AX·NEWS
Sign in

Artificial Intelligence

Artificial Intelligence news today

80 stories from 9 sources in the last 48 hours, every one in English. Each links to its publisher.

Open in the live feedRSS feedOn a TV

  1. France "Supermarket Fridge Hack Excavation Iran"...EU Extra Edition Promotion
    KRAI Times (Korea) Translated from Korean

    The French government has overseeed the Iran Revolution Squadron as the back of the cyber attacks that have been fighting the refrigeration and refrigeration of hundreds of supermarkets last month. In France, the food waste and fire…

  2. Vidu Reveals ‘Animal Ivove’ to automate AI research... Improvement of 6x Speed
    KRAI Times (Korea) Translated from Korean

    China Baidu has published an automatic research system that can utilize AI development by establishing a research direction and repeating experiments. We have shown the technical progress towards recurring self improvement (RSI) in a…

  3. [AI Model Story] 6 types from virtual world real-time expedition to 22MB image analysis AI
    KRAI Times (Korea) Translated from Korean

    ■ WorldLabs, an AI-produced virtual world real-time-exported "Frameboy" released the virtual world created by the world model "Atlas" and the "Frameboy" that can be explored in real time like the game. When the user moves the camera, AI is…

  4. The cost of "Hiku" performance is 3 minutes...Meta, an agent "Muse Ember" released
    KRAI Times (Korea) Translated from Korean

    Meta revealed a lightweight AI model ‘Muse Ember’ designed to handle repeated actions of AI agents fast and affordable. From large-scale documentation analysis to code modification, tool calling, and work automation that connects multiple…

  5. Cultivation of Cultivation of Cultivation of Cultivation of Cultivation of Cultivation of Cultivation
    KRAI Times (Korea) Translated from Korean

    AX business with crowdworks and in-hands. Crowdworks (CEO Kim Gwang-il) announced that it will participate as an official partner in the ‘Manufacturing Ammuji-based AI model development business’ in Humans and Industry. The industrial…

  6. Google, AI Agent·Security Specialization TPU 2 types public...Environment
    KRAI Times (Korea) Translated from Korean

    Google has expanded its own Tensourcing Device (TPU) family to meet ENVIDIA in the AI chip market. We add two new products to our corporate security system that handles AI agents and sensitive data, and strengthen the AI infrastructure…

  7. OpenFormating, AI-based SQL automatic tuning solution ‘Quartik 2.0’ released
    KRAI Times (Korea) Translated from Korean

    Open Maid Consulting (CEO Choi Young-cheol) announced that it will be launched on the 16th day of the query Medic 2.0, which will automatically analyze the SQL of the database (DB) and improve the performance. QueryMedicine is a solution…

  8. Antropic, 93 trillion IPO Plan Passed Directly...Excludes 35 trillion investments
    KRAI Times (Korea) Translated from Korean

    Antropicy has delivered an IPO plan to an investment bank that raises up to $693 billion (approximately 93 trillion won), and it has been known that the public price information is incorrect and withdrawn. The data includes seven major…

  9. The U.S. Congress, the AI Data Center Construction Promotion Act...
    KRAI Times (Korea) Translated from Korean

    The U.S. Congress mitted the construction permit of the AI data center at the heart of the second bill to strengthen the cost of the operator's power grid. However, in the middle of the democrats, the democrats enforced the data center…

  10. Trump "FBI, Jenson Huang Case Study"...Is there an AI event
    KRAI Times (Korea) Translated from Korean

    President Donald Trump said that the President of the U.S. President of the United States was investigated by the Federal Bureau of FBI. In the event that the CEO of Huang has been in connection with the Chinese Information Agency, the…

  11. Nvidia·Jenson Huang, a group of shareholders faced...Spy suspicious 'Herit Explanation' Contest
    KRAI Times (Korea) Translated from Korean

    Nvidia and Jenson Huang CEO faced the shareholders’ group societies for the reason that the information related to Chinese spy fraud was not disclosed properly to investors. It is a key issue that has provided information to investors, who…

  12. DeepX, Ultra Low Power AI Semiconductor ‘DX-M2’ design completed... Mass production of the second half of the year
    KRAI Times (Korea) Translated from Korean

    The AI Semiconductor Professional DeepX (CEO Kim Cheonwon) was designed for the next generation generation of AI semiconductor ‘DX-M2’ based on Samsung Electronics’ 2Nano process, and it was in charge for mass production in the second half…

  13. Dental Clinic, Dental Clinic, Dental Clinic
    KRAI Times (Korea) Translated from Korean

    Dental AI specialized Dencom (CEO, Im-Jeon) has established an AI-based medical record system in the Western and the hospital of Seoul. It is a key to automatically record the conversation of the patient’s ancestor and the patient’s…

  14. Employment increase in South Korean occupation 88.5% in the Tajiki area...
    KRAI Times (Korea) Translated from Korean

    This year's insurance for the job seeker has increased by 78 people than last year, but 88.5% of the increase appeared in the number of people working on different attempts. In the same period, the employer in the same period has increased…

  15. Governments, 25,000 tons of barley special purchases...In order to be delivered to the next year
    KRAI Times (Korea) Translated from Korean

    In order to respond to the rapid barley production of the government this year, the next year's assets will be adjusted to the production center of the contract. The domestic barley market, which has undergone a lack of supply in…

  16. Tokyo High-rise Building, Susuk production and storage of solar...
    KRAI Times (Korea) Translated from Korean

    The system that converts hydrogen into the rooftop solar into the large office building in Tokyo, Japan will be introduced in November 2027. The solar power generation time and the power usage time of the building is a business to solve…

  17. TURKEYY, 2040 Solar & Wind Power 175GW Target
    KRAI Times (Korea) Translated from Korean

    TURKEYY has presented the goal of expanding solar and wind power plants by 2040 to 175 GW. As of September this year, the installation capacity is about 43.8GW, and it plans to invest about 1000 billion dollars (approximately 135 trillion…

  18. North America Solar PPA Present Price 4.9% Lift...America Dec.
    KRAI Times (Korea) Translated from Korean

    North America Solar Power Purchasing Contract (PPA) announced price has risen 4.9% earlier than Q3 of this year. It is an analysis that the expected cost of solar materials and parts import regulations to be implemented in December of the…

  19. The first year of insurance for the fishing season...the government, 160 hires are checked
    KRAI Times (Korea) Translated from Korean

    The government has spent about 160 countries in the country where foreign fishing seasons are hired. This year, the insurance affiliate obligations conducted in February confirm that they are kept on the site. The Ministry of Justice and…

  20. Aristudio Receives CES 2027 Innovation Award...Recognized AI Film Technology
    KRAI Times (Korea) Translated from Korean

    AI content specialist Aria studio (CEO) won the ‘CES 2027 Innovation Award’ with the technology that the audience can interact with the characters in the film and participate in the entire story. This is an interactive film technology that…

  21. [AIReader's Book] Claw Code Perfect Guide for Team
    KRAI Times (Korea) Translated from Korean

    Clad code for the team Perfect Guide Yun Dae Hee | WikiBooks2026 October 29 recommendedThis book covers the working process of converting Clad code into a standardized infrastructure in a personal tool. From pilot design to cost…

  22. What is AI? "Temporary error"
    KRAI Times (Korea) Translated from Korean

    ■ Industrial specialized AI agent, manufacturing AI’s competitive standards are changed, and the competition criteria of the company AI are moving to industry-specific work performance in the performance of the universal model. This is the…

  23. [Nov. 9] Chinese AI, American model, but...There is an open AI, and tropicen unopened model
    KRAI Times (Korea) Translated from Korean

    When Chinese AI companies are launching a new product, it will increase the emphasis of following the cutting-edge models of OpenAI and Andropic, or in some performances. Andt Group announced the last five days of the Multimodal AI model…

  24. Meta’s Conductor deal puts enterprise AI coding workflows on deadline
    GBAI Business

    The social media giant’s acquisition leaves developers facing a cloud shutdown and highlights that enterprises need to protect AI coding workflows and plan vendor exits.

  25. ANTIQUE 'FIVE 5.2', AI performance evaluation No. 1 deversion...API price also 20%
    KRAI Times (Korea) Translated from Korean

    Andropic has released a new top-level model ‘Claude Fable 5.2’ and ‘Claude Mythos 5.2’. Payable 5.2 has been deploying Google in the performance evaluation (AAII) of the AttiFicial Analyst, and the API price is also 20%, and the…

  26. Mi Parliament, Jenson Sulfide Press...China AI Chip Deal·Government Responsibility
    KRAI Times (Korea) Translated from Korean

    Mr. Jenson Huang, China’s spying suspicious around Nvidia CEO, leads to the supervisory issues of China AI semiconductor export policies and Trump administration of the U.S. Congress. Since the public auditor demands a non-public briefing…

  27. Jensen Huang’s future raises questions for AI infrastructure industry
    GBAI Business

    Possible foreign intelligence targeting at Nvidia raises concerns about CEO succession and chipmaker’s influence over the AI market.

  28. "There is no zenson hugyeom"...Announcement to the influence of NVIDIA children
    KRAI Times (Korea) Translated from Korean

    As Jenson Huang Nvidia CEO is involved in the Chinese spying suspicion, the Board of Directors will be revised and the influence of the company of the latter and the family. As a member of the Board of Directors, the Board of Directors of…

  29. 8.8% of the week of Nvidia...the shareholders support Jenson Huang
    KRAI Times (Korea) Translated from Korean

    Jenson Huang Nvidia CEO’s cheating recognition and the Chinese spy hug, Nvidia’s share was up to 8.8% faster. However, major investors are open to publicly support management sustainability while disappointing personal stolen. In the…

  30. Jailbreak week: how SharkNinja is hacking its own business to unlock AI
    GBAI Business

    Some organizations are bold, while others are more measured in their approach.

  31. The backlash about data centers might really just be about electricity
    USMIT Technology Review Biotech

    This story originally appeared in The Algorithm, our weekly newsletter on AI. To get stories like this in your inbox first, sign up here . As voters went to the polls across the US last week, I was in Ashburn, Virginia, part of what’s…

  32. ChatGPT Enterprise now available in AWS Marketplace
    USOpenAI Blog

    ChatGPT Enterprise is now available through AWS Marketplace, helping organizations simplify procurement through existing AWS purchasing and billing relationships.

  33. Nancy Pelosi’s replacement is ready to regulate AI
    USMIT Technology Review Biotech

    As a California state senator, Scott Wiener authored legislation aimed at curbing the risks of advanced AI. Now, after winning a tight congressional race for former Speaker of the House Nancy Pelosi’s open San Francisco seat, he could play…

  34. Continuous Batching Beats Static and Dynamic by Up to 2.91× in vLLM
    USTowards Data Science

    For anyone running an inference server: a zero-variance control, five traffic traces, and an engine-vs-scheduler split show where the gains come from. The post Continuous Batching Beats Static and Dynamic by Up to 2.91× in vLLM appeared…

  35. Nemotron-Personas-Finance-Japan: A Synthetic Dataset Capturing Japan's Demographics and Financial Context
    USHugging Face Blog
  36. Ask AI to Refine Your Ideas, Not Invent Them
    USTowards Data Science

    Studies of short stories, metaphors, and 486 participants point one way, an 800-person experiment points the other, and the author weighs both. The post Ask AI to Refine Your Ideas, Not Invent Them appeared first on Towards Data Science .

  37. Import AI 480: Self-sustaining space economies; AI for science progress, and Meta builds perfect hands for a superintelligence
    USImport AI Newsletter

    Welcome to Import AI, a newsletter about AI research. Import AI runs on arXiv, cappuccinos, and feedback from readers. If you’d like to support this, please subscribe. Subscribe now AI Science startup converts a quantum conjecture into a…

  38. How to Estimate the Recall of AI Agents with Capture-Recapture
    USTowards Data Science

    A fisheries estimator from the 1890s, tested on 294 CUAD contracts, landed within 2 points of true recall for most agents with a 10% hand audit. The post How to Estimate the Recall of AI Agents with Capture-Recapture appeared first on…

  39. The Download: empty AI climate claims, and an Nvidia investigation
    USMIT Technology Review Biotech

    This is today’s edition of The Download , our weekday newsletter that provides a daily dose of what’s going on in the world of technology. Opinion: The climate case for AI has not been made —By Mary Nichols, distinguished counsel at the…

  40. Innovating with AI in the special chemicals industry
    USMIT Technology Review Biotech

    As companies look to accelerate innovation with artificial intelligence, cloud computing, and advanced modeling, the primary challenge is connecting new technologies to reliable data, scientific expertise, and the practical realities of…

  41. Introducing Control Room: Human control over AI workloads
    USOpenAI Blog

    Learn how OpenAI’s Control Room tracks model workloads across our research infrastructure and can rapidly stop them when safety or security concerns arise.

  42. Control Room: A central system to track and stop model workloads across our research infrastructure
    USOpenAI Blog

    Learn how OpenAI’s Control Room tracks model workloads across our research infrastructure and can rapidly stop them when safety or security concerns arise.

  43. Your Experts Have Partners. There's HOPE for Keeping Them Together
    USHugging Face Blog
  44. Quantile Regression Caught the House Price OLS Missed by $55K
    USTowards Data Science

    On 88 house sales, OLS predicted $164K for a $219K home, while a 90% interval from 5th and 95th percentile models spanned $136K to $227K. The post Quantile Regression Caught the House Price OLS Missed by $55K appeared first on Towards Data…

  45. Jenson Huangnon on Nvidia Silence...GPU supply without discrimination"
    KRAI Times (Korea) Translated from Korean

    In the first open main official event, Mr. Jenson Huang Nvidia CEO, Nvidia’s affiliates asked questions. The head of the head of the head of the head of the head of the head of the head of the head of the head of the head of the head of…

  46. Kakao Mobility-Environmentia, Level 4 Self-Governing Cooperation...Kia PBV Preferred
    KRAI Times (Korea) Translated from Korean

    Kakao Mobility and Nvidia are integrated into the vehicle (PBV) for the purpose of driving the technology and computing·sensing·safety technology. Kakao Mobility (CEO R&D) announced that it concludes a work cooperative agreement (MOU) for…

  47. NVIDIA, three domestic AI startups.. . Adress: No. 5 Yard, 7th Guanggu Street, Badaling Economic Development Zone, Yanqing District, Beijing
    KRAI Times (Korea) Translated from Korean

    Nvidia is investing in three domestic AI startups including real world, business status, and motif technology. Nvidia has released the same investment plan at the ‘AI Startup Partnership Forum’ held at the ‘AI Startup Partnership Forum’…

  48. The climate case for AI has not been made
    USMIT Technology Review Biotech

    This week, as countries meet in Antalya, Turkey for the United Nations climate conference, one topic likely to dominate the conversation is artificial intelligence. The host has already called on countries to commit to ensuring that AI is…

  49. [Bulletin] Upstay, Series C 2nd investment, Nvidia joins
    KRAI Times (Korea) Translated from Korean

    ■ Upstay paper (CEO Kim Sung-hun) announced that it will receive an investment from Nvidia at the round of the second investment of the series C. The amount of investment is private. Once the investment is completed, it will be secured as…

  50. PSI “Proof of 18 Years of Quantum Flame Proof” with Non-Public and Trophy AI
    KRAI Times (Korea) Translated from Korean

    Research results have been demonstrated by the ovulation "Aaronson-Ambainis speculation" in Yangjae Computing field that has not been pulled for 18 years. AI creates a mathematical proof and passes through to validation through computers…

  51. AI learning data is only 0.1% manipulated so you can deepen the remote code backdoor to the agent
    KRAI Times (Korea) Translated from Korean

    A study found that only 0.1% of AI pre-learning data can be contaminated by the AI agent developed on this basis. It is not possible to remove the backdoors from the normal data to fine-tuning and reinforcing learning, and it has been…

  52. Google opens the ‘elephant-boldfish’ plug-in with coded pre-design error
    KRAI Times (Korea) Translated from Korean

    Google has released the ‘elephant-goldfish model (EGM)’ plugin to improve the design and review process of AI coding agent as an open source. Before creating a code, create a detailed design specification, and it is a key that a separate…

  53. Nvidia reveals 5 times faster ‘VFlash’ for Robot AI
    KRAI Times (Korea) Translated from Korean

    Nvidia revealed the technology that reduces route prediction errors while increasing the AI harbor speed to plan the movement of the robot. Instead of creating a coordinate of a moving path, it is a description that can be used to improve…

  54. Nus Research, Political Bias Answer of Chinese AI Model 22%→0.9%
    KRAI Times (Korea) Translated from Korean

    Nassau Resach published ‘Arete-27B(Arete-27B)’, a politic response in the open-way model of Alibaba. It is characterized by almost maintaining the coding and agent performance of the existing model, while the lead microsecond is reduced…

  55. Motif, NVidia Investment, develops ‘Motif 4’
    KRAI Times (Korea) Translated from Korean

    Motif Technologies (CEO Im Jeong-hwan) announced that the next-generation foundation has completed the investment attracting investment for the development of large-language model (LLM) ‘Motif 4’ and was completed in full development. This…

  56. BAIR at CoRL 2026
    USBAIR Blog

    The Conference on Robot Learning (CoRL) takes place November 9–12, 2026, in Austin, Texas. BAIR faculty, postdocs and students will be presenting research at various sessions, including an Early Career Keynote from Professor Negar Mehr…

  57. If you are staying in Denmark and the photovoltaic power transmission limits, you will be exempt from the burden of your connection
    KRAI Times (Korea) Translated from Korean

    We are committed to protecting the unauthorized collection of e-mail addresses posted on this website. If you do not agree to the use of e-mail address, please contact us directly. In order to increase the first priority in the power…

  58. Amazon, 'Astro' fails and re-interpretation of household robots...We look forward to the development of dermal AI
    KRAI Times (Korea) Translated from Korean

    Amazon recalled the next generation robot development even in the absence of Astro. Unlike the past, AI is a determining the real environment and the practicality of household robots is also higher, while the ‘Physical AI’ technology is…

  59. NVIDIA·SK Telecom·Leebelion, Iggijong AI Computing Match 3
    KRAI Times (Korea) Translated from Korean

    NVIDIA·SKTelecom·Rebelion collects hands to implement fusion AI computing system that connects NPU and NVIDIA GPU. These 3 companies concluded ‘NV Link Fusion’ cooperation agreement (MOU) at the Nvidia ‘AI Day’ event held at COEX in Seoul…

  60. LG, ‘K-Exam 2.5’ API Free open front... Lower barriers to enter the developer
    KRAI Times (Korea) Translated from Korean

    LG AI researchers will open the API of ‘K-exam 2.5’, which is the reader AI foundation model (e.g. poison), and will open it for free. LG’s decision is that the model performance has been developed for a long time, reflecting the…

  61. India, Agricultural Irrigation Pump Solar Switching Speed...Opens 116GW market with local duty
    KRAI Times (Korea) Translated from Korean

    In India, if you switch to the solar power for agricultural irrigation pumps, solar modules have been able to have an additional demand for 116 GW. This is an analysis of the transition of irrigation facilities while the Government of…

  62. Google AI discovers two spherical RSA password weaknesses...Insufficient cost 30x lowering algorithm development
    KRAI Times (Korea) Translated from Korean

    Google has developed a new algorithm that reduces the cost of acquiring multiple password keys at one time by using AI agent. AI collaborates with human experts to demonstrate complex password analysis and mathematical fixation. Google…

  63. LGU+, Seoul, KAIST·Development of voice AI that understands the next generation and ‘Momto·Macro
    KRAI Times (Korea) Translated from Korean

    LG UPlus will develop next-generation voice AI specialized in domestic universities and Korean. LGU+ announced that it will develop ‘K-V2V’ technology-based ‘K-Voice Engine’, which will give speech and voice to the University of Seoul…

  64. US solar industry, after mid-line policy support promotion...Restore of three-components
    KRAI Times (Korea) Translated from Korean

    In the middle of the U.S., the Democratic Republic of the U.S. demands the development of solar and energy storage devices (ESS) and supports the U.S. solar manufacturing. However, the restoration of clean energy deficiency is also…

  65. Kuken Industries searches 70 years of drawings in minutes
    USOpenAI Blog

    Kuken Industries uses the OpenAI API to search 70 years of cooling tower drawings. Engineers can find the drawing for a repair in 4.5 minutes, down from nearly an hour.

  66. CodeRabbit scales agentic software development and ships 20x more
    USOpenAI Blog

    CodeRabbit uses OpenAI models to govern and trust AI-generated code and Codex to ship 20 times more software with the same engineering team.

  67. OpenAI, Life Science AI Benchmark Public...GPT-6 Astra Performance 23%P Upgrade
    KRAI Times (Korea) Translated from Korean

    OpenAI revealed a new benchmark that measures AI agent’s data analysis and error detection ability in real life science labs. “GPT-6 Astra”, the latest model, has recorded a high score of 22.9% than previous models, and this study is also…

  68. Nvidia "No need for a long context on the LLM"...Continue reading
    KRAI Times (Korea) Translated from Korean

    Nvidia Research Institute announced that it is not necessary to have a long context window for large language model (LLM). If you want to compress the conversation history or read only the part you need in a large document, you can…

  69. The As-If Machine: Personalized AI counterfactuals make distant policy risks feel close
    USarXiv cs.AI

    arXiv:2610.33301v2 Announce Type: replace-cross Abstract: Life experiences can shape policy attitudes, but we only live one life. Citizens may therefore fail to support policies addressing risks they have never personally experienced…

  70. From Trajectories to Prefixes: Reusing Teacher Trajectories via Replayed Prefixes and Online Continuation
    USarXiv cs.AI

    arXiv:2607.19395v2 Announce Type: replace-cross Abstract: Small language models are attractive backbones for interactive agents, but direct distillation from strong teacher trajectories often turns rich multi-turn behavior into one-shot…

  71. Broken Gates: Re-evaluating Web Bot Defenses in the Age of LLM Agents
    USarXiv cs.AI

    arXiv:2607.18659v2 Announce Type: replace-cross Abstract: The rapid emergence of web automation across commercial crawlers, solver services, and LLM-based browser agents has created new challenges for defenses that must distinguish…

  72. Curved Knowledge in Flat Models: Building Expert AI at Scale
    USarXiv cs.AI

    arXiv:2607.17513v3 Announce Type: replace-cross Abstract: Intelligence that must act before outside correction arrives must carry the structure of its expert knowledge, and hyperbolic geometry represents hierarchy with a fidelity no…

  73. Beyond Bayesian Nash: Learning Minimax-Regret Equilibria for Adversarial Team Games under Asymmetric Information
    USarXiv cs.AI

    arXiv:2607.09993v2 Announce Type: replace-cross Abstract: Adversarial team games (ATGs) with asymmetric information, such as adversarial path-finding, goal search, and reachability games on graphs, require strategies that are robust to…

  74. Decomposing Yes-No Bias in LLM Moral Judgment: Stance, Position, Word, and Verdict
    USarXiv cs.AI

    arXiv:2607.05552v2 Announce Type: replace-cross Abstract: Large language models (LLMs) increasingly issue binary verdicts - safety gates, judge decisions, value probes - and the answer is read as the model's judgment. A settled judgment…

  75. Steering Training Trajectories in Diffusion Representation Learning
    USarXiv cs.AI

    arXiv:2607.05319v2 Announce Type: replace-cross Abstract: Across training runs of a diffusion autoencoder, we observe similar image reconstruction quality, but substantially different levels of disentanglement in the learned…

  76. MCP Server Architecture Patterns for LLM-Integrated Applications
    USarXiv cs.AI

    arXiv:2606.30317v2 Announce Type: replace-cross Abstract: The Model Context Protocol (MCP), introduced by Anthropic in November 2024, defines a standardized interface for connecting large language models (LLMs) to external tools, data…

  77. Exposing the Illusion of Erasure in Knowledge Editing for LLMs
    USarXiv cs.AI

    arXiv:2606.23276v3 Announce Type: replace-cross Abstract: Knowledge Editing (KE) has emerged as a frontier for updating specific facts in LLMs without costly retraining, but its reliability and underlying mechanisms remain poorly…

  78. SDS-LoRA: Overcoming Anisotropic Gradient Scaling in Low-Rank Adaptation
    USarXiv cs.AI

    arXiv:2606.16454v4 Announce Type: replace-cross Abstract: Low-Rank Adaptation (LoRA) enables efficient adaptation of large pretrained models to downstream tasks by parameterizing weight updates with low-rank matrices. In this paper, we…

  79. Real-Time Execution with Autoregressive Policies
    USarXiv cs.AI

    arXiv:2606.13355v2 Announce Type: replace-cross Abstract: Real-time execution, enabled by asynchronous inference that ensures both smooth action trajectories and fast reactivity, is critical for realistic deployments of large-scale…

  80. Subspace-Aware Sparse Autoencoders for Effective Mechanistic Interpretability
    USarXiv cs.AI

    arXiv:2606.06333v2 Announce Type: replace-cross Abstract: Sparse Autoencoders (SAEs) are widely used for mechanistic interpretability in large language models, yet their standard formulation assigns each latent feature a single decoder…