System translated (Gemini)

🤖 AI 速览

Today’s focus shifts from ‘capability’ to ‘controllability’: OpenAI is simultaneously pushing for AI governance baselines at state and federal levels and releasing the automated red-teaming GPT-Red, demonstrating the accelerated engineering of safety evaluations. …
📋 文章元数据
发布时间
2026-07-16
类型
ai-daily
字数
5802
阅读时长
28 min

2026-07-16 AI Daily | AI Competition Enters Foundational Safety Phase: GPT-Red Goes Live as Governance Frameworks Heat Up Link to heading

Today’s focus shifts from “capability” to “controllability”: OpenAI is simultaneously pushing for AI governance baselines at the state and federal levels while releasing the automated red-teaming tool, GPT-Red, indicating that safety evaluations are rapidly becoming more engineered. Meanwhile, the data and permission risks associated with AI programming tools are once again highlighted. The prerequisite for deployment is no longer just performance, but also boundaries, audits, and trust.

📖 In-depth Guide to This Issue’s Watch List Link to heading

The most important thread to follow today is “safety and interpretability”: OpenAI is discussing AI governance baselines at the US state and federal levels while also releasing GPT-Red, using automated red-teaming to expand vulnerability discovery. Combined with content related to Claude’s mechanistic interpretability, this is essential system-level reading for security and platform teams.

The second thread is “the reliability of models in professional scenarios.” Papers on CANDI-QA, G-SHARE, Braille proficiency evaluation, and disinformation detection all point to the same issue: general capabilities do not equal deployable capabilities. Scenarios like healthcare, finance, human factors diagnostics, and accessibility require more fine-grained contextual alignment and evaluation.

The third area of focus is “efficiency and engineering.” Techniques like RAG reranking distillation, text dataset distillation, prompt optimization stability, and point-in-time language models all address issues of cost, timeliness, and reliable backtesting. Additionally, it’s worth listening to the former Intel CEO’s retrospective and Lovable’s Vibe Coding interview to get the industry’s real perspective on AI programming and the transformation of large tech companies.

🌐 AI Hot Topics on X Link to heading

Topic 1: DeepMind CEO Predicts AGI in Few Years, Proposes Safety Body Link to heading

  • Category: AI · News
  • Overview: Trending for: 1 day ago, Related posts: 50,000
  • What happened: The CEO of DeepMind publicly predicted that AGI could emerge within a few years and proposed the establishment of a dedicated safety body to oversee related risks.
  • Why it’s important: This reflects that top AI institutions’ timelines for AGI are becoming a matter of practical discussion, pushing both “capability advancement” and “governance and safety” to the forefront of the industry’s core agenda.
  • Discussion summary: The main discussion on X revolves around whether the AGI timeline is overly optimistic, whether companies like DeepMind should lead the creation of safety rules, and whether the “safety body” is a practical governance solution or a PR move. Supporters emphasize the need for proactive planning, while skeptics worry that the capability race will outpace regulation.

Topic 2: Security Researcher Exposes Grok Build’s Unintended Repo Uploads Link to heading

  • Category: AI · News
  • Overview: Trending for: 1 day ago, Related posts: 8,300
  • What happened: A security researcher claims that xAI’s Grok Build feature may upload content from code repositories without the user’s knowledge, raising concerns about data leaks.
  • Why it’s important: This incident highlights the security risks of AI programming and building tools concerning access to local code, handling of private repositories, and permission boundaries, potentially affecting enterprise trust and adoption of AI development tools.
  • Discussion summary: Discussions on X are focused on whether this constitutes a serious security vulnerability, if user authorization prompts were adequate, xAI’s responsibility and response time, and how AI tools should balance convenience with privacy protection.

Topic 3: Booster Robotics Unveils T2 Humanoid for AI Developers Link to heading

  • Category: AI · News
  • Overview: Trending for: 13 hours ago, Related posts: 512
  • What happened: Booster Robotics has released the T2 humanoid robot platform for AI developers, designed for embodied intelligence, robot control, and application development.
  • Why it’s important: The product provides a hardware platform for AI models to transition from virtual environments to the real physical world, helping to advance the implementation of embodied intelligence, robot learning, and human-computer interaction applications.
  • Discussion summary: The discussion on X is mainly focused on T2’s openness, price, developer-friendliness, motor capabilities, and mass production feasibility. Some also question whether it has a competitive advantage over other humanoid robots like Tesla’s Optimus and Figure.

Summary of Today’s AI Public Opinion on X Link to heading

Today’s main narrative is that AI is simultaneously moving from “more powerful models” to “real-world deployment.” On one hand, the AGI timeline is being pushed forward by top institutions, forcing safety governance into the core agenda; on the other, AI programming tools and humanoid robots are beginning to interact with real code assets and the physical world. The consensus is that the industry can no longer only discuss capability improvements; permission boundaries, data security, regulatory mechanisms, and deployment reliability must be designed in parallel. The main points of disagreement are who is qualified to set safety rules, whether AGI predictions are overly optimistic, and whether new hardware platforms are a crucial gateway for embodied intelligence or just another case of concepts preceding reality. The potential risks are that the capability race may outpace the implementation of governance, development tools might expand the surface area for data leaks under the guise of convenience, and the premature commercialization of robot platforms could introduce software uncertainties into the real world.

💡 Influencer Insights Link to heading

AI Industry Pulse Daily: Insights from X Influencers Link to heading

Below is an in-depth analysis based on the past 24 hours of activity from key opinion leaders in the AI domain on the X platform.

1. Core Trend: AI Programming Enters the Deep End of “Workflow Automation” Link to heading

Today’s most significant common focus is not on the release of new models, but rather on the ecologization and practical application of AI Agent skills and workflows. Industry leaders are shifting from the initial exploration of “what models can do” to “how to make models consistently and efficiently work for me.”

  • Explosion of Agent Skill Ecosystem: Multiple bloggers have, by coincidence, focused on the creation, sharing, and commercialization of Skills.

    • @dotey shared in detail his journey of developing BaoCut (a video translation and editing Skill), emphasizing the best practice of a hybrid workflow: Agent + Skill + CLI + GUI. He also demonstrated a personal end-to-end AI-powered closed loop from “design -> prototype -> coding -> release.”
    • @Pluvio9yte followed up with a share, releasing a Skill for parsing WeChat official account articles (rn-wechat-extract), solving the “last mile” problem when Agents access content within specific ecosystems. He also recommended a search infrastructure designed for AI Agents, AnySearch, highlighting the pain point that traditional search returns summaries which Agents then need to parse a second time.
    • @ruanyf keenly observed that Xiaohongshu is attempting to build a Skill ecosystem (REDSkill), combining a social media platform with a Skill Hub. This is seen as laying the foundation for the community’s AI transformation and offers developers a new channel to reach a massive user base.
    • @vista8 also released his own GIF compression Skill.
    • Analyst Insight: This indicates that the competition in AI programming has moved beyond model capabilities to a contest centered on the toolchains, plugin ecosystems, and workflow integration capabilities of Agent platforms. Developers are encapsulating fragmented AI capabilities into reusable, shareable, and task-oriented Skills, which is a crucial step in Agents transitioning from “toys” to “productivity tools.”

  • The Battle of Approaches: Codex vs. Claude Code:

    • @dotey continuously monitored Codex’s developments, mentioning its user base has reached 8 million, and it integrates a complete workflow including GPT-5.6, Ultra multi-agent mode, and Cloudflare one-click publishing. He noted that GPT 5.6 Sol ranked second in the Agent arena, closely approaching Claude Fable 5.
    • @zhixianio, in a real-world scenario, deeply experienced Claude Fable 5, expressing profound impression with its stunning coding capabilities (completing 70% of work and autonomously optimizing design in 40 minutes), believing it has changed the development paradigm.
    • Analyst Insight: The market has formed two major camps: OpenAI’s Codex attracts users through Workspace integration and ecosystem expansion; Anthropic’s Claude Code (Fable 5) earns its reputation with top-tier single-task reasoning and code generation quality. Users (like @dotey) are beginning to develop their own combination strategies: using Opus/Fable for UI design and GPT 5.6 Sol for logic implementation.

2. Unique Perspectives and Industry Foresight Link to heading

  • The Ongoing Game Between “Knowledge Payment” and “AI Replacement Theory”:

    • @gefei55 deeply reflected on the value of community operations and knowledge payment, refuting the notion of their uselessness. He believes that in the AI era, paying to learn from experienced individuals saves valuable time, which is a necessary investment.
    • He also warned the programmer community: AI has improved production efficiency, but if demand research and product promotion are not done, it will only transform from “writing one unused App a month” to “spending ten thousand tokens a month to write 37 unused Apps,” accurately pointing out the trap of “Vibe Coding” and emphasizing the irreplaceable role of humans in demand validation, marketing, and other aspects.

  • In-Depth Reflection on Model Capabilities:

    • @zhixianio’s in-depth review of the edge model Gemma 4 12B Coder is highly valuable. He pointed out that this model performs on par with 35B models on simple tasks but collapses on complex, long-chain tasks (like generating a complete Tetris game) due to its parameter ceiling. This perspective precisely defines the capability boundaries of different scale models.
    • @Pluvio9yte strongly recommended Tencent’s Hy3 model, believing it possesses a generational advantage in its comprehensive capabilities across front-end development, aesthetic design, and content arrangement. It can generate websites comparable to several days’ work of a professional designer in 8 minutes, with outstanding output quality in terms of demand alignment and playability.
    • @vista8 raised a noteworthy compliance question: Why do many domestic model APIs no longer allow custom temperature parameters? He speculates this might be related to distillation technology, pointing to potential implicit technical limitations in model deployment within the industry.

  • Philosophical reflections on the relationship between humans and AI:

    • @vista8 quotes IBM’s famous 1979 saying: “A computer can never be held accountable, therefore a computer must never make a management decision.” He emphasizes that what will be truly scarce in the future are people who can identify real problems, make judgments with incomplete information, and take responsibility for the consequences, a view that resonates with @gefei55.
    • @lijigang points out that after long-term use of a specific model, your language style will acquire its “flavor” because the human brain’s neural network is highly influenced by context. He advises people to be wary of being molded by models and to curate their information inputs as carefully as they would curate books to maintain intellectual “heterogeneity.”

3. Recommended Tools & Resources Link to heading

  • AI Agent Workflow Tools:

    • AnySearch (recommended by @Pluvio9yte): A search infrastructure designed for AI Agents that can convert web pages into structured Markdown, addressing the pain point of traditional searches that only return links.
    • rn-wechat-extract Skill (open-sourced by @Pluvio9yte): Allows an Agent to bypass WeChat restrictions to directly download and parse the content of official account articles.
    • REDSkill (observed by @ruanyf): Xiaohongshu’s official Skill platform, which can be used to host and distribute your Agent Skills to reach a massive user base.
    • BaoCut Skill (from @dotey): Works with an Agent to implement a full workflow for video transcription, polishing (removing filler words), translation, and editing.
    • qiaomu-tiny-gif Skill (open-sourced by @vista8): A GIF compression tool specifically designed to meet the 10MB size and frame rate limits for WeChat official accounts.

  • Design & Frontend Resources:

    • Vibe Coding Design Terminology Website (recommended by @vista8): A website that catalogues the names of common Web/App components, animations, and design styles (like Liquid Glass, Neumorphism). It helps you accurately describe design requirements to an AI, improving the quality of Vibe Coding.

  • Noteworthy Models:

    • Tencent Hunyuan Hy3 Model (jointly recommended by @Pluvio9yte @ruanyf): API pricing is approximately $0.15 for input / $0.59 for output per million tokens. It performs excellently in frontend development, design, and Agent tasks. Community testing feedback indicates it surpasses several models with larger parameter counts, offering extremely high overall cost-effectiveness.
    • On-device models Gemma 4 + Gemma 4 Coder (in-depth testing by @zhixianio): Suitable for local deployment and efficient for small tasks, but not ideal for generating complex programs in a single pass. Google’s QAT (Quantization Aware Training) version is a key optimization direction for on-device deployment.

📚 Appendix: Today’s Watch List Source Updates Link to heading

Timeframe: Last 3 days; covers 22 sources; 25 updates in total

All-In Podcast (A_full) Link to heading

  • Former Intel CEO on What Went Wrong, What’s Next + Lovable CEO on the Real Promise of Vibe Coding
    • Published: 2026-07-16 04:55 Beijing Time
    • Abstract: - Airwallex is a leading global payment and financial platform for modern businesses, providing trusted solutions to manage everything from corporate accounts, payments, treasury, and expense management to embedded finance.
      • Plaud - If your work relies on conversations (meetings, deal flows, interviews, customer calls), Plaud helps you capture and organize everything with highly accurate AI-generated notes that go beyond simple summaries to highlight pain points, key decisions, next steps, and customizable summary templates.
      • Check out Plaud and enjoy up to 20% off with code ALLIN!
      • Also available on Amazon: (code: ALLIN20X).
      • Special thanks to the Raise Summit for hosting us in Paris.
    • EN Key Points:
      • (0:00) Former Intel CEO Pat Gelsinger joins Jason
      • (1:41) What Went Wrong at Intel
      • (15:19) Why a Taiwan Blockade Would Cripple the US Economy
      • (25:00) Lovable’s Anton Osika: One Million New Apps a Week

Stratechery by Ben Thompson (A_full) Link to heading

  • IBM Misses, IBM’s Mainframe Moat, IBM’s Many AI Problems
    • Publication Time: 2026-07-15 18:00 Beijing Time
    • Abstract: - IBM announced preliminary results that spooked the software market generally; this is a story, however, specifically about IBM and its mainframe franchise.
      • $15/month* or *$150/year.
      • Substantive analysis of the day’s news via three emails or podcasts per week.
      • Strategy interviews.
      • Interviews with leading public market CEOs, private company founders, and discussions with fellow analysts.
    • EN Key Points:
      • IBM announced preliminary results that spooked the software market generally; this is a story, however, specifically about IBM and its mainframe franchise.

OpenAI Blog (A_full) Link to heading

  • The US is advancing AI safety through state and federal action

    • Publication Time: 2026-07-15 20:00 Beijing Time
    • Abstract: - By Chris Lehane, Chief Global Affairs Officer.
      • From state capitals across the country to Washington D.C. and at international gatherings, a serious approach to the governance of frontier AI is taking shape.
      • Together they are advancing a democratic vision for AI.
      • California, New York, and, most recently, Illinois have forged sophisticated frontier safety legislation that helps move the country toward a common baseline for managing the most powerful AI systems.
      • These efforts reflect the dynamic behind what OpenAI calls reverse federalism: states helping to build a common direction through a common framework.
    • EN Key Points:
      • OpenAI outlines a “reverse federalism” approach to AI governance, where state laws help build a national framework for safe, democratic AI.

  • GPT-Red: Unlocking Self-Improvement for Robustness

    • Publication Time: 2026-07-15 18:00 Beijing Time
    • Abstract: - * Red teaming is crucial for finding vulnerabilities and improving the robustness of our models.
      • However, current approaches are not scalable, creating a bottleneck.
        • Commonly used robustness evaluations are already saturated by our latest models.
        • We need to develop methods that allow safety and alignment to scale along with model capabilities.
        • We trained GPT‑Red, an automated red teaming model that scales our ability to find vulnerabilities so we can fix them before wider deployment.
    • EN Key Points:
      • Explore GPT-Red, OpenAI’s automated red teaming system that uses self-play to improve AI safety, alignment, and prompt injection robustness.

Two Minute Papers (B_intro+search) Link to heading

  • Claude’s Brain Has A Secret… And Scientists Found It
    • Publication Time: 2026-07-15 21:58 Beijing Time
    • Abstract: - ❤️ Check out Lambda and sign up for their GPU Cloud here:.
      • 📝 The paper is available here:.
      • Paper on the vision change of reindeers -.
      • Adam Bridges, Benji Rabhan, B Shang, Cameron Navor, Charles Ian Norman Venn, Christian Ahlin, Eric T, Fred R, Gordon Child, Juan Benet, Michael Tedder, Owen Skarpness, Richard Sundvall, Ryan Stankye, Shawn Becker, Steef, Taras Bobrovytsky, Tazaur Sagenclaw, Tybie Fitzhugh, Ueli Gallizzi.
    • EN Key Points:
  • ❤️ Check out Lambda here and sign up for their GPU Cloud:
  • 📝 The paper is available here:
  • Paper for reindeer vision change -
  • 🙏 We would like to thank our generous Patreon supporters who make Two Minute Papers possible:

ArXiv cs.CL (B_intro+search) Link to heading

  • Scaling Point-in-Time Language Models

    • Publication Time: 2026-07-15 12:00 Beijing Time
    • Abstract: - arXiv:2607.11889v1 Announcement Type: New.
      • Abstract: Large language models trained on unrestricted internet corpora inevitably embed information from the future, introducing lookahead bias that compromises the validity of backtesting and causal inference in finance and the social sciences.
      • Point-in-time language models (trained exclusively on text available up to each calendar date) eliminate this leakage by construction, but existing efforts typically yield models that lag significantly behind their unconstrained counterparts.
      • We show that this performance gap can be substantially narrowed through scale.
    • EN Key Points:
      • arXiv:2607.11889v1 Announce Type: new
      • Abstract: Large language models trained on unrestricted internet corpora inevitably embed information from the future, introducing lookahead bias that compromis…
      • Point-in-time language models–trained exclusively on text available up to each calendar date–eliminate this leakage by construction, but existing efforts typi…
      • We show that this performance gap can be substantially narrowed through scale

  • CANDI: Contextual Alignment for Niche Domains Question Answering

    • Publication Time: 2026-07-15 12:00 Beijing Time
    • Abstract: - arXiv:2607.11891v1 Announcement Type: New.
      • Abstract: The deployment of Large Language Models (LLMs) in specialized domains like medical diagnostics and financial advisory requires evaluating capabilities that go beyond general knowledge.
      • Traditional question-answering benchmarks often fail to capture the nuanced contextual grounding, user awareness, and domain understanding these fields require.
      • To address this, we introduce CANDI-QA (Contextual Alignment for Niche Domains Question Answering), a novel dataset for evaluating LLMs on delivering accurate, context-sensitive, and user-aligned answers in professional environments.
    • EN Key Points:
      • arXiv:2607.11891v1 Announce Type: new
      • Abstract: The deployment of large language models (LLMs) in specialized domains like medical diagnostics and financial advisory necessitates evaluating capabili…
      • Traditional question-answering benchmarks often fail to capture the nuanced contextual grounding, user awareness, and domain understanding these fields require
      • To address this, we introduce CANDI-QA (Contextual Alignment for Niche Domains Question Answering), a novel dataset evaluating LLMs on delivering accurate, cont…

  • G-SHARE: A Guideline-Based Structured Reasoning Framework for Human-Factor Event Diagnosis

    • Publication Time: 2026-07-15 12:00 Beijing Time
    • Abstract: - arXiv:2607.11892v1 Announcement Type: New.
      • Abstract: Human-factor event diagnosis is crucial for learning from operational events in nuclear power plants, but its quality largely depends on experts’ interpretation of narrative reports and guideline-based reasoning. Existing data-driven or one-shot Large Language Model approaches often lack structured reasoning, have limited consistency with formal diagnostic guidelines, and may produce logically inconsistent conclusions.
      • To address this issue, this study proposes G-SHARE, a guideline-based structured reasoning framework that operationalizes the CNNP nine-step human-factor event diagnosis guideline as a multi-stage diagnostic process. Comprising evidence extraction, stepwise diagnostic reasoning, and post-hoc consistency repair, the framework enables explicit use of reported evidence, generation of intermediate rationales, and logical validation of diagnostic outputs.
      • A dataset of real human-factor event reports was constructed from Chinese nuclear industry sources and evaluated using a gold-standard subset annotated by domain experts.
    • EN Key Points:
      • arXiv:2607.11892v1 Announce Type: new
      • Abstract: Human-factor event diagnosis is essential for learning from operational events in nuclear power plants, yet its quality depends strongly on expert int…
      • To address this issue, this study proposes G-SHARE, a guideline-based structured reasoning framework that operationalizes the CNNP nine-step human-factor event…
      • A dataset of real human-factor event reports was constructed from Chinese nuclear industry sources, and a gold-standard subset annotated by domain experts was u…

  • I’m Sorry, but I Can’t Help with Braille: Revealing Accessibility Failures in State-of-the-Art LLMs

    • Publication Time: 2026-07-15 12:00 Beijing Time
    • Abstract: - arXiv:2607.11893v1 Announcement Type: New.
      • Abstract: Large Language Models (LLMs) perform strongly on many language tasks, but their capabilities in structurally constrained, accessibility-critical modalities (e.g., Braille) remain unclear.
      • We evaluate state-of-the-art LLMs on bidirectional Korean Braille translation using a human-annotated dataset.
      • Despite expectations that multilingual, instruction-tuned models can generalize to Braille via text representations, we find consistently poor, unstable outputs, and significant divergence from human judgment.
    • EN Key Points:
      • arXiv:2607.11893v1 Announce Type: new
      • Abstract: Large Language Models (LLMs) perform strongly on many language tasks, but their capability in structurally constrained, accessibility-critical modalit…
      • We evaluate state-of-the-art LLMs on bidirectional Korean-Braille translation using a human-annotated dataset
      • Despite expectations that multilingual, instruction-tuned models can generalize to Braille via text representations, we find consistently poor, unstable outputs…

  • Graph-Based Detection of Disinformation Narrative Diffusion between Russian and Ukrainian Telegram Channels

    • Publish Date: 2026-07-15 12:00 Beijing Time
    • Abstract: - arXiv:2607.11894v1 Announce Type: New.
      • Abstract: Detecting disinformation narratives on social media is challenging due to the expanding scale, rapid evolution, and linguistic variability of online content.
      • We propose a graph-based framework for identifying and analyzing disinformation narratives in Telegram ecosystems by combining weak supervision with propagation graph analysis.
      • The method aggregates semantically related claims into narrative-level clusters and models their diffusion across interconnected channels.
    • EN Key Points:
      • arXiv:2607.11894v1 Announce Type: new
      • Abstract: Detecting disinformation narratives on social media is challenging due to the scale of amplification, rapid evolution, and linguistic variability of o…
      • We propose a graph-based framework for identifying and analyzing disinformation narratives in Telegram ecosystems by combining weak supervision with propagation…
      • The approach aggregates semantically related claims into narrative-level clusters and models their diffusion across interconnected channels

  • TAKE: Trajectory-Aware Knowledge Estimation for Text Dataset Distillation

    • Publish Date: 2026-07-15 12:00 Beijing Time
    • Abstract: - arXiv:2607.11898v1 Announce Type: New.
      • Abstract: Large-scale text corpora have become a quiet bottleneck in modern NLP, not just in storage, but in the accumulated costs of training, fine-tuning, and continuous learning.
      • We propose a text dataset distillation framework that reduces corpora to as little as 0.1% of their original size while preserving downstream task fidelity.
      • We approach distillation through the lens of influence functions, which quantify each sample’s contribution to the downstream objective, a natural and principled basis for selection.
    • EN Key Points:
      • arXiv:2607.11898v1 Announce Type: new
      • Abstract: Large-scale text corpora have become a quiet bottleneck in modern NLP, not just in storage, but in the accumulated cost of training, fine-tuning, and…
      • We propose a text dataset distillation framework that reduces corpora to as little as 0.1% of their original size while preserving downstream task fidelity
      • We approach distillation through the lens of influence functions, which quantify each sample’s contribution to the downstream objective, a natural and principle…

  • Transforming LLMs into Efficient Cross-Encoders via Knowledge Distillation for RAG Reranking

    • Publish Date: 2026-07-15 12:00 Beijing Time
    • Abstract: - arXiv:2607.11933v1 Announce Type: New.

  • Abstract: Cross-encoders achieve high reranking accuracy in Retrieval-Augmented Generation (RAG) pipelines but impose quadratic inference costs that limit real-time deployment.

  • We address this by fine-tuning LLaMA 3 (8B) as a drop-in reranker using a two-stage pipeline: supervised fine-tuning on a custom query-document relevance dataset via the Unsloth framework and LoRA adapters, followed by 4-bit quantization for efficient inference.

  • The resulting model replaces the cross-encoder in a dual-retriever RAG pipeline combining BM25 and dense vector search.

    • EN Key Points:
      • arXiv:2607.11933v1 Announce Type: new
      • Abstract: Cross-encoders achieve high reranking accuracy in Retrieval-Augmented Generation (RAG) pipelines but impose quadratic inference costs that limit real-…
      • We address this by fine-tuning LLaMA 3 (8B) as a drop-in reranker using a two-stage pipeline: supervised fine-tuning on a custom query-document relevance datase…
      • The resulting model replaces the cross-encoder in a dual-retriever RAG pipeline combining BM25 and dense vector search

  • MAGE: Understanding Stability-Performance Trade-offs in Multi-component Prompt Optimization

    • Publication Time: 2026-07-15 12:00 Beijing Time
    • Abstract: - arXiv:2607.11944v1 Announce Type: new.
      • Abstract: How do different components of iterative prompt optimization interact, and what happens when they are combined?
      • We investigate this through MAGE (Memory-Augmented Goal-directed Prompt Evolution), a controlled analysis framework for studying component interaction in prompt optimization.
      • MAGE is not proposed as a superior optimizer in absolute terms; it integrates episodic memory, multi-objective Pareto selection, and adaptive evaluation as a platform for controlled ablations.
    • EN Key Points:
      • arXiv:2607.11944v1 Announce Type: new
      • Abstract: How do different components of iterative prompt optimization interact, and what happens when they are combined
      • We investigate this through MAGE (Memory-Augmented Goal-directed Prompt Evolution), a controlled analysis framework for studying component interaction in prompt…
      • MAGE is not proposed as a superior optimizer in absolute terms; it integrates episodic memory, multi-objective Pareto selection, and adaptive evaluation as a pl…

  • Belief-reality separation lives in routing over a shared value slot in language models

    • Publication Time: 2026-07-15 12:00 Beijing Time
    • Abstract: - arXiv:2607.11945v1 Announce Type: new.
      • Abstract: Capable language models separate characters’ beliefs from facts: told “Anna believes the cup is blue; in fact it is red,” they answer blue for Anna and red for the world.
      • Where in the computation does this separation live?
      • We show it lives on two separable mechanisms in two locations.
    • EN Key Points:
      • arXiv:2607.11945v1 Announce Type: new

  • Abstract: Capable language models hold what a character believes apart from what is true: told “Anna believes the cup is blue; in reality it is red,” they answe…

  • Where in the computation does that separation live

  • We show it rests on two separable mechanisms at two positions

  • Hybrid Continual Learning for Low-Resource Australian Aboriginal Language Identification

    • Release Time: 2026-07-15 12:00 Beijing Time
    • Abstract: - arXiv:2607.11946v1 Announcement Type: new.
      • Abstract: Language identification is an important step toward integrating endangered Australian Aboriginal languages (AALs) into speech technologies supporting language revitalization and digital inclusion.
      • However, extreme data scarcity limits model performance.
      • Transfer learning from high-resource languages shows promise but often suffers from catastrophic forgetting when adapting to new languages.
    • EN Highlights:
      • arXiv:2607.11946v1 Announce Type: new
      • Abstract: Language identification is an important step toward integrating endangered Australian Aboriginal languages (AALs) into speech technologies supporting…
      • However, extreme data scarcity limits model performance
      • Transfer learning from high-resource languages shows promise but often suffers from catastrophic forgetting when adapting to new languages

ArXiv cs.LG (B_intro+search) Link to heading

  • OmniPMNet: Bridging discrete and gridded PM10 forecasts via omni-query neural processes

    • Release Time: 2026-07-15 12:00 Beijing Time
    • Abstract: - arXiv:2607.11896v1 Announcement Type: new.
      • Abstract: Forecasting particulate matter (PM10) requires both station-scale accuracy and continuous spatial fields, especially during severe dust storms.
      • Chemical transport models (CTMs) provide gridded forecasts but retain local biases, whereas graph neural networks (GNNs) track monitoring sites well for short periods but do not produce gridded output.
      • Here, we propose OmniPM-Net, a fusion model based on Convolutional Conditional Neural Processes (ConvCNP), which reconciles these two forecast types in a shared spatial representation.
    • EN Highlights:
      • arXiv:2607.11896v1 Announce Type: new
      • Abstract: Forecasting particulate matter (PM10) requires both station-scale accuracy and continuous spatial fields, especially during severe dust storms
      • Chemical transport models (CTMs) provide gridded forecasts but retain local biases, whereas graph neural networks (GNNs) track monitoring sites well at short le…

  • Here we present OmniPM-Net, a Convolutional Conditional Neural Process (ConvCNP)-based fusion model that reconciles these two forecast types within a shared spa…

  • Semidirect Fourier Delta Attention: Phase-Controlled Delta Memory with Constructive Chunk-WY Kernels

    • Published: 2026-07-15 12:00 Beijing Time
    • Abstract: - arXiv:2607.11897v1 Announce Type: new.
      • Abstract: Linear attention replaces softmax attention’s growing KV cache with a fixed recurrent state, but this compression limits exact state tracking and long context memory.
      • We introduce \emph{Semidirect Fourier Delta Attention} (SFDA), a phase-controlled generalization of Kimi Delta Attention that replaces real diagonal decay with block-rotational Fourier control: [ S_t=(I-\beta_t k_tk_t^)\Lambda_tS_{t-1}+\beta_tk_tv_t^, \qquad \Lambda_t=\diag(\alpha_t\odot e^{i\theta_t}).
      • ] Our main result is a constructive chunk-WY factorization for the product (A_t=\Lambda_t-u_tr_t^), giving [ A_t\cdots A_1=\Gamma_t-Y_tM_tW_t^ ], with rank growth confined within a fixed block.
    • EN Key Points:
      • arXiv:2607.11897v1 Announce Type: new
      • Abstract: Linear attention replaces softmax attention’s growing KV cache with a fixed recurrent state, but this compression limits exact state tracking and long…
      • We introduce \emph{Semidirect Fourier Delta Attention} (SFDA), a phase-controlled generalization of Kimi Delta Attention that replaces real diagonal decay with…
      • ] Our main result is a constructive chunk-WY factorization for products (A_t=\Lambda_t-u_tr_t^), giving [ A_t\cdots A_1=\Gamma_t-Y_tM_tW_t^ ] with rank g…

  • Repairing Shape-Prior Shortcuts in Long-Range Single-Shot Fringe Projection Profilometry

    • Published: 2026-07-15 12:00 Beijing Time
    • Abstract: - arXiv:2607.11928v1 Announce Type: new.
      • Abstract: Single-shot fringe projection profilometry (FPP) networks that directly regress depth can exploit shape-prior shortcuts, recovering depth from object boundaries instead of from the fringe phase.
      • On a photorealistic synthetic benchmark (15,600 fringe images, 50 objects at 1.5-2.1 m range), the best UNet baseline plateaus at 14.54 mm object mean absolute error (MAE), and neither more data nor more capacity removes the shortcut, as neither changes the hypothesis space searched by the optimizer.
      • We introduce PhiCalNet, which outputs a wrapped phase representation $(\sin\phi, \cos\phi)$ and maps it to depth through a fixed, differentiable calibration layer, architecturally eliminating the shape-prior solution instead of penalizing it through loss.
    • EN Key Points:
      • arXiv:2607.11928v1 Announce Type: new

  • Abstract: Single-shot fringe projection profilometry (FPP) networks that regress depth directly can exploit a shape-prior shortcut, recovering depth from object…

  • On a photorealistic synthetic benchmark (15,600 fringe images, 50 objects at 1.5-2.1 m standoff), the best such UNet baseline plateaus at 14.54 mm object mean a…

  • We introduce PhiCalNet, which outputs a wrapped-phase representation $(\sin\phi, \cos\phi)$ and maps it to depth through a fixed differentiable calibration laye…

  • Qubit-Efficient Quantum Search for Hyperdimensional Decomposition via Logarithmic Encoding

    • Publication Time: 2026-07-15 12:00 Beijing Time
    • Abstract: - arXiv:2607.11936v1 Announce Type: new.
      • Hyperdimensional Computing (HDC) represents symbols using high-dimensional hypervectors of dimension $D$.
      • In hypervector decomposition, the objective is to recover $F$ constituent hypervectors, each drawn from a codebook of size $N$, from a bound target hypervector.
      • This requires searching over $N^F$ candidate tuples, making the task computationally prohibitive at scale.
    • EN Key Points:
      • arXiv:2607.11936v1 Announce Type: new
      • Abstract: Hyperdimensional Computing (HDC) represents symbols using high-dimensional hypervectors of dimension $D$
      • In hypervector decomposition, the objective is to recover $F$ constituent hypervectors, each drawn from a codebook of size $N$, from a bound target hypervector
      • This requires searching over $N^F$ candidate tuples, making the task computationally prohibitive at scale

  • Mirror Horizon: Viable Path Entropy as a Measure of Bounded Reflection

    • Publication Time: 2026-07-15 12:00 Beijing Time
    • Abstract: - arXiv:2607.11937v1 Announce Type: new.
      • Mirror Theory proposes that an intelligent system should be studied not only by what it represents, but also by what coherent continuations it can sustain under repeated reflection.
      • We operationalize this claim via \emph{viable path entropy} (VPE), a finite-budget measure of verified continuation capability.
      • Given a mirror state, a rollout protocol, a verifier, and a pattern map, VPE decomposes bounded capability into two components: the probability of reaching a viable continuation, and the diversity of verified continuation patterns achieved in successful rollouts.
    • EN Key Points:
      • arXiv:2607.11937v1 Announce Type: new
      • Abstract: Mirror Theory proposes that an intelligent system should be studied not only by what it represents, but by what coherent continuations it can sustain…

  • We make this claim operational through \emph{viable path entropy} (VPE), a finite-budget measure of verified continuation capacity

  • Given a mirror state, a rollout protocol, a verifier, and a mode map, VPE decomposes bounded capability into two parts: the probability of reaching a viable con…

  • Mathematics of Data Science

    • Publication Time: 2026-07-15 12:00 Beijing Time
    • Abstract:- arXiv:2607.11938v1 Announce Type: new.
      • Abstract: This book is about the mathematical foundations of data science.
      • The Curses, Blessings, and Surprises of High Dimensions.
      • Singular Value Decomposition and Principal Component Analysis.
    • EN Key Points:
      • arXiv:2607.11938v1 Announce Type: new
      • Abstract: This book is about the mathematical foundations of data science
      • Introduction
      • Curses, Blessings, and Surprises in High Dimensions

  • CARE-LoRA: Compressed Activation REconstruction for Memory-Efficient LoRA

    • Publication Time: 2026-07-15 12:00 Beijing Time
    • Abstract:- arXiv:2607.11940v1 Announce Type: new.
      • Abstract: As the scale of large pre-trained models continues to grow, fine-tuning them under limited memory budgets has become increasingly challenging.
      • Low-Rank Adaptation (LoRA), currently one of the most widely adopted parameter-efficient fine-tuning (PEFT) methods, mitigates this challenge by optimizing only low-rank adaptation matrices, thereby significantly reducing the number of trainable parameters.
      • With the parameter overhead substantially reduced, the activations retained for backpropagation have emerged as the primary remaining memory bottleneck during LoRA fine-tuning.
    • EN Key Points:
      • arXiv:2607.11940v1 Announce Type: new
      • Abstract: As the scale of large pre-trained models continues to grow, fine-tuning them under limited memory budgets has become increasingly challenging
      • Low-Rank Adaptation (LoRA), currently one of the most widely adopted parameter-efficient fine-tuning (PEFT) methods, mitigates this challenge by optimizing only…
      • With the parameter overhead substantially reduced, the activations retained for backpropagation have emerged as the primary remaining memory bottleneck during L…

  • How Query Visibility Changes KV-Cache Compression Rankings: A Matched-Budget Audit

    • Publication Time: 2026-07-15 12:00 Beijing Time
    • Abstract:- arXiv:2607.11942v1 Announce Type: new.
      • Abstract: KV-cache compression methods are primarily evaluated by compressing queries attached to the context beforehand—a query-aware protocol.
      • However, the economic case for compressing KV-caches is reuse: compress a document once, and many future questions can be answered against it.
      • In this deployment, compression must be query-agnostic—before any questions arise.
    • EN Key Points:
      • arXiv:2607.11942v1 Announce Type: new

  • Abstract: KV-cache compression methods are predominantly evaluated with the query appended to the context before compression – a query-aware protocol

  • Yet the economic case for a compressed KV cache is reuse: compress a document once, answer many future questions against it

  • In that deployment, compression must happen query-agnostic – before any question is seen

  • BattVAE-GP: Generative Modeling of Long-Horizon Battery Degradation with Uncertainty Quantification

    • Publication Time: 2026-07-15 12:00 Beijing Time
    • Abstract: - arXiv:2607.11943v1 Announcement Type: new.
      • Abstract: Physics-based long-horizon battery degradation simulations provide mechanistic insights, but remain computationally expensive, limiting their use for intensive exploration of operating conditions during extended cycle life.
      • Here, we propose a hybrid physics-probabilistic learning framework for surrogate modeling of lithium-ion battery degradation trajectories at unseen charging rates.
      • Cycle-resolved degradation data generated with a DFN/P2D electrochemical model in PyBaMM is first transformed into capacity-aligned voltage and derivative features and encoded using a Variational Autoencoder (VAE).
    • EN Highlights:
      • arXiv:2607.11943v1 Announce Type: new
      • Abstract: Long-horizon physics-based simulations of battery degradation provide mechanistic insight but remain computationally expensive, limiting their use for…
      • Here, we propose a hybrid physics-probabilistic learning framework for surrogate modeling of lithium-ion battery degradation trajectories at unseen charging rat…
      • Cycle-resolved degradation data generated with a DFN/P2D electrochemical model in PyBaMM are first transformed into capacity-aligned voltage and derivative feat…

  • Generalized Distribution-Free Semi-Supervised Learning with Risk Rewrite

    • Publication Time: 2026-07-15 12:00 Beijing Time
    • Abstract: - arXiv:2607.11947v1 Announcement Type: new.
      • Abstract: Typical semi-supervised learning (SSL) methods rely on distributional assumptions, and their performance degrades when these assumptions are violated.
      • While PNU learning (a risk rewriting method) offers a distribution-free alternative, it is limited to binary classification, and its variance optimality remains unclear.
      • In this paper, we propose a general framework that constructs an unbiased risk estimator using a linear combination of component risks, incorporating PNU learning and extending it to multi-class classification.
    • EN Highlights:
      • arXiv:2607.11947v1 Announce Type: new
      • Abstract: Typical semi-supervised learning (SSL) methods rely on distributional assumptions, and their performance degrades when these are violated

  • While PNU learning, a risk rewriting method, offers a distribution-free alternative, it is restricted to binary classification and its variance optimality remai…

  • In this paper, we propose a generalized framework that constructs unbiased risk estimators using linear combinations of component risks, subsuming PNU learning…