00+08:00",
  "type": "ai-daily",
  "kind": "page",
  "language": "en",
  "description": "The key development today is not about new benchmark scores, but whether AI can stably and cost-effectively complete long-term tasks. New research is beginning to evaluate the marginal value of agent memory on practical execution and transforming generative evaluation into actionable feedback for improvement. Meanwhile, OpenAI is inviting alignment researchers into its foundation\u0026rsquo;s safety governance system, indicating that the capability expansion of frontier models is advancing in tandem with accountability mechanisms.",
  "keywords": null,
  "tags": [],
  "categories": [],
  "author": "Mark (Miao) Kong",
  "image": "https://miaok.ong/images/avatar.jpg",
  "content": "\u003ch1 id=\"2026-09-10-ai-daily--as-ai-starts-long-term-work-memory-costs-process-feedback-and-safety-governance-emerge\"\u003e\n  2026-09-10 AI Daily | As AI Starts Long-Term Work: Memory Costs, Process Feedback, and Safety Governance Emerge\n  \u003ca class=\"heading-link\" href=\"#2026-09-10-ai-daily--as-ai-starts-long-term-work-memory-costs-process-feedback-and-safety-governance-emerge\"\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\u003eThe key shift today isn\u0026rsquo;t about models breaking new score records, but whether AI can stably and cost-effectively complete long-term tasks. New research is beginning to assess the marginal value of agent memory on actual execution and transform generation evaluations into actionable improvement feedback. Meanwhile, OpenAI\u0026rsquo;s invitation for alignment researchers to join its foundation\u0026rsquo;s safety governance system shows that the capability expansion of frontier models is advancing in step with accountability mechanisms.\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 theme to watch today is the shift of AI agents from \u0026ldquo;capability showcases\u0026rdquo; to \u0026ldquo;long-term governance.\u0026rdquo; OpenAI is bringing Paul Christiano into its foundation and safety affairs while releasing GPT-6 Astra for workplace scenarios. Related research further asks: who will maintain agent skills, how to design governance mechanisms for autonomous goals, and why safety monitoring with \u0026ldquo;high recall doesn\u0026rsquo;t equal protection.\u0026rdquo; Meta\u0026rsquo;s release of its personal agent also warrants examination in light of these governance issues.\u003c/p\u003e\n\u003cp\u003eThe second major theme is the refinement of evaluation systems. Medical speech recognition, Tamil speech endpoint detection, and opioid use identification show that real-world scenarios care more about critical errors and reproducibility than a single average score. At the same time, work on RAG query rewriting, parallel decoding, and CriticGen is advancing evaluation from simply \u0026ldquo;judging right or wrong\u0026rdquo; to \u0026ldquo;explaining and improving the generation process.\u0026rdquo;\u003c/p\u003e\n\u003ch2 id=\"-ai-hotspots-on-x\"\u003e\n  🌐 AI Hotspots on X\n  \u003ca class=\"heading-link\" href=\"#-ai-hotspots-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-deepseek-v41-flash-delivers-top-ai-performance-at-low-cost\"\u003e\n  Topic 1: DeepSeek V4.1 Flash Delivers Top AI Performance at Low Cost\n  \u003ca class=\"heading-link\" href=\"#topic-1-deepseek-v41-flash-delivers-top-ai-performance-at-low-cost\"\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 1 day ago, 14,000 related posts\u003c/li\u003e\n\u003cli\u003eWhat it is: According to the headline, DeepSeek V4.1 Flash is reported to have achieved leading AI performance at a low cost.\u003c/li\u003e\n\u003cli\u003eWhy it matters: This pertains to the inference costs of high-performance AI models, their commercial accessibility, and the impact of open-source or low-cost models on the existing market landscape.\u003c/li\u003e\n\u003cli\u003eDiscussion summary: Discussions on X likely focus on the reliability of its performance benchmarks, the actual extent of its cost advantage, comparisons with other mainstream models, and whether this result is sustainable. Specific points of disagreement cannot be further confirmed as no representative posts were provided.\u003c/li\u003e\n\u003c/ul\u003e\n\u003ch3 id=\"topic-2-ai-researcher-quits-anthropic-over-superintelligence-risks\"\u003e\n  Topic 2: AI Researcher Quits Anthropic Over Superintelligence Risks\n  \u003ca class=\"heading-link\" href=\"#topic-2-ai-researcher-quits-anthropic-over-superintelligence-risks\"\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 23 hours ago, 723,000 related posts\u003c/li\u003e\n\u003cli\u003eWhat it is: Anthropic researcher Jacob Coxon resigned over concerns that self-improving superintelligence could lead to loss of control and human extinction, publicly criticizing the industry\u0026rsquo;s race to build it.\u003c/li\u003e\n\u003cli\u003eWhy it matters: The event highlights the conflict within frontier AI labs between capability competition, alignment research, risk disclosure, and regulatory responsibility. It also brings the probability of superintelligence risks and the adequacy of existing safety mechanisms back into the public debate.\u003c/li\u003e\n\u003cli\u003eDiscussion summary: Discussions on X primarily revolve around whether Coxon\u0026rsquo;s warning is credible, Anthropic\u0026rsquo;s probability assessment of future AI-induced disasters, and whether companies are weakening safety regulations due to competition and lobbying. The main point of disagreement is whether to immediately slow down or strengthen regulation, or to continue development while relying on safety research to mitigate risks.\u003c/li\u003e\n\u003c/ul\u003e\n\u003ch3 id=\"topic-3-omarchy-linux-distro-gains-traction-with-ai-agents-and-sleek-design\"\u003e\n  Topic 3: Omarchy Linux Distro Gains Traction with AI Agents and Sleek Design\n  \u003ca class=\"heading-link\" href=\"#topic-3-omarchy-linux-distro-gains-traction-with-ai-agents-and-sleek-design\"\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, 11,000 related posts\u003c/li\u003e\n\u003cli\u003eWhat it is: Omarchy Linux, a distribution aimed at developers, has gained attention on X for its support of AI agent workflows and its clean desktop design.\u003c/li\u003e\n\u003cli\u003eWhy it matters: It reflects a trend where AI agent applications are driving developers to refocus on controllable, lightweight, and easily automatable operating system environments, potentially influencing local AI toolchains and the developer experience.\u003c/li\u003e\n\u003cli\u003eDiscussion summary: Discussions mainly focus on Omarchy\u0026rsquo;s interface design, installation and configuration experience, the practical extent of its support for AI agents, hardware compatibility, system stability, and its advantages compared to existing Linux distributions. Some users believe its popularity stems more from its design aesthetic than from technical breakthroughs.\u003c/li\u003e\n\u003c/ul\u003e\n\u003ch3 id=\"topic-4-meme-juxtaposes-ai-doomsday-warnings-with-apples-camera-upgrades\"\u003e\n  Topic 4: Meme Juxtaposes AI Doomsday Warnings with Apple\u0026rsquo;s Camera Upgrades\n  \u003ca class=\"heading-link\" href=\"#topic-4-meme-juxtaposes-ai-doomsday-warnings-with-apples-camera-upgrades\"\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 hours ago, 1,400 related posts\u003c/li\u003e\n\u003cli\u003eWhat it is: A meme juxtaposes AI doomsday warnings with Apple\u0026rsquo;s continuous camera upgrades, creating a stark contrast.\u003c/li\u003e\n\u003cli\u003eWhy it matters: This reflects the gap between the grand public discourse on AI safety risks and the incremental innovation in consumer electronics. It helps in observing how the AI boom affects tech industry topics and public attention.\u003c/li\u003e\n\u003cli\u003eDiscussion summary: The focus of discussion on X is mainly on whether AI doomsday theories are overhyped, whether Apple\u0026rsquo;s camera upgrades lack innovation, and whether there is an imbalance in tech companies\u0026rsquo; focus on high-risk AI issues versus actual product experience.\u003c/li\u003e\n\u003c/ul\u003e\n\u003ch3 id=\"topic-5-meme-reimagines-job-interview-as-robot-apocalypse-stand-off\"\u003e\n  Topic 5: Meme Reimagines Job Interview as Robot Apocalypse Stand-Off\n  \u003ca class=\"heading-link\" href=\"#topic-5-meme-reimagines-job-interview-as-robot-apocalypse-stand-off\"\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 · Entertainment\u003c/li\u003e\n\u003cli\u003eOverview: Trending Time: , Related Posts: 2700\u003c/li\u003e\n\u003cli\u003eWhat it is: An internet meme dramatically reframes a job interview as a standoff between humans and robots in an apocalyptic scenario.\u003c/li\u003e\n\u003cli\u003eWhy it matters: It presents the impact of AI entering the workplace on employment, job competition, and human-machine relationships in an entertaining way, reflecting the social imagination sparked by AI\u0026rsquo;s evolution from a conversational tool to an intelligent agent and enterprise workflow component.\u003c/li\u003e\n\u003cli\u003eDiscussion Summary: Discussions on X primarily revolve around whether this setup is a commentary on AI taking jobs, interview anxiety, and human fears of automation. It also extends to the anthropomorphic expression of robots and whether the entertaining narrative exaggerates AI\u0026rsquo;s impact on the job market.\u003c/li\u003e\n\u003c/ul\u003e\n\u003ch3 id=\"topic-6-teaser-trailer-drops-for-guadagninos-sam-altman-drama-artificial\"\u003e\n  Topic 6: Teaser Trailer Drops for Guadagnino’s Sam Altman Drama Artificial\n  \u003ca class=\"heading-link\" href=\"#topic-6-teaser-trailer-drops-for-guadagninos-sam-altman-drama-artificial\"\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 · Entertainment\u003c/li\u003e\n\u003cli\u003eOverview: Trending Time: 1 day ago, Related Posts: 52000\u003c/li\u003e\n\u003cli\u003eWhat it is: The first teaser trailer for \u0026ldquo;Artificial,\u0026rdquo; a drama directed by Luca Guadagnino focusing on the story of Sam Altman, has been released, drawing attention on the X platform.\u003c/li\u003e\n\u003cli\u003eWhy it matters: The film brings the OpenAI power dynamics, AI commercialization, and the image of tech leaders into mainstream cinematic narrative, showing that generative AI has expanded from a technical issue to a topic of popular culture and public governance.\u003c/li\u003e\n\u003cli\u003eDiscussion Summary: Discussions on X are focused on whether the film will accurately portray the internal crisis at OpenAI, whether Sam Altman\u0026rsquo;s public image has been overly dramatized, and whether Hollywood is prematurely adapting events from the still-developing AI industry into entertainment.\u003c/li\u003e\n\u003c/ul\u003e\n\u003ch4 id=\"ai-public-opinion-summary-on-x-today\"\u003e\n  AI Public Opinion Summary on X Today\n  \u003ca class=\"heading-link\" href=\"#ai-public-opinion-summary-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 revolves around the tension between the expansion of AI capabilities and society\u0026rsquo;s capacity to adapt. On one hand, topics like DeepSeek V4.1 Flash and Omarchy Linux reinforce the consensus that low-cost, high-performance models, localized toolchains, and AI agent workflows are rapidly becoming widespread. On the other hand, the departure of a researcher from Anthropic, the job interview vs. robot meme, and related film projects have amplified concerns about superintelligence getting out of control, job displacement, corporate governance, and the power of tech. The main point of contention is whether the industry should continue to push the capabilities race while relying on safety research to mitigate risks, or if it should immediately slow down R\u0026amp;D and strengthen regulation. Meanwhile, the public is also debating whether the \u0026ldquo;AI apocalypse\u0026rdquo; narrative is a necessary wake-up call or excessive hype. Potential risks include the lack of reproducible verification for model performance and cost claims, companies downplaying safety responsibilities amidst commercial competition, and the tendency for entertainment and memes to amplify anxiety, blurring the line between real risks and technological hype.\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\u003cblockquote\u003e\n\u003cp\u003eInfluencer insights are unavailable today. We recommend reading the in-depth content from the Watch List.\u003c/p\u003e\n\u003c/blockquote\u003e\n\u003ch2 id=\"-appendix-todays-watch-list-update-source-list\"\u003e\n  📚 Appendix: Today\u0026rsquo;s Watch List Update Source List\n  \u003ca class=\"heading-link\" href=\"#-appendix-todays-watch-list-update-source-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\u003cblockquote\u003e\n\u003cp\u003eTimeframe: Last 3 days; 22 sources covered; 33 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/openai-does-math-reward-hacking-meta-launches-personal-agent/\"  class=\"external-link\" target=\"_blank\" rel=\"noopener\"\u003eOpenAI Does Math, Reward-Hacking, Meta Launches Personal Agent\u003c/a\u003e\u003c/strong\u003e\n\u003cul\u003e\n\u003cli\u003ePublication Time: 2026-09-09 18:00 Beijing Time\u003c/li\u003e\n\u003cli\u003eSummary: [TO BE TRANSLATED] - OpenAI solving one of the most famous math problems is extremely impressive, and of little impact to most people’s lives; Meta’s Muse agent launch has the potential to be the exact opposite.\n\u003cul\u003e\n\u003cli\u003e\u003cstrong\u003e$15\u003c/strong\u003e / month \u003cem\u003eor\u003c/em\u003e \u003cstrong\u003e$150\u003c/strong\u003e / year.\u003c/li\u003e\n\u003cli\u003eSubstantial analysis of the news of the day delivered via three weekly emails or podcasts.\u003c/li\u003e\n\u003cli\u003e\u003cstrong\u003eStratechery Interviews\u003c/strong\u003e.\u003c/li\u003e\n\u003cli\u003eInterviews with leading public CEOs, private company founders, and discussions with fellow analysts.\u003c/li\u003e\n\u003c/ul\u003e\n\u003c/li\u003e\n\u003cli\u003eEN Key Points:\n\u003cul\u003e\n\u003cli\u003eOpenAI solving one of the most famous math problems is extremely impressive, and of little impact to most people\u0026rsquo;s lives; Meta\u0026rsquo;s Muse agent launch has the poten…\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/paul-christiano-joins-openai-foundation-board\"  class=\"external-link\" target=\"_blank\" rel=\"noopener\"\u003ePaul Christiano joins OpenAI Foundation Board\u003c/a\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cul\u003e\n\u003cli\u003ePublish Time: 2026-09-10 01:00 Beijing Time\u003c/li\u003e\n\u003cli\u003eSummary: We’re announcing the appointment of Paul Christiano to the OpenAI Foundation Board.\n\u003cul\u003e\n\u003cli\u003eHe will be a non-voting observer on the OpenAI Group PBC Board.\u003c/li\u003e\n\u003cli\u003ePaul will also join the Safety and Security Committee (SSC) of the Foundation Board, working alongside its chair, Zico Kolter.\u003c/li\u003e\n\u003cli\u003eThe SSC provides governance over safety and security practices across all of OpenAI, including OpenAI Group PBC.\u003c/li\u003e\n\u003cli\u003eDepartment of Commerce, where he is a Senior Tech Advisor\u003ca href=\"#citation-bottom-1\" \u003e1\u003c/a\u003e.\u003c/li\u003e\n\u003c/ul\u003e\n\u003c/li\u003e\n\u003cli\u003eEN Key Points:\n\u003cul\u003e\n\u003cli\u003ePaul Christiano joins the OpenAI Foundation Board and its Safety and Security Committee, bringing experience in AI alignment, safety, and standards.\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/gpt-6-astra-next-generation-work\"  class=\"external-link\" target=\"_blank\" rel=\"noopener\"\u003eGPT-6 Astra: The next generation in intelligence for work\u003c/a\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cul\u003e\n\u003cli\u003ePublish Time: 2026-09-09 19:00 Beijing Time\u003c/li\u003e\n\u003cli\u003eSummary: Last week we introduced GPT‑6 Astra, the world’s most intelligent and aligned model, now available in ChatGPT Work, Codex, and the API.\n\u003cul\u003e\n\u003cli\u003eAstra is state-of-the-art on computer use, browsing, professional work, software engineering, cybersecurity, and science, so teams can take on the most demanding professional work with unmatched speed, accuracy, and judgment.\u003c/li\u003e\n\u003cli\u003e\n\u003ch2 id=\"the-worlds-best-model-for-complex-work\"\u003e\n  The world’s best model for complex work.\n  \u003ca class=\"heading-link\" href=\"#the-worlds-best-model-for-complex-work\"\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\u003c/li\u003e\n\u003cli\u003eMost AI systems require businesses to prepare their data, redesign workflows, and build custom integrations before they can deliver value.\u003c/li\u003e\n\u003cli\u003eIn ChatGPT Work and Codex, it can write code and work through the same applications people use every day—even when those applications don’t have an API.\u003c/li\u003e\n\u003c/ul\u003e\n\u003c/li\u003e\n\u003cli\u003eEN Key Points:\n\u003cul\u003e\n\u003cli\u003eMeet GPT-6 Astra, OpenAI’s most capable model for business, with advanced reasoning, computer use, and stronger writing and design judgment.\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-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/2609.05437\"  class=\"external-link\" target=\"_blank\" rel=\"noopener\"\u003eBeyond Right and Wrong: Evaluating Second-order Social Reasoning in Large Language Models\u003c/a\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cul\u003e\n\u003cli\u003ePublish Time: 2026-09-09 12:00 Beijing Time\u003c/li\u003e\n\u003c/ul\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003eSummary: [To be translated] - arXiv:2609.05437v1 Announce Type: new.\u003c/p\u003e\n\u003cul\u003e\n\u003cli\u003eAbstract: Previous AI alignment efforts have focused primarily on first-order social norms \u0026ndash; teaching models what is socially acceptable or unacceptable (e.g., `do not steal\u0026rsquo;).\u003c/li\u003e\n\u003cli\u003eHowever, social intelligence depends not only on norm recognition, but also on anticipating who will enforce it and how (e.g., public shame or even imprisonment).\u003c/li\u003e\n\u003cli\u003eThese second-order expectations, known as metanorms, govern how people respond when social rules are broken.\u003c/li\u003e\n\u003cli\u003eEN Key Points:\n\u003cul\u003e\n\u003cli\u003earXiv:2609.05437v1 Announce Type: new\u003c/li\u003e\n\u003cli\u003eAbstract: Previous AI alignment efforts have focused primarily on first-order social norms \u0026ndash; teaching models what is socially acceptable or unacceptable (e.g.,…\u003c/li\u003e\n\u003cli\u003eHowever, social intelligence depends not only on norm recognition, but also on anticipating who will enforce it and how (e.g., public shame or even imprisonment…\u003c/li\u003e\n\u003cli\u003eThese second-order expectations, known as metanorms, govern how people respond when social rules are broken\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/2609.05439\"  class=\"external-link\" target=\"_blank\" rel=\"noopener\"\u003eCriticGen: Generation-Aware Evaluation as Actionable Feedback\u003c/a\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cul\u003e\n\u003cli\u003ePublication Time: 2026-09-09 12:00 Beijing Time\u003c/li\u003e\n\u003cli\u003eSummary: [To be translated] - arXiv:2609.05439v1 Announce Type: new.\n\u003cul\u003e\n\u003cli\u003eAbstract: Current evaluation methods for large language models are coarse-grained and decoupled from generation, producing generic explanations that fail to provide actionable feedback for model improvement.\u003c/li\u003e\n\u003cli\u003eWe propose CriticGen, a fine-grained, generation-aware evaluation framework that turns evaluation into actionable control for answer improvement.\u003c/li\u003e\n\u003cli\u003eCriticGen first generates sample-specific evaluation dimensions and scoring criteria under high-level categories such as subjective, objective, and self-derived constraints.\u003c/li\u003e\n\u003c/ul\u003e\n\u003c/li\u003e\n\u003cli\u003eEN Key Points:\n\u003cul\u003e\n\u003cli\u003earXiv:2609.05439v1 Announce Type: new\u003c/li\u003e\n\u003cli\u003eAbstract: Current evaluation methods for large language models are coarse-grained and decoupled from generation, producing generic explanations that fail to pro…\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 CriticGen, a fine-grained, generation-aware evaluation framework that turns evaluation into actionable control for answer improvement\u003c/p\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003eCriticGen first generates sample-specific evaluation dimensions and scoring criteria under high-level categories such as subjective, objective, and self-derived…\u003c/p\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003e\u003cstrong\u003e\u003ca href=\"https://arxiv.org/abs/2609.05441\"  class=\"external-link\" target=\"_blank\" rel=\"noopener\"\u003eWhen Does Memory Help? A Cost-Aware Evaluation of Long-Term Memory in Tool-Using LLM Agents\u003c/a\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cul\u003e\n\u003cli\u003ePublication Time: 2026-09-09 12:00 Beijing Time\u003c/li\u003e\n\u003cli\u003eSummary: [Translation Pending] - arXiv:2609.05441v1 Announce Type: new.\n\u003cul\u003e\n\u003cli\u003eAbstract: Long-term memory for LLM agents is evaluated today by conversational recall benchmarks (LoCoMo, LongMemEval), which measure question answering over dialogue history, not whether remembered facts change what a tool-using agent does.\u003c/li\u003e\n\u003cli\u003eWe present MERIT (Memory Evaluation for Realistic Instrumented Tasks), a benchmark and harness that measures the marginal utility of memory for task-executing agents under explicit cost accounting.\u003c/li\u003e\n\u003cli\u003eMERIT provides episodic tool-use tasks in three domains whose dependence on earlier-episode facts is verified by an automated leak check; a difficulty ladder ending in updated-fact recall; controlled memory corruption; and full token and dollar metering of every memory operation.\u003c/li\u003e\n\u003c/ul\u003e\n\u003c/li\u003e\n\u003cli\u003eEN Key Points:\n\u003cul\u003e\n\u003cli\u003earXiv:2609.05441v1 Announce Type: new\u003c/li\u003e\n\u003cli\u003eAbstract: Long-term memory for LLM agents is evaluated today by conversational recall benchmarks (LoCoMo, LongMemEval), which measure question answering over di…\u003c/li\u003e\n\u003cli\u003eWe present MERIT (Memory Evaluation for Realistic Instrumented Tasks), a benchmark and harness that measures the marginal utility of memory for task-executing a…\u003c/li\u003e\n\u003cli\u003eMERIT provides episodic tool-use tasks in three domains whose dependence on earlier-episode facts is verified by an automated leak check; a difficulty ladder en…\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/2609.05446\"  class=\"external-link\" target=\"_blank\" rel=\"noopener\"\u003eAutoFyn Technical Report: Non-Parametric Expert Iteration for Long-Horizon Agents\u003c/a\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cul\u003e\n\u003cli\u003ePublication Time: 2026-09-09 12:00 Beijing Time\u003c/li\u003e\n\u003c/ul\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003eAbstract: [Translation Pending] - arXiv:2609.05446v1 Announce Type: new.\u003c/p\u003e\n\u003cul\u003e\n\u003cli\u003eAbstract: We introduce AutoFyn, an agent harness inspired by the Expert Iteration algorithm, adapting a frozen model across many rounds by updating persistent state from verified reward signals rather than model weights.\u003c/li\u003e\n\u003cli\u003eEach round begins from a fresh model session, and durable information is reintroduced only through explicit interfaces such as persistent memory files, reports, and repository state.\u003c/li\u003e\n\u003cli\u003eWithin a round, an orchestrator explores, plans and builds many alternative approaches with specialized agents, while a task-grounded verifier verifies the work and supplies an objective reward for measuring progress.\u003c/li\u003e\n\u003cli\u003eEN Highlights:\n\u003cul\u003e\n\u003cli\u003earXiv:2609.05446v1 Announce Type: new\u003c/li\u003e\n\u003cli\u003eAbstract: We introduce AutoFyn, an agent harness inspired by the Expert Iteration algorithm, adapting a frozen model across many rounds by updating persistent s…\u003c/li\u003e\n\u003cli\u003eEach round begins from a fresh model session, and durable information is reintroduced only through explicit interfaces such as persistent memory files, reports,…\u003c/li\u003e\n\u003cli\u003eWithin a round, an orchestrator explores, plans and builds many alternative approaches with specialized agents, while a task-grounded verifier verifies the work…\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/2609.05448\"  class=\"external-link\" target=\"_blank\" rel=\"noopener\"\u003eDamage-Aware Bandit Pruning for Vision and Language Transformers\u003c/a\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cul\u003e\n\u003cli\u003ePublication Time: 2026-09-09 12:00 Beijing Time\u003c/li\u003e\n\u003cli\u003eAbstract: [Translation Pending] - arXiv:2609.05448v1 Announce Type: new.\n\u003cul\u003e\n\u003cli\u003eAbstract: Structured post-training pruning of transformers requires selecting complete functional units whose suppression causes limited degradation.\u003c/li\u003e\n\u003cli\u003eWe formulate structured-unit selection for language and vision transformers as a damage-aware multi-armed bandit problem under a fixed candidate-evaluation budget.\u003c/li\u003e\n\u003cli\u003eAttention heads and MLP channel groups are temporarily masked on calibration batches.\u003c/li\u003e\n\u003c/ul\u003e\n\u003c/li\u003e\n\u003cli\u003eEN Highlights:\n\u003cul\u003e\n\u003cli\u003earXiv:2609.05448v1 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: Structured post-training pruning of transformers requires selecting complete functional units whose suppression causes limited degradation\u003c/p\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003eWe formulate structured-unit selection for language and vision transformers as a damage-aware multi-armed bandit problem under a fixed candidate-evaluation budg…\u003c/p\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003eAttention heads and MLP channel groups are temporarily masked on calibration batches\u003c/p\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003e\u003cstrong\u003e\u003ca href=\"https://arxiv.org/abs/2609.05459\"  class=\"external-link\" target=\"_blank\" rel=\"noopener\"\u003eCompiling VGDL into Causal Models\u003c/a\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cul\u003e\n\u003cli\u003ePublished: 2026-09-09 12:00 Beijing Time\u003c/li\u003e\n\u003cli\u003eAbstract: [Translation Pending] - arXiv:2609.05459v1 Announce Type: new.\n\u003cul\u003e\n\u003cli\u003eAbstract: Reinforcement learning and large language models often struggle to accurately capture the causal mechanics of game environments.\u003c/li\u003e\n\u003cli\u003eStandard reinforcement learning agents tend to rely on spurious correlations, while large language models are prone to hallucinating game rules.\u003c/li\u003e\n\u003cli\u003eAlthough causal reinforcement learning improves interpretability, there is currently no formal methodology to map complex game mechanics directly into causal models.\u003c/li\u003e\n\u003c/ul\u003e\n\u003c/li\u003e\n\u003cli\u003eEN Key Points:\n\u003cul\u003e\n\u003cli\u003earXiv:2609.05459v1 Announce Type: new\u003c/li\u003e\n\u003cli\u003eAbstract: Reinforcement learning and large language models often struggle to accurately capture the causal mechanics of game environments\u003c/li\u003e\n\u003cli\u003eStandard reinforcement learning agents tend to rely on spurious correlations, while large language models are prone to hallucinating game rules\u003c/li\u003e\n\u003cli\u003eAlthough causal reinforcement learning improves interpretability, there is currently no formal methodology to map complex game mechanics directly into causal mo…\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/2609.05461\"  class=\"external-link\" target=\"_blank\" rel=\"noopener\"\u003eARC-Bench: Closed-Loop Replanning Masks Broken Action Ranking in Frozen JEPA World Models\u003c/a\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cul\u003e\n\u003cli\u003ePublished: 2026-09-09 12:00 Beijing Time\u003c/li\u003e\n\u003cli\u003eAbstract: [Translation Pending] - arXiv:2609.05461v1 Announce Type: new.\n\u003cul\u003e\n\u003cli\u003eAbstract: Reward-free latent world models plan by scoring candidate actions with distances in a frozen latent space: an action is preferred if its predicted future embedding lands closer to the goal embedding.\u003c/li\u003e\n\u003c/ul\u003e\n\u003c/li\u003e\n\u003c/ul\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003eThis silently assumes that latent closeness is action-rankable, i.e., that ordering candidates by latent distance agrees with ordering them by true cost.\u003c/p\u003e\n\u003cul\u003e\n\u003cli\u003eWe audit this assumption directly.\u003c/li\u003e\n\u003cli\u003eEN Key Points:\n\u003cul\u003e\n\u003cli\u003earXiv:2609.05461v1 Announce Type: new\u003c/li\u003e\n\u003cli\u003eAbstract: Reward-free latent world models plan by scoring candidate actions with distances in a frozen latent space: an action is preferred if its predicted fut…\u003c/li\u003e\n\u003cli\u003eThis silently assumes that latent closeness is action-rankable, i.e., that ordering candidates by latent distance agrees with ordering them by true cost\u003c/li\u003e\n\u003cli\u003eWe audit this assumption directly\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/2609.05481\"  class=\"external-link\" target=\"_blank\" rel=\"noopener\"\u003eRAPID: Reliability-Aware Pair Importance Distillation\u003c/a\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cul\u003e\n\u003cli\u003eRelease Time: 2026-09-09 12:00 Beijing Time\u003c/li\u003e\n\u003cli\u003eAbstract: [To be translated] - arXiv:2609.05481v1 Announce Type: new.\n\u003cul\u003e\n\u003cli\u003eAbstract: Inter example relational distillation transfers a teacher\u0026rsquo;s representation geometry by matching relations among examples within a mini batch.\u003c/li\u003e\n\u003cli\u003eComputing all pairs has quadratic complexity in the batch size, whereas uniform subsampling may use a limited relation budget inefficiently.\u003c/li\u003e\n\u003cli\u003eWe introduce Reliability Aware Pair Importance Distillation, or RAPID, which separates a reliability gated relational target from a full support adaptive pair proposal.\u003c/li\u003e\n\u003c/ul\u003e\n\u003c/li\u003e\n\u003cli\u003eEN Key Points:\n\u003cul\u003e\n\u003cli\u003earXiv:2609.05481v1 Announce Type: new\u003c/li\u003e\n\u003cli\u003eAbstract: Inter example relational distillation transfers a teacher\u0026rsquo;s representation geometry by matching relations among examples within a mini batch\u003c/li\u003e\n\u003cli\u003eComputing all pairs has quadratic complexity in the batch size, whereas uniform subsampling may use a limited relation budget inefficiently\u003c/li\u003e\n\u003cli\u003eWe introduce Reliability Aware Pair Importance Distillation, or RAPID, which separates a reliability gated relational target from a full support adaptive pair 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/2609.05488\"  class=\"external-link\" target=\"_blank\" rel=\"noopener\"\u003ePGP-Clinical-TimeKAN: Prior-Guided Joint Probabilistic Forecasting of Clinical Trajectories\u003c/a\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cul\u003e\n\u003cli\u003eRelease Time: 2026-09-09 12:00 Beijing Time\u003c/li\u003e\n\u003cli\u003eAbstract: [To be translated] - arXiv:2609.05488v1 Announce Type: new.\u003c/li\u003e\n\u003c/ul\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003eAbstract: Clinical deterioration unfolds through coupled, partially observed trajectories, not a single diagnostic label.\u003c/p\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003eWe introduce PGP-Clinical-TimeKAN, a trajectory-first framework for joint probabilistic forecasting of multivariate physiology.\u003c/p\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003eIt combines missingness-aware temporal encoders, a soft organ-system prior, patient-specific relations, nonlinear Kolmogorov-Arnold messages, and a low-rank multivariate Student-t head.\u003c/p\u003e\n\u003cul\u003e\n\u003cli\u003eEN Highlights:\n\u003cul\u003e\n\u003cli\u003earXiv:2609.05488v1 Announce Type: new\u003c/li\u003e\n\u003cli\u003eAbstract: Clinical deterioration unfolds through coupled, partially observed trajectories, not a single diagnostic label\u003c/li\u003e\n\u003cli\u003eWe introduce PGP-Clinical-TimeKAN, a trajectory-first framework for joint probabilistic forecasting of multivariate physiology\u003c/li\u003e\n\u003cli\u003eIt combines missingness-aware temporal encoders, a soft organ-system prior, patient-specific relations, nonlinear Kolmogorov-Arnold messages, and a low-rank mul…\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/2609.05505\"  class=\"external-link\" target=\"_blank\" rel=\"noopener\"\u003eSciLitBench: Benchmark and Design Principles for LLM-Powered Systematic Literature Reviews\u003c/a\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cul\u003e\n\u003cli\u003ePublished: 2026-09-09 12:00 Beijing Time\u003c/li\u003e\n\u003cli\u003eAbstract: [To be translated] - arXiv:2609.05505v1 Announce Type: new.\n\u003cul\u003e\n\u003cli\u003eAbstract: Systematic reviews require sustained human judgment across thousands of records, yet existing evaluations of large language models (LLMs) typically examine review stages in isolation.\u003c/li\u003e\n\u003cli\u003eWe introduce SciLitBench, a multi-stage benchmark spanning title and abstract screening, full-text screening, and schema-guided data extraction, with 42,981 retrieved records, 1,012 full texts, and annotations for 888 included papers.\u003c/li\u003e\n\u003cli\u003eAcross 22 open-weight LLMs from six model families, explicit inclusion and exclusion criteria improve title and abstract screening $F_2$ by 28.8%, while researcher-authored rationales improve full-text screening by 15%.\u003c/li\u003e\n\u003c/ul\u003e\n\u003c/li\u003e\n\u003cli\u003eEN Highlights:\n\u003cul\u003e\n\u003cli\u003earXiv:2609.05505v1 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: Systematic reviews require sustained human judgment across thousands of records, yet existing evaluations of large language models (LLMs) typically ex…\u003c/p\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003eWe introduce SciLitBench, a multi-stage benchmark spanning title and abstract screening, full-text screening, and schema-guided data extraction, with 42,981 ret…\u003c/p\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003eAcross 22 open-weight LLMs from six model families, explicit inclusion and exclusion criteria improve title and abstract screening $F_2$ by 28.8%, while resear…\u003c/p\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/2609.05531\"  class=\"external-link\" target=\"_blank\" rel=\"noopener\"\u003eWhen Agent Governance Helps\u003c/a\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cul\u003e\n\u003cli\u003ePublished: 2026-09-09 12:00 Beijing Time\u003c/li\u003e\n\u003cli\u003eAbstract: [To be translated] - arXiv:2609.05531v1 Announce Type: new.\n\u003cul\u003e\n\u003cli\u003eAbstract: No specification says how a governed autotelic AI agent organization, where agents pursue self-generated goals inside guardrails, should be designed and evaluated.\u003c/li\u003e\n\u003cli\u003eFirst, we synthesize the Governed Autotelic Multi-Agent Product Organization (GAMPO) framework from a document-based qualitative evidence synthesis of 321 sources, integrating agency, agile, platform, and governance theory into a runnable specification.\u003c/li\u003e\n\u003cli\u003eSecond, we probe a prompt-layer instantiation of GAMPO on CHI-Bench, a long-horizon healthcare benchmark, across open and frontier models.\u003c/li\u003e\n\u003c/ul\u003e\n\u003c/li\u003e\n\u003cli\u003eEN Key Points:\n\u003cul\u003e\n\u003cli\u003earXiv:2609.05531v1 Announce Type: new\u003c/li\u003e\n\u003cli\u003eAbstract: No specification says how a governed autotelic AI agent organization, where agents pursue self-generated goals inside guardrails, should be designed a…\u003c/li\u003e\n\u003cli\u003eWe answer in two parts\u003c/li\u003e\n\u003cli\u003eFirst, we synthesize the Governed Autotelic Multi-Agent Product Organization (GAMPO) framework from a document-based qualitative evidence synthesis of 321 sourc…\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/2609.05631\"  class=\"external-link\" target=\"_blank\" rel=\"noopener\"\u003eTamilEOT: A Dataset and Model for Semantic End-of-Turn Detection in Tamil Telephone Speech\u003c/a\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cul\u003e\n\u003cli\u003ePublished: 2026-09-09 12:00 Beijing Time\u003c/li\u003e\n\u003cli\u003eAbstract: [To be translated] - arXiv:2609.05631v1 Announce Type: new.\n\u003cul\u003e\n\u003cli\u003eAbstract: A voice agent has to decide, at every pause, whether the user has finished speaking.\u003c/li\u003e\n\u003c/ul\u003e\n\u003c/li\u003e\n\u003c/ul\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003eWithout a model of the language that decision falls back to a fixed silence timeout: set it short and the agent interrupts, set it long and every turn pays the full wait.\u003c/p\u003e\n\u003cul\u003e\n\u003cli\u003eOpen semantic end-of-turn detectors exist, but to our knowledge none covers a South Indian language.\u003c/li\u003e\n\u003cli\u003eEN Highlights:\n\u003cul\u003e\n\u003cli\u003earXiv:2609.05631v1 Announce Type: new\u003c/li\u003e\n\u003cli\u003eAbstract: A voice agent has to decide, at every pause, whether the user has finished speaking\u003c/li\u003e\n\u003cli\u003eWithout a model of the language that decision falls back to a fixed silence timeout: set it short and the agent interrupts, set it long and every turn pays the…\u003c/li\u003e\n\u003cli\u003eOpen semantic end-of-turn detectors exist, but to our knowledge none covers a South Indian language\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/2609.05637\"  class=\"external-link\" target=\"_blank\" rel=\"noopener\"\u003eBetter Together: Complementary Query Rewriting Under a Strong RAG Baseline\u003c/a\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cul\u003e\n\u003cli\u003ePublished: 2026-09-09 12:00 Beijing Time\u003c/li\u003e\n\u003cli\u003eAbstract: [To be translated] - arXiv:2609.05637v1 Announce Type: new.\n\u003cul\u003e\n\u003cli\u003eAbstract: A popular way to improve Retrieval-Augmented Generation (RAG) is to rewrite the user\u0026rsquo;s question into several variants and search with all of them.\u003c/li\u003e\n\u003cli\u003eWe test whether this actually helps once the underlying search is already strong.\u003c/li\u003e\n\u003cli\u003eUnder one fixed, competitive pipeline (BGE dense retrieval, cross-encoder reranking, and MMR diversification), we compare four query-rewriting strategies (S1-S4) against two strong LLM baselines (HyDE, Query2Doc) on three datasets (HotpotQA, AmbigNQ, and the 512K-document EnterpriseRAG-Bench) over three seeds with paired-bootstrap significance tests.\u003c/li\u003e\n\u003c/ul\u003e\n\u003c/li\u003e\n\u003cli\u003eEN Highlights:\n\u003cul\u003e\n\u003cli\u003earXiv:2609.05637v1 Announce Type: new\u003c/li\u003e\n\u003cli\u003eAbstract: A popular way to improve Retrieval-Augmented Generation (RAG) is to rewrite the user\u0026rsquo;s question into several variants and search with all of them\u003c/li\u003e\n\u003cli\u003eWe test whether this actually helps once the underlying search is already strong\u003c/li\u003e\n\u003cli\u003eUnder one fixed, competitive pipeline (BGE dense retrieval, cross-encoder reranking, and MMR diversification), we compare four query-rewriting strategies (S1-S4…\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/2609.05677\"  class=\"external-link\" target=\"_blank\" rel=\"noopener\"\u003eWho Maintains Agent Skills? A Longitudinal Study of Human-Governed, AI-Assisted Skill Maintenance\u003c/a\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cul\u003e\n\u003cli\u003ePublished: 2026-09-09 12:00 Beijing Time\u003c/li\u003e\n\u003cli\u003eSummary: [Translation Pending] - arXiv:2609.05677v1 Announce Type: new.\n\u003cul\u003e\n\u003cli\u003eAbstract: Lifelong LLM agents increasingly rely on external skill artifacts as one element for preserving and reusing capabilities over time.\u003c/li\u003e\n\u003cli\u003eThese skills (usually portable Markdown files such as SKILL.md) describe when and how to apply a capability and must be corrected, expanded, and consolidated as tools and usage patterns shift over deployment.\u003c/li\u003e\n\u003cli\u003eRecent work seeks to automate skill curation, but it largely evaluates against automated baselines and treats human maintenance as an unmeasured bottleneck.\u003c/li\u003e\n\u003c/ul\u003e\n\u003c/li\u003e\n\u003cli\u003eEN Highlights:\n\u003cul\u003e\n\u003cli\u003earXiv:2609.05677v1 Announce Type: new\u003c/li\u003e\n\u003cli\u003eAbstract: Lifelong LLM agents increasingly rely on external skill artifacts as one element for preserving and reusing capabilities over time\u003c/li\u003e\n\u003cli\u003eThese skills (usually portable Markdown files such as SKILL.md) describe when and how to apply a capability and must be corrected, expanded, and consolidated as…\u003c/li\u003e\n\u003cli\u003eRecent work seeks to automate skill curation, but it largely evaluates against automated baselines and treats human maintenance as an unmeasured bottleneck\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/2609.05682\"  class=\"external-link\" target=\"_blank\" rel=\"noopener\"\u003eA Rubric-Guided Large Language Model Solution for Opioid Use Disorder Computable Phenotyping\u003c/a\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cul\u003e\n\u003cli\u003ePublished: 2026-09-09 12:00 Beijing Time\u003c/li\u003e\n\u003cli\u003eSummary: [Translation Pending] - arXiv:2609.05682v1 Announce Type: new.\n\u003cul\u003e\n\u003cli\u003eAbstract: Opioid use disorder (OUD) remains a public health crisis in the United States, yet it is difficult to identify from electronic health records (EHRs) because missing diagnosis codes and supporting evidence are buried in clinical narratives.\u003c/li\u003e\n\u003cli\u003eAccurate OUD identification is critical to support interventions and improve health outcomes.\u003c/li\u003e\n\u003cli\u003eThis study developed a rubric-guided large language model (LLM) that incorporated Optimization by PROmpting (OPRO) for OUD computable phenotyping (CP).\u003c/li\u003e\n\u003c/ul\u003e\n\u003c/li\u003e\n\u003cli\u003eEN Highlights:\u003c/li\u003e\n\u003c/ul\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003earXiv:2609.05682v1 Announce Type: new\u003c/p\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003eAbstract: Opioid use disorder (OUD) remains a public health crisis in the United States, yet it is difficult to identify from electronic health records (EHRs) b…\u003c/p\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003eAccurate OUD identification is critical to support interventions and improve health outcomes\u003c/p\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003eThis study developed a rubric-guided large language model (LLM) that incorporated Optimization by PROmpting (OPRO) for OUD computable phenotyping (CP)\u003c/p\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003e\u003cstrong\u003e\u003ca href=\"https://arxiv.org/abs/2609.05707\"  class=\"external-link\" target=\"_blank\" rel=\"noopener\"\u003eIntra-Prompt Parallel Decoding for Common-Context Question Answering\u003c/a\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cul\u003e\n\u003cli\u003ePublished: 2026-09-09 12:00 Beijing Time\u003c/li\u003e\n\u003cli\u003eAbstract: - arXiv:2609.05707v1 Announce Type: new.\n\u003cul\u003e\n\u003cli\u003eAbstract: In common-context question answering (CCQA) tasks, multiple input questions share a common context to base their answers from.\u003c/li\u003e\n\u003cli\u003eHowever, Large Language Models typically generate each answer using an independent prompt.\u003c/li\u003e\n\u003cli\u003eWhile existing batching and caching techniques help improve parallelism and reduce repeated computations, the separation of questions across prompts limits the achievable speedup, as modern GPUs are underutilized due to a memory bottleneck during attention.\u003c/li\u003e\n\u003c/ul\u003e\n\u003c/li\u003e\n\u003cli\u003eKey Points:\n\u003cul\u003e\n\u003cli\u003earXiv:2609.05707v1 Announce Type: new\u003c/li\u003e\n\u003cli\u003eAbstract: In common-context question answering (CCQA) tasks, multiple input questions share a common context to base their answers from\u003c/li\u003e\n\u003cli\u003eHowever, Large Language Models typically generate each answer using an independent prompt\u003c/li\u003e\n\u003cli\u003eWhile existing batching and caching techniques help improve parallelism and reduce repeated computations, the separation of questions across prompts limits the…\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/2609.05728\"  class=\"external-link\" target=\"_blank\" rel=\"noopener\"\u003eMedWER: A Reproducible, Model-Free Evaluation Protocol for Medical Speech Recognition\u003c/a\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cul\u003e\n\u003cli\u003ePublished: 2026-09-09 12:00 Beijing Time\u003c/li\u003e\n\u003cli\u003eAbstract: - arXiv:2609.05728v1 Announce Type: new.\n\u003cul\u003e\n\u003cli\u003eAbstract: Overall word error rate hides clinically critical errors: a transcript can be 95% correct and still swap one drug for another.\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 usual fix weights errors on medical entities, and almost always depends on an evaluation-time named-entity recognition (NER) model or cloud API, which makes the metric\u0026rsquo;s denominator a versioned black box.\u003c/p\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003eWe present MedWER, an evaluation protocol and open-source tool for medical ASR whose denominator is a fixed, license-clean term list: 19,373 drug, diagnosis, symptom, and injury-mechanism entries projected from public sources.\u003c/p\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003eEN Highlights:\u003c/p\u003e\n\u003cul\u003e\n\u003cli\u003earXiv:2609.05728v1 Announce Type: new\u003c/li\u003e\n\u003cli\u003eAbstract: Overall word error rate hides clinically critical errors: a transcript can be 95% correct and still swap one drug for another\u003c/li\u003e\n\u003cli\u003eThe usual fix weights errors on medical entities, and almost always depends on an evaluation-time named-entity recognition (NER) model or cloud API, which makes…\u003c/li\u003e\n\u003cli\u003eWe present MedWER, an evaluation protocol and open-source tool for medical ASR whose denominator is a fixed, license-clean term list: 19,373 drug, diagnosis, sy…\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/2609.05743\"  class=\"external-link\" target=\"_blank\" rel=\"noopener\"\u003eSome Tokens Behave like Magnets: Revealing Linguistic Organization in the Layers of Language Models\u003c/a\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cul\u003e\n\u003cli\u003ePublished: 2026-09-09 12:00 Beijing Time\u003c/li\u003e\n\u003cli\u003eAbstract: [Translation Pending] - arXiv:2609.05743v1 Announce Type: new.\n\u003cul\u003e\n\u003cli\u003eAbstract: We identify a special group of token vectors inside large language models (LLMs), which we term magnetic vectors, that organize the surrounding tokens by either attracting or repelling them.\u003c/li\u003e\n\u003cli\u003eParticularly, tokens pointing the same way as an attracting magnet are elongated; tokens pointing the same way as a repelling magnet are compressed.\u003c/li\u003e\n\u003cli\u003eJust as physical magnets pull or push away the iron filings around them, these vectors organize their surroundings through two opposing polarities.\u003c/li\u003e\n\u003c/ul\u003e\n\u003c/li\u003e\n\u003cli\u003eEN Highlights:\n\u003cul\u003e\n\u003cli\u003earXiv:2609.05743v1 Announce Type: new\u003c/li\u003e\n\u003cli\u003eAbstract: We identify a special group of token vectors inside large language models (LLMs), which we term magnetic vectors, that organize the surrounding tokens…\u003c/li\u003e\n\u003c/ul\u003e\n\u003c/li\u003e\n\u003c/ul\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003eParticularly, tokens pointing the same way as an attracting magnet are elongated; tokens pointing the same way as a repelling magnet are compressed\u003c/p\u003e\n\u003cul\u003e\n\u003cli\u003eJust as physical magnets pull or push away the iron filings around them, these vectors organize their surroundings through two opposing polarities\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/2609.05757\"  class=\"external-link\" target=\"_blank\" rel=\"noopener\"\u003eCrisisKD: Five-Stage Knowledge Distillation for Aspect-Level Sentiment and Emotion Analysis in Crisis Discourse\u003c/a\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cul\u003e\n\u003cli\u003eRelease Time: 2026-09-09 12:00 Beijing Time\u003c/li\u003e\n\u003cli\u003eAbstract: [To be translated] - arXiv:2609.05757v1 Announce Type: new.\n\u003cul\u003e\n\u003cli\u003eAbstract: Identifying the target of emotional words or phrases in crisis situations, especially health-related ones, is important for understanding public concerns across cultural and linguistic contexts.\u003c/li\u003e\n\u003cli\u003eWe propose CrisisKD, a five-stage teacher\u0026ndash;student knowledge distillation framework for aspect-level sentiment and emotion analysis on unannotated social media data.\u003c/li\u003e\n\u003cli\u003eA teacher LLM generates aspect-level labels and reasoning traces that supervise a smaller student model across aspect extraction, syntactic parsing, opinion extraction, sentiment classification, and emotion classification.\u003c/li\u003e\n\u003c/ul\u003e\n\u003c/li\u003e\n\u003cli\u003eEN Key Points:\n\u003cul\u003e\n\u003cli\u003earXiv:2609.05757v1 Announce Type: new\u003c/li\u003e\n\u003cli\u003eAbstract: Identifying the target of emotional words or phrases in crisis situations, especially health-related ones, is important for understanding public conce…\u003c/li\u003e\n\u003cli\u003eWe propose CrisisKD, a five-stage teacher\u0026ndash;student knowledge distillation framework for aspect-level sentiment and emotion analysis on unannotated social media…\u003c/li\u003e\n\u003cli\u003eA teacher LLM generates aspect-level labels and reasoning traces that supervise a smaller student model across aspect extraction, syntactic parsing, opinion ext…\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/2609.05797\"  class=\"external-link\" target=\"_blank\" rel=\"noopener\"\u003eRecall Is Not Protection: Evaluating Safety Monitors Against Model Compliance\u003c/a\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cul\u003e\n\u003cli\u003eRelease Time: 2026-09-09 12:00 Beijing Time\u003c/li\u003e\n\u003cli\u003eAbstract: [To be translated] - arXiv:2609.05797v1 Announce Type: new.\u003c/li\u003e\n\u003c/ul\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003eAbstract: Safety monitors screen prompts sent to deployed language models, flagging harmful requests so they are never answered.\u003c/p\u003e\n\u003cul\u003e\n\u003cli\u003eThey are evaluated by recall against harmfulness labels, but a catch only prevents harm if the model would otherwise have complied.\u003c/li\u003e\n\u003cli\u003eWe measure the difference directly: we sample repeated responses from the target model, call a harmful prompt \\emph{elicitable} if the model complies at least once, and report monitor recall separately on elicitable and non-elicitable prompts.\u003c/li\u003e\n\u003cli\u003eEN Highlights:\n\u003cul\u003e\n\u003cli\u003earXiv:2609.05797v1 Announce Type: new\u003c/li\u003e\n\u003cli\u003eAbstract: Safety monitors screen prompts sent to deployed language models, flagging harmful requests so they are never answered\u003c/li\u003e\n\u003cli\u003eThey are evaluated by recall against harmfulness labels, but a catch only prevents harm if the model would otherwise have complied\u003c/li\u003e\n\u003cli\u003eWe measure the difference directly: we sample repeated responses from the target model, call a harmful prompt \\emph{elicitable} if the model complies at least o…\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-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/2609.05435\"  class=\"external-link\" target=\"_blank\" rel=\"noopener\"\u003eAhaBench: Do Agents Learn from Prior Experience? A Benchmark for Long-Horizon Continual Learning\u003c/a\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cul\u003e\n\u003cli\u003eRelease Time: 2026-09-09 12:00 Beijing Time\u003c/li\u003e\n\u003cli\u003eAbstract: [Translation Pending] - arXiv:2609.05435v1 Announce Type: new.\n\u003cul\u003e\n\u003cli\u003eAbstract: Modern language agents are expected to operate over long horizons: they ask follow-up questions, reuse worked examples, handle tool feedback, and adapt to delayed consequences.\u003c/li\u003e\n\u003cli\u003eMost evaluations still reset the agent after a prompt or score only the final state of one trajectory.\u003c/li\u003e\n\u003cli\u003eAhaBench asks a more operational question: when a fixed model receives useful experience, does its later behavior improve under a related evaluation condition where the obvious support has been removed, changed, or delayed?\u003c/li\u003e\n\u003c/ul\u003e\n\u003c/li\u003e\n\u003cli\u003eEN Highlights:\n\u003cul\u003e\n\u003cli\u003earXiv:2609.05435v1 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: Modern language agents are expected to operate over long horizons: they ask follow-up questions, reuse worked examples, handle tool feedback, and adap…\u003c/p\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003eMost evaluations still reset the agent after a prompt or score only the final state of one trajectory\u003c/p\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003eAhaBench asks a more operational question: when a fixed model receives useful experience, does its later behavior improve under a related evaluation condition w…\u003c/p\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003e\u003cstrong\u003e\u003ca href=\"https://arxiv.org/abs/2609.05508\"  class=\"external-link\" target=\"_blank\" rel=\"noopener\"\u003eWhen Do Options Help? Policy Necrosis and Redundant Coverage in Option-Critic\u003c/a\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cul\u003e\n\u003cli\u003ePublished: 2026-09-09 12:00 Beijing Time\u003c/li\u003e\n\u003cli\u003eAbstract: [Pending Translation] - arXiv:2609.05508v1 Announce Type: new.\n\u003cul\u003e\n\u003cli\u003eAbstract: Option-critic learns options: sub-policies together with a learned rule for when each one hands control back.\u003c/li\u003e\n\u003cli\u003eIts headline result is that performance improves as options are added.\u003c/li\u003e\n\u003cli\u003eWe explain that result, with theory and experiment.\u003c/li\u003e\n\u003c/ul\u003e\n\u003c/li\u003e\n\u003cli\u003eEN Highlights:\n\u003cul\u003e\n\u003cli\u003earXiv:2609.05508v1 Announce Type: new\u003c/li\u003e\n\u003cli\u003eAbstract: Option-critic learns options: sub-policies together with a learned rule for when each one hands control back\u003c/li\u003e\n\u003cli\u003eIts headline result is that performance improves as options are added\u003c/li\u003e\n\u003cli\u003eWe explain that result, with theory and experiment\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/2609.05574\"  class=\"external-link\" target=\"_blank\" rel=\"noopener\"\u003eMulti-granularity Adaptive Hypergraph Representation Learning via Granular-ball\u003c/a\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cul\u003e\n\u003cli\u003ePublished: 2026-09-09 12:00 Beijing Time\u003c/li\u003e\n\u003cli\u003eAbstract: [Pending Translation] - arXiv:2609.05574v1 Announce Type: new.\n\u003cul\u003e\n\u003cli\u003eAbstract: Hypergraph representation learning aims to capture high-order information in graphs by constructing hyperedges that simultaneously connect multiple nodes.\u003c/li\u003e\n\u003cli\u003eThese hyperedges adapt to the graph\u0026rsquo;s topological features, facilitating the extraction of high-order relationships at multiple granularities.\u003c/li\u003e\n\u003cli\u003eMost prior work relies on predefined definitions to generate hyperedges, overlooking the diversity in graph topological structures and the multi-granularity characteristics of hyperedges.\u003c/li\u003e\n\u003c/ul\u003e\n\u003c/li\u003e\n\u003cli\u003eEN Highlights:\n\u003cul\u003e\n\u003cli\u003earXiv:2609.05574v1 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: Hypergraph representation learning aims to capture high-order information in graphs by constructing hyperedges that simultaneously connect multiple no…\u003c/p\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003eThese hyperedges adapt to the graph\u0026rsquo;s topological features, facilitating the extraction of high-order relationships at multiple granularities\u003c/p\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003eMost prior work relies on predefined definitions to generate hyperedges, overlooking the diversity in graph topological structures and the multi-granularity cha…\u003c/p\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003e\u003cstrong\u003e\u003ca href=\"https://arxiv.org/abs/2609.05575\"  class=\"external-link\" target=\"_blank\" rel=\"noopener\"\u003eCapsule Lens: Locating and Tracking Concept Geometry in Model Representations\u003c/a\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cul\u003e\n\u003cli\u003ePublication Time: 2026-09-09 12:00 Beijing Time\u003c/li\u003e\n\u003cli\u003eAbstract: [Pending Translation] - arXiv:2609.05575v1 Announce Type: new.\n\u003cul\u003e\n\u003cli\u003eAbstract: Understanding how concepts are encoded in the internal representations of machine learning models is a central problem in mechanistic interpretability, essential both for the science of deep learning and for the trustworthy deployment of increasingly capable models.\u003c/li\u003e\n\u003cli\u003eExisting approaches to interpret model representations mainly map representations onto more interpretable spaces and do not directly characterize how concepts occupy representation space; various hypotheses have been proposed, but often lack of rigorous validation and largely focus on static representations.\u003c/li\u003e\n\u003cli\u003eIn this work, we introduce Capsule Lens, a framework that matches the region a concept occupies with a simple, trackable geometric form, a capsule, defined by several interpretable parameters, fitted in closed form to each concept\u0026rsquo;s geometry and validated on held-out samples.\u003c/li\u003e\n\u003c/ul\u003e\n\u003c/li\u003e\n\u003cli\u003eEN Key Points:\n\u003cul\u003e\n\u003cli\u003earXiv:2609.05575v1 Announce Type: new\u003c/li\u003e\n\u003cli\u003eAbstract: Understanding how concepts are encoded in the internal representations of machine learning models is a central problem in mechanistic interpretability…\u003c/li\u003e\n\u003cli\u003eExisting approaches to interpret model representations mainly map representations onto more interpretable spaces and do not directly characterize how concepts o…\u003c/li\u003e\n\u003c/ul\u003e\n\u003c/li\u003e\n\u003c/ul\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003eIn this work, we introduce Capsule Lens, a framework that matches the region a concept occupies with a simple, trackable geometric form, a capsule, defined by s…\u003c/p\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003e\u003cstrong\u003e\u003ca href=\"https://arxiv.org/abs/2609.05582\"  class=\"external-link\" target=\"_blank\" rel=\"noopener\"\u003eHB-PVI: A Hierarchical Bayesian Personalization and Value-of-Information Framework for Complex Activity Recognition\u003c/a\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cul\u003e\n\u003cli\u003ePublication Time: 2026-09-09 12:00 Beijing Time\u003c/li\u003e\n\u003cli\u003eAbstract: [To be translated] - arXiv:2609.05582v1 Announce Type: new.\n\u003cul\u003e\n\u003cli\u003eAbstract: Personalization can improve activity-recognition performance, but participant-specific gains are heterogeneous, and every additional calibration label has an acquisition cost.\u003c/li\u003e\n\u003cli\u003eThis study presents HB-PVI, a hierarchical Bayesian personalization and value-of-information framework jointly modeling participant heterogeneity, the benefit and harm of four personalization mechanisms, and the economic value of an additional label, for the 47-participant MUSIC-CAR complex-activity cohort.\u003c/li\u003e\n\u003cli\u003eA leakage-safe, leave-one-participant-out evaluation combines a sequential-Monte-Carlo participant-effect updater with a Student-$t$ hierarchical gain model and a one-step expected-value-of-sample-information (EVSI) stopping rule.\u003c/li\u003e\n\u003c/ul\u003e\n\u003c/li\u003e\n\u003cli\u003eEN Key Points:\n\u003cul\u003e\n\u003cli\u003earXiv:2609.05582v1 Announce Type: new\u003c/li\u003e\n\u003cli\u003eAbstract: Personalization can improve activity-recognition performance, but participant-specific gains are heterogeneous, and every additional calibration label…\u003c/li\u003e\n\u003cli\u003eThis study presents HB-PVI, a hierarchical Bayesian personalization and value-of-information framework jointly modeling participant heterogeneity, the benefit a…\u003c/li\u003e\n\u003cli\u003eA leakage-safe, leave-one-participant-out evaluation combines a sequential-Monte-Carlo participant-effect updater with a Student-$t$ hierarchical gain model and…\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/2609.05650\"  class=\"external-link\" target=\"_blank\" rel=\"noopener\"\u003eEndogenous Exploration in Reinforcement Learning with Intrinsic Curiosity\u003c/a\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cul\u003e\n\u003cli\u003ePublication Time: 2026-09-09 12:00 Beijing Time\u003c/li\u003e\n\u003cli\u003eAbstract: [To be translated] - arXiv:2609.05650v1 Announce Type: new.\u003c/li\u003e\n\u003c/ul\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003eAbstract: We propose a reinforcement learning framework in which exploration is driven by intrinsic curiosity, designed for scenarios where environments are non-stationary and rewards are sparse, delayed, uninformative, or absent.\u003c/p\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003eIn our model, action selection is guided by a combination of external rewards and an epistemic motivation mechanism that biases the agent toward structured exploratory directions.\u003c/p\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003eThe central hypothesis is that effective exploration emerges at intermediate levels of incoherence, while performance degrades under both overly rigid and overly disordered dynamics.\u003c/p\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003eEN Key Points:\u003c/p\u003e\n\u003cul\u003e\n\u003cli\u003earXiv:2609.05650v1 Announce Type: new\u003c/li\u003e\n\u003cli\u003eAbstract: We propose a reinforcement learning framework in which exploration is driven by intrinsic curiosity, designed for scenarios where environments are non…\u003c/li\u003e\n\u003cli\u003eIn our model, action selection is guided by a combination of external rewards and an epistemic motivation mechanism that biases the agent toward structured expl…\u003c/li\u003e\n\u003cli\u003eThe central hypothesis is that effective exploration emerges at intermediate levels of incoherence, while performance degrades under both overly rigid and overl…\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/2609.05658\"  class=\"external-link\" target=\"_blank\" rel=\"noopener\"\u003eRobustness of LLM-Generated SystemVerilog Assertions to Semantics-Preserving RTL Transformations\u003c/a\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cul\u003e\n\u003cli\u003ePublication Time: 2026-09-09 12:00 Beijing Time\u003c/li\u003e\n\u003cli\u003eAbstract: [Translation Pending] - arXiv:2609.05658v1 Announce Type: new.\n\u003cul\u003e\n\u003cli\u003eAbstract: Large language models (LLMs) are increasingly being explored for automating SystemVerilog Assertion (SVA) generation, yet most evaluations report correctness on a single syntactic representation of an input.\u003c/li\u003e\n\u003cli\u003eSuch point accuracy does not reveal whether a model\u0026rsquo;s correct output is stable when the same RTL behavior is written differently.\u003c/li\u003e\n\u003cli\u003eThis paper presents a controlled metamorphic evaluation of LLM-based SVA generation under semantics-preserving RTL transformations.\u003c/li\u003e\n\u003c/ul\u003e\n\u003c/li\u003e\n\u003cli\u003eEN Key Points:\n\u003cul\u003e\n\u003cli\u003earXiv:2609.05658v1 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 language models (LLMs) are increasingly being explored for automating SystemVerilog Assertion (SVA) generation, yet most evaluations report corr…\u003c/p\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003eSuch point accuracy does not reveal whether a model\u0026rsquo;s correct output is stable when the same RTL behavior is written differently\u003c/p\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003eThis paper presents a controlled metamorphic evaluation of LLM-based SVA generation under semantics-preserving RTL transformations\u003c/p\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003e\u003cstrong\u003e\u003ca href=\"https://arxiv.org/abs/2609.05676\"  class=\"external-link\" target=\"_blank\" rel=\"noopener\"\u003ePAC-Private Autoregressive Generation: Calibrating Noise to Ensemble Disagreement\u003c/a\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cul\u003e\n\u003cli\u003ePublication Time: 2026-09-09 12:00 Beijing Time\u003c/li\u003e\n\u003cli\u003eAbstract: [Translation pending] - arXiv:2609.05676v1 Announce Type: new.\n\u003cul\u003e\n\u003cli\u003eAbstract: Language models adapted on private text are often served through APIs, so privacy leakage occurs through generated outputs rather than exposed weights.\u003c/li\u003e\n\u003cli\u003ePrivate prediction protects these releases.\u003c/li\u003e\n\u003cli\u003eMethods such as PMixED incur privacy cost at each release and increasingly rely on the public model over long horizons.\u003c/li\u003e\n\u003c/ul\u003e\n\u003c/li\u003e\n\u003cli\u003eEN Key Points:\n\u003cul\u003e\n\u003cli\u003earXiv:2609.05676v1 Announce Type: new\u003c/li\u003e\n\u003cli\u003eAbstract: Language models adapted on private text are often served through APIs, so privacy leakage occurs through generated outputs rather than exposed weights\u003c/li\u003e\n\u003cli\u003ePrivate prediction protects these releases\u003c/li\u003e\n\u003cli\u003eMethods such as PMixED incur privacy cost at each release and increasingly rely on the public model over long horizons\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/2609.05688\"  class=\"external-link\" target=\"_blank\" rel=\"noopener\"\u003eConnecting Score Matching, Maximum Likelihood, and Expectation-Maximization in Mixed Linear Regression\u003c/a\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cul\u003e\n\u003cli\u003ePublication Time: 2026-09-09 12:00 Beijing Time\u003c/li\u003e\n\u003cli\u003eAbstract: [Translation pending] - arXiv:2609.05688v1 Announce Type: new.\n\u003cul\u003e\n\u003cli\u003eAbstract: We study variance-preserving diffusion of the response in mixed linear regression (MLR) with unknown mixing weights.\u003c/li\u003e\n\u003cli\u003eOur analysis separates the statistical guarantees of score matching from the loss geometry and optimization signal at a fixed diffusion noise level.\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 KL divergence links the denoising score matching objective integrated over the diffusion path with the likelihood and a terminal discrepancy.\u003c/p\u003e\n\u003cul\u003e\n\u003cli\u003eEN Key Points:\n\u003cul\u003e\n\u003cli\u003earXiv:2609.05688v1 Announce Type: new\u003c/li\u003e\n\u003cli\u003eAbstract: We study variance-preserving diffusion of the response in mixed linear regression (MLR) with unknown mixing weights\u003c/li\u003e\n\u003cli\u003eOur analysis separates the statistical guarantees of score matching from the loss geometry and optimization signal at a fixed diffusion noise level\u003c/li\u003e\n\u003cli\u003eThe KL divergence links the denoising score matching objective integrated over the diffusion path with the likelihood and a terminal discrepancy\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/2609.05694\"  class=\"external-link\" target=\"_blank\" rel=\"noopener\"\u003eGraphNOSE: A Graph Transformer in Olfaction\u003c/a\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cul\u003e\n\u003cli\u003ePublication Time: 2026-09-09 12:00 Beijing Time\u003c/li\u003e\n\u003cli\u003eSummary: [Translation Pending] - arXiv:2609.05694v1 Announce Type: new.\n\u003cul\u003e\n\u003cli\u003eAbstract: Predicting olfactory qualities from molecular structure is an open problem in chemoinformatics.\u003c/li\u003e\n\u003cli\u003eAlthough linear models can link molecular features to odor descriptors, they often fail when extrapolating to novel chemical scaffolds, extreme molecular weights, or complex odor mixtures.\u003c/li\u003e\n\u003cli\u003eTo address this, we introduce GraphNOSE, an open-source graph transformer framework that predicts multi-label odor descriptors from simplified molecular-input line-entry system (SMILES) strings for single molecules and binary mixtures.\u003c/li\u003e\n\u003c/ul\u003e\n\u003c/li\u003e\n\u003cli\u003eEN Key Points:\n\u003cul\u003e\n\u003cli\u003earXiv:2609.05694v1 Announce Type: new\u003c/li\u003e\n\u003cli\u003eAbstract: Predicting olfactory qualities from molecular structure is an open problem in chemoinformatics\u003c/li\u003e\n\u003cli\u003eAlthough linear models can link molecular features to odor descriptors, they often fail when extrapolating to novel chemical scaffolds, extreme molecular weight…\u003c/li\u003e\n\u003cli\u003eTo address this, we introduce GraphNOSE, an open-source graph transformer framework that predicts multi-label odor descriptors from simplified molecular-input l…\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": 6456,
  "readingTime": 31,
  "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=\"#-ai-hotspots-on-x\"\u003e🌐 AI Hotspots on X\u003c/a\u003e\n      \u003cul\u003e\n        \u003cli\u003e\u003ca href=\"#topic-1-deepseek-v41-flash-delivers-top-ai-performance-at-low-cost\"\u003eTopic 1: DeepSeek V4.1 Flash Delivers Top AI Performance at Low Cost\u003c/a\u003e\u003c/li\u003e\n        \u003cli\u003e\u003ca href=\"#topic-2-ai-researcher-quits-anthropic-over-superintelligence-risks\"\u003eTopic 2: AI Researcher Quits Anthropic Over Superintelligence Risks\u003c/a\u003e\u003c/li\u003e\n        \u003cli\u003e\u003ca href=\"#topic-3-omarchy-linux-distro-gains-traction-with-ai-agents-and-sleek-design\"\u003eTopic 3: Omarchy Linux Distro Gains Traction with AI Agents and Sleek Design\u003c/a\u003e\u003c/li\u003e\n        \u003cli\u003e\u003ca href=\"#topic-4-meme-juxtaposes-ai-doomsday-warnings-with-apples-camera-upgrades\"\u003eTopic 4: Meme Juxtaposes AI Doomsday Warnings with Apple\u0026rsquo;s Camera Upgrades\u003c/a\u003e\u003c/li\u003e\n        \u003cli\u003e\u003ca href=\"#topic-5-meme-reimagines-job-interview-as-robot-apocalypse-stand-off\"\u003eTopic 5: Meme Reimagines Job Interview as Robot Apocalypse Stand-Off\u003c/a\u003e\u003c/li\u003e\n        \u003cli\u003e\u003ca href=\"#topic-6-teaser-trailer-drops-for-guadagninos-sam-altman-drama-artificial\"\u003eTopic 6: Teaser Trailer Drops for Guadagnino’s Sam Altman Drama Artificial\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    \u003cli\u003e\u003ca href=\"#-appendix-todays-watch-list-update-source-list\"\u003e📚 Appendix: Today\u0026rsquo;s Watch List Update Source List\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      \u003c/ul\u003e\n    \u003c/li\u003e\n    \u003cli\u003e\u003ca href=\"#the-worlds-best-model-for-complex-work\"\u003eThe world’s best model for complex work.\u003c/a\u003e\n      \u003cul\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
}