System translated (Gemini)

🤖 AI 速览

Today’s main theme is AI moving from capability demonstration to engineering acceptance. GPT-5.6 enters Kiro, strengthening the integrated process of planning, building, reviewing, and testing; small models and local inference continue to push up cost re-evaluation. Meanwhile, research into …
📋 文章元数据
发布时间
2026-08-25
类型
ai-daily
字数
7991
阅读时长
38 min

2026-08-25 AI Daily | AI Agents Are No Longer Just About Capability: Kiro Closes the Development Loop, While Small Models and Reliability Emerge as New Variables Link to heading

Today’s main theme is the shift of AI from capability showcases to engineering acceptance. GPT-5.6’s integration into Kiro enhances the unified process of planning, building, reviewing, and testing. Concurrently, small models and local inference continue to prompt a re-evaluation of costs. Meanwhile, research into long-context efficiency, training transfer, and security backdoors serves as a reminder to the industry: after usability, trust is the next major hurdle.

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

The most noteworthy topic today is “The Cost-Effectiveness Tipping Point for AI Programming Agents.” Kiro’s introduction of GPT-5.6 (Sol/Terra/Luna) shifts the focus from one-off capability showcases to creating an engineering workflow that integrates planning, building, reviewing, and testing into a process with fewer iterations. At the same time, discussions around the small model Qwen3.8-27B are prompting teams to re-evaluate the boundaries of localized, low-cost inference.

The second major theme is “Long Context and Model Reliability.” BF1 sparse attention, subconscious feature transfer in optimizer states, and “wrong-physics” backdoors in neural PDE operators each serve as reminders from the perspectives of efficiency, training dynamics, and security poisoning that a usable model is not necessarily a trustworthy one.

On the application front, the focus should be on XAI and high-risk scenarios. Applications like bankruptcy prediction, public health forecasting, thermal comfort control, mental health bots, and personality extraction for digital twins are pushing AI into front-line decision-making roles. The safety assessment of therapy bots for adolescents is particularly crucial reading for product and compliance teams.

🌐 AI Hot Topics on X Link to heading

Topic 1: Anthropic Launches Enterprise-Managed Auth for MCP Connectors Link to heading

  • Category: AI · News
  • Overview: Hot Topic Time:, Related Posts: 87
  • What it is: Anthropic has launched enterprise-managed authentication for MCP connectors, allowing businesses to centrally manage permissions and access control when AI connects to external tools and data sources.
  • Why it’s important: This is significant for the AI field because MCP is becoming a universal interface for models to access enterprise systems. Authentication and permission governance are critical infrastructure for deploying AI agents and tool calls at scale and securely.
  • Discussion summary: Discussions on X are focused on whether this will significantly boost Claude’s competitiveness in enterprise scenarios, if it can lower the barrier for deploying MCP in highly regulated industries like finance, and whether Anthropic’s recent frequent releases of enterprise and agent capabilities are merely feature stacking or are building a genuine ecosystem advantage.

Topic 2: Square Root Puzzles Trick Social Media Solvers Link to heading

  • Category: AI · Entertainment
  • Overview: Hot Topic Time: 9 hours ago, Related Posts: 66,000
  • What it is: A type of square root puzzle that appears simple but contains hidden traps related to order of operations, domain, or symbols has gone viral on X, prompting a large number of users and AIs to attempt to solve it.
  • Why it’s important: These kinds of problems expose weaknesses in AI’s mathematical reasoning, problem comprehension, and ability to avoid intuitive answers. They are often used to test whether a model truly possesses robust logical reasoning capabilities.
  • Discussion summary: The discussion on X centers on whether the correct answer depends on mathematical conventions or the wording of the problem. Some see it as a fun brain teaser, while others criticize it for using ambiguity to create controversy. Other users are comparing the performance of different AI models in solving the puzzle.

Topic 3: Maye Musk Enjoys Fashion and Sights on Shanghai Visit Link to heading

  • Category: AI · Entertainment
  • Overview: Hot Topic Time: 5 hours ago, Related Posts: 2,900
  • What it is: Maye Musk visited city attractions and participated in fashion-related events in Shanghai, drawing attention on the X platform.
  • Why it’s important: The event itself is not an AI technological advancement, but because of its connection to the public image of Elon Musk and his AI, automotive, and technology businesses, it has been interpreted by some users as a signal of interaction between the Musk family and the Chinese market.
  • Discussion summary: The discussion on X is mainly focused on Maye Musk’s Shanghai itinerary, her fashion choices, and the presentation of the city’s image. Some also connect the visit to the business relationships of Tesla, xAI, and Musk in China, though there is disagreement on whether it holds any real industry significance.

Topic 4: Apodex Releases 1.1 with Multi-Agent Teams for Complex Tasks Link to heading

  • Category: AI · News
  • Overview: Hot Topic Time: 6 hours ago, Related Posts: 1,100
  • What it is: Apodex has released version 1.1, introducing multi-agent teams that can collaborate to handle complex tasks.
  • Why it’s important: Multi-agent collaboration is seen as a key direction for enhancing AI’s capabilities in task planning, division of labor, and complex problem-solving, which could impact the implementation of enterprise-level automation and agent applications.
  • Discussion summary: Discussions on X are focused on whether this feature genuinely improves the completion rate of complex tasks, what advantages it has over existing agent frameworks, and the challenges of multi-agent systems in terms of cost, stability, and controllability.

Topic 5: Nvidia Partners with Poolside on $6 Billion Deal for Powerful Open AI Model Link to heading

  • Category: AI · News
  • Overview: Trending for: 1 day ago, Related posts: 4800
  • What it is: Nvidia has reportedly partnered with AI startup Poolside in a deal valued at approximately $6 billion to develop a powerful open AI model.
  • Why it matters: This indicates that computing power giants are further entering the foundational model competition. The open model approach may also gain stronger momentum with Nvidia’s funding, chips, and ecosystem support.
  • Discussion summary: Discussions on X are mainly focused on whether the deal will change the landscape of open models, whether Poolside can compete with leading companies like OpenAI and Anthropic, and whether Nvidia’s influence in the AI industry chain is becoming overly concentrated.

Topic 6: Xiaomi Unveils AI Cube Prototype with Custom Chips for Local AI Link to heading

  • Category: AI · News
  • Overview: Trending for: 16 hours ago, Related posts: 4500
  • What it is: Xiaomi showcased an AI Cube prototype device featuring its self-developed custom chips, focused on local, on-device AI computing.
  • Why it matters: This shows that consumer electronics manufacturers are accelerating the shift of AI capabilities from the cloud to local devices to enhance privacy, reduce latency, and decrease reliance on cloud computing. It also intensifies the competition in on-device AI chips and ecosystems.
  • Discussion summary: Discussions on X are centered on whether Xiaomi’s self-developed chip capabilities are sufficient to support high-quality local AI, the practical application scenarios and price of the AI Cube, and its competitiveness compared to on-device AI solutions from Apple, Nvidia, and Qualcomm.

Topic 7: AI Rankings Favor Cheaper Models Over Anthropic’s Premium Options Link to heading

  • Category: AI · News
  • Overview: Trending for: 2 days ago, Related posts: 21000
  • What it is: Recent AI rankings circulating on X show that lower-priced models are outperforming Anthropic’s high-priced flagship models on the leaderboards.
  • Why it matters: This reflects a growing emphasis in the AI field on cost-effectiveness, real-world performance, and cost efficiency, which could impact model procurement, product positioning, and the competitive landscape of the industry.
  • Discussion summary: The discussion focuses on whether the leaderboards genuinely represent actual capabilities, whether cheaper models are now “good enough,” and if the additional performance and safety advantages of high-priced models are worth the premium. Some also question the ranking methodology’s bias towards cost factors.

Topic 8: 49ers Owner Jed York Pleads No Contest in Ohio Prostitution Sting Link to heading

  • Category: AI · Sports
  • Overview: Trending for: 9 hours ago, Related posts: 89000
  • What it is: San Francisco 49ers owner Jed York was reportedly arrested in a prostitution sting in Ohio and pleaded no contest to lesser charges.
  • Why it matters: While the incident itself has no direct connection to AI technology or the industry, its inclusion in the “AI · Sports” trending list reflects potential misjudgment or generalization issues in the platform’s topic classification, automated tagging, and information recommendation systems.
  • Discussion summary: Discussions on X center on the changing wording of media headlines, whether the case is being downplayed, the accountability of public figures, and the accuracy of the reporting. Some also question why the trending topic classification associated a sports scandal with AI.

Topic 9: Enes Kanter Freedom Ejected Over Women’s Sports Shirt at WNBA Game Link to heading

  • Category: AI · Sports
  • Overview: Trending for: 1 day ago, Related posts: 827000
  • What it is: Former NBA player Enes Kanter Freedom was removed by security from a WNBA game between the Chicago Sky and Indiana Fever after wearing a T-shirt that read “Woman: adult human female” and getting into a verbal altercation with player Natasha Cloud.
  • Why it matters: While the incident itself pertains to sports and gender issues, its significance in the AI domain lies in its highlighting of how social media algorithms can amplify highly polarized content, and the boundary challenges faced by content moderation, hate speech detection, fact-checking, and public issue recommendation systems.
  • Discussion summary: Discussions on X are mainly split into two camps: one side believes Kanter was simply expressing his stance on fairness in women’s sports and that his ejection reflects a shrinking space for speech; the other side argues his actions were provocative and exclusionary towards the transgender community, and that the league and players have the right to maintain an inclusive environment. The debate centers on who initiated the conflict, whether the T-shirt’s content constitutes an offense, the rules for transgender athletes’ participation, and whether the platform is amplifying division.

Topic 10: Manchester City Joins Tottenham in Race for Liverpool’s Cody Gakpo Link to heading

  • Category: AI · Sports
  • Overview:Trending since:1 day ago,Related posts:50000
  • What it is:Manchester City has reportedly joined Tottenham in the race to sign Liverpool forward Cody Gakpo, sparking discussions on X.
  • Why it matters:While not directly related to AI advancements, this event highlights the value of sports transfer rumors for public opinion monitoring, recommendation algorithms, and sports data analysis on social media platforms.
  • Discussion summary:The discussion focuses on the credibility of the rumor, whether Liverpool would sell Gakpo, the actual needs of Manchester City and Tottenham, and the potential impact of the transfer on the Premier League’s competitive landscape.

Topic 11:Manchester City Agree €40M Deal for Palmeiras Winger Allan Elias Link to heading

  • Category:AI · Sports
  • Overview:Trending since:1 day ago,Related posts:21000
  • What it is:According to ESPN, Manchester City has reached an agreement with Palmeiras for winger Allan Elias for a fee of around €40 million.
  • Why it matters:High-profile sports transfer news like this quickly generates extensive social media discussion, making it a good case for observing AI’s performance in real-time information extraction, event aggregation, public opinion monitoring, and multi-language summarization.
  • Discussion summary:Discussions on X are mainly about the authenticity of the news, discrepancies in the transfer fee, the player’s skill and suitability for Manchester City’s system, and whether this deal could trigger a chain reaction of subsequent transfers.

Topic 12:Chelsea Edge Fulham 3-2 in Dramatic Season Opener Under Alonso Link to heading

  • Category:AI · Sports
  • Overview:Trending since:4 hours ago,Related posts:249000
  • What it is:Chelsea narrowly defeated Fulham 3-2 in a dramatic season opener, drawing attention to Alonso’s debut as manager.
  • Why it matters:Such high-profile sports events demonstrate the value of AI in real-time trend identification, match summary generation, and public opinion analysis. It also provides a typical scenario for multi-modal sports content understanding.
  • Discussion summary:Discussions on X focus on Chelsea’s crucial goals and defensive issues, Alonso’s coaching performance, and whether referee decisions and the match’s tempo affected the final outcome.

Topic 13:Hermes Agent’s HUD Lets Gamers Chat with AI Over WoW Link to heading

  • Category:AI · Entertainment
  • Overview:Trending since:9 hours ago,Related posts:245
  • What it is:Hermes Agent has launched a HUD tool for World of Warcraft that allows players to talk directly with an AI assistant during gameplay.
  • Why it matters:This shows that AI is evolving from a standalone chatbot to being further embedded in real-time entertainment scenarios, potentially changing in-game guidance, quest assistance, and the player interaction experience.
  • Discussion summary:Discussions on X are focused on whether the tool can enhance the gaming experience, if it might break immersion or fairness, and to what extent AI assistants should be allowed in multiplayer online games.

Topic 14:Designer Trains AI to Paint Watercolors via Editable Code Link to heading

  • Category:AI · News
  • Overview:Trending since:1 day ago,Related posts:2100
  • What it is:A designer has trained an AI to generate watercolor painting effects using editable code, gaining attention on X.
  • Why it matters:This demonstrates the potential of combining the controllability of code with generative image models, which can enhance the editability, reproducibility, and creator control of AI-generated art.
  • Discussion summary:The discussion centers on whether this method allows artists to better control the AI’s output, whether code-based creation diminishes the value of traditional painting, and the boundaries of copyright and originality for AI-generated watercolor works.

Topic 15:Ben Affleck Boat Meme Roasts Remote Work Habits Link to heading

  • Category:AI · Entertainment
  • Overview:Trending since:6 hours ago,Related posts:1300
  • What it is:A meme of Ben Affleck on a boat has gone viral on X, used to poke fun at slacking off or lazy behavior during remote work.
  • Why it matters:This type of viral meme reflects how generative AI and social media can rapidly amplify entertainment content and influence public perception of remote work culture and digital collaboration methods.
  • Discussion summary:The discussion focuses on whether the meme is a humorous jab at remote work or if it reinforces stereotypes about working from home. Some are also using it to debate the balance between remote work efficiency and workplace freedom.

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

The main AI narrative on X today has clearly shifted from “whose model is stronger” to “who can better integrate AI into enterprises, devices, and real-world scenarios.” Topics like MCP enterprise certification, multi-agent collaboration, on-device AI, and in-game assistants are all centered on AI infrastructure and practical application capabilities. The general consensus is that governance, controllability, cost-effectiveness, and scenario integration have become more important than the sheer parameter race. At the same time, the fact that low-cost models are outperforming expensive flagships on leaderboards reinforces the idea that “cost-performance and real-world effectiveness” are becoming the new standard. The main points of disagreement are twofold: first, whether these announcements represent substantial breakthroughs or are merely feature stacking and ecosystem narratives; and second, whether the security and performance premiums of high-cost models are truly justified, and if leaderboards are an adequate measure of actual capabilities. The potential risks are also clear: inadequate governance in enterprise connectors could lead to data leaks and compliance issues; multi-agent systems could increase costs, and introduce risks of losing control and instability; and if platform algorithms and trending mechanisms continue to amplify controversial and polarizing content, the AI discourse will be further hijacked by noise, bias, and misjudgment.

💡 Influencer Insights Link to heading

No influencer insights for today. We recommend reading the in-depth content on the Watch List.

📚 Appendix: Today’s Watch List Update Sources Link to heading

Time window: Last 3 days; 22 sources covered; 33 updates in total.

Stratechery by Ben Thompson (A_full) Link to heading

  • Autonomy and Innovation
    • Publication Time: 2026-08-24 18:00 Beijing Time
    • Abstract: - Listen to this post.
      • While not every Western followed this trope, by the 1930s, cowboy serials had developed a consistent visual signal: the hero of the story wore a white hat, and the villain wore a black hat.
      • Ultimately, however, they were both cowboys in cowboy hats.
      • Westerns are hardly a cultural reference point today, but the terms “white hat” and “black hat” are very important in the tech world: hackers who focus on fixing vulnerabilities and protecting software are “white hat hackers,” while those who focus on exploiting vulnerabilities for malicious purposes are “black hat hackers.”
      • Of course, this quickly becomes complicated: a government might hire hackers to break into an enemy’s software systems—are they white hats or black hats?
    • EN Key Points:
      • Listen to this post :
      • Log in to listen
      • While not every Western followed the cliché, by the 1930s cowboy serials had landed on a consistent visual cue: the hero of the show wore a white hat, and the v…
      • At the end of the day, however, they both were cowboys with cowboy hats

OpenAI Blog (A_full) Link to heading

  • Advancing price-performance for developers with GPT‑5.6 in Kiro
    • Publication Time: 2026-08-24 20:00 Beijing Time
    • Abstract: The GPT-5.6 model series is now available in Kiro, a software development agent that brings engineering rigor and high quality to large-scale, AI-native coding. For Kiro users, this update brings OpenAI’s latest flagship model series (including Sol, Terra, and Luna) into the development workflow where teams plan, build, review, and test software. Together, these models help developers generate higher-quality code in fewer iterations and achieve better value per token. GPT-5.6 extracts more useful work from each token, offering stronger performance-per-dollar and on-demand capabilities for complex tasks. Within Kiro, developers can apply these capabilities to long-term development work based on their requirements, codebase, and team standards.
    • EN Key Points:
      • GPT‑5.6 is now available in Kiro, helping developers plan, build, review, and test software with better price-performance.

Two Minute Papers (B_intro+search) Link to heading

  • This Small AI Will Change Everything
    • Publication Time: 2026-08-25 00:48 Beijing Time
    • Abstract: ❤️ Check out Lambda and sign up for their GPU cloud: 📝 Qwen3.8-27b is available here: Adam Bridges, B Shang, Carlos Galarza, Christian Ahlin, Eric Tyson, Juan Benet, Lukas Biewald, Michael Tedder, Owen Skarpness, Ryan Stankye, Shawn Becker, Steef, Taras Bobrovytsky, Tazaur Sagenclaw, Tybie Fitzhugh, Ueli Gallizzi. This little AI will change everything.
    • EN Highlights:
      • ❤️ Check out Lambda here and sign up for their GPU Cloud:
      • 📝 The Qwen3.8-27b is available here:
      • 🙏 We would like to thank our generous Patreon supporters who make Two Minute Papers possible:
      • Adam Bridges, B Shang, Carlos Galarza, Christian Ahlin, Eric Tyson, Juan Benet, Lukas Biewald, Michael Tedder, Owen Skarpness, Ryan Stankye, Shawn Becker, Steef…

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

  • SDAD: Spec-Driven Agentic Development for the AI-Native SDLC

    • Published: 2026-08-24 12:00 Beijing Time

    • Abstract: arXiv:2608.20341v1 Announcement Type: New.

      Abstract: Frontier programming agents based on large language models, with context windows spanning hundreds of thousands to millions of tokens, are reshaping the Software Development Life Cycle (SDLC). Rich context handling and multi-step reasoning capabilities enable the complete integration of extensive Functional Requirement Documents (FRDs) and code repository contexts within a single workflow, making specification quality the fuel for autonomous delivery. This report formalizes Spec-Driven Agentic Development (SDAD) as a comprehensive approach that combines rigorous upfront formalization with high-velocity implementation: intent capture, machine-readable specifications, agent synthesis, and independent multi-agent verification under human sign-off.

    • EN Highlights:

      • arXiv:2608.20341v1 Announce Type: new
      • Abstract: Frontier coding agents backed by large language models with context windows from hundreds of thousands to millions of tokens are restructuring the Sof…
      • Rich context handling and multi-step reasoning now allow substantial Functional Requirement Documents (FRDs) and repository context to be ingested in a single w…
      • This report formalises Spec-Driven Agentic Development (SDAD) as a synthesis of disciplined up-front formalisation and high-velocity implementation: intent capt…

  • PrimeAgentOrchestrator: Memory-Primed Agent Spawning for Personal AI Infrastructure

    • Published: 2026-08-24 12:00 Beijing Time
    • Abstract: - arXiv:2608.20342v1 Announcement Type: New.
      • Abstract: Large Language Model (LLM) coding agents start each session with an empty context window, discarding knowledge accumulated in previous work.
      • We introduce PrimeAgentOrchestrator (PAO), a system that spawns new instances of Claude Code—Anthropic’s terminal-based coding agent—pre-loaded with relevant memories compiled from the user’s existing personal database.

  • At spawn time, PAO queries two independently-operated memory backends in parallel (a PostgreSQL entity-observation database and a Cloudflare Worker semantic search index), fuses the results using backend-specific retrieval strategies, and passes the compiled briefing to the host agent via filesystem injection, leveraging its configuration for automatic reading behavior.

    • EN Key Points:
      • arXiv:2608.20342v1 Announce Type: new
      • Abstract: Large language model (LLM) coding agents start each session with an empty context window, discarding accumulated knowledge from prior work
      • We present PrimeAgentOrchestrator (PAO), a system that spawns new instances of Claude Code – Anthropic’s terminal-based coding agent – pre-loaded with relevan…
      • At spawn time, PAO queries two independently-operated memory backends in parallel (a PostgreSQL entity-observation database and a Cloudflare Worker semantic sea…

  • Truth Lies Deep: Countering Semantic Camouflage via Latent Intent Verification

    • Publication Time: 2026-08-24 12:00 Beijing Time

    • Abstract: arXiv:2608.20378v1 Announce Type: new.

      Abstract: Safety alignment in Large Language Models (LLMs) is often superficial, relying on refusal mechanisms that trigger only at the final stages of generation and do not eliminate the foundational knowledge of harmful concepts acquired during pre-training. This study demonstrates that this architectural disconnect leaves models vulnerable to “semantic camouflage” attacks — an adversarial attack that wraps harmful intent in benign narrative contexts (e.g., creative writing), effectively bypassing standard input and output safeguards. By analyzing the latent activation trajectories of three distinct Small Language Model (SLM) families (Phi-3, Qwen2.5, and Gemma-2b) under adversarial stress, this research identifies a universal “intent horizon” — a critical depth (typically 15%–20% of total layers) at which the model’s unique pre-trained representations of harmful intent disintegrate when queries are placed in a “safe” narrative context.

    • EN Key Points:

      • arXiv:2608.20378v1 Announce Type: new
      • Abstract: Safety alignment in Large Language Models (LLMs) is often superficial, relying on refusal mechanisms that trigger only at the final stages of generati…
      • This study demonstrates that this architectural disconnect leaves models vulnerable to Semantic Camouflage – adversarial attacks that wrap harmful intent in be…
      • By analyzing the latent activation trajectories of three distinct Small Language Model (SLM) families (Phi-3, Qwen2.5, and Gemma-2b) under adversarial stress, t…

  • A Survey on Foundations and Frontiers of Multimodal Agentic Frameworks: Techniques and Applications

    • Publication Time: 2026-08-24 12:00 Beijing Time
    • Abstract: - arXiv:2608.20379v1 Announce Type: new paper.
      • Abstract: The progress of Large Language Models (LLMs) has driven a wave of research into agentic capabilities, namely the abilities to reason, plan, and act.

  • This effort has produced agentic frameworks that orchestrate perception, memory, and decision-making around powerful LLM backbones.

  • With the advent of large multimodal models (LMMs), these systems can process and integrate diverse modalities, including images, audio, and video, thereby improving their applicability in the real world.

    • EN Points:
      • arXiv:2608.20379v1 Announce Type: new
      • Abstract: Advances in large language models (LLMs) have fueled a wave of research into agency: the ability to reason, plan, and act
      • This effort has produced agentic frameworks that orchestrate perception, memory, and decision-making around powerful LLM backbones
      • With the advent of large multimodal models (LMMs), these systems can process and integrate diverse modalities, including images, audio, and video, thereby impro…

  • Interpretable Multimodal Classification with Linear Discriminant Tree Ensembles

    • Publication Time: 2026-08-24 12:00 Beijing Time
    • Abstract: arXiv:2608.20384v1 Announce Type: New Paper. Abstract: Multimodal affect and behavior classifiers that fuse heterogeneous text, audio, and visual streams must simultaneously achieve competitive accuracy and generate human-understandable explanations to clarify the cues driving their decisions—a dual objective only partially addressed by current high-capacity models, especially Transformers. While Transformers achieve strong predictive performance, their distributed representations and deep non-linearities make it difficult to assign meaningful weights to individual multimodal features, limiting their use in trust-sensitive applications such as clinical emotion monitoring and educational assessment. We address this gap by developing a framework based on tree-based ensembles that balances accuracy and interpretability.
    • EN Points:
      • arXiv:2608.20384v1 Announce Type: new
      • Abstract: Multimodal affect and behaviour classifiers that fuse heterogeneous text, audio, and visual streams must simultaneously achieve competitive accuracy a…
      • While Transformers attain strong predictive performance, their distributed representations and deep nonlinearity make it difficult to assign meaningful importan…
      • We address this gap by developing a framework based on tree-based ensembles that balances accuracy and interpretability

  • Representation Affects Retrieval: A Case Study of Skill Discovery and Routing in a Multimodal Agent Harness

    • Publication Time: 2026-08-24 12:00 Beijing Time

    • Abstract: arXiv:2608.20389v1 Announcement Type: New.

      Abstract: A production-grade agent framework must discover and rank the most suitable skills from an ever-expanding skill library for a user’s task. At a small scale, this selection occurs in-context: the large language model planner chooses from skill representations exposed in its system prompt, without an explicit embedding-based retrieval step. We view this in-context selection as the small-N counterpart to large-scale embedding-based skill retrieval and present a case study of how a production multimodal video agent framework, Tinycloud, represents skills for its planner.

    • EN Points:

  • arXiv:2608.20389v1 Announce Type: new

  • Abstract: A production agent harness must discover and rank, from a growing library of skills, the one most appropriate for a user’s task

  • At small scale this selection happens in context: the LLM planner chooses among skill representations exposed in its system prompt, without an explicit embeddin…

  • We treat this in-context selection as the small-N counterpart to embedding-based skill retrieval at scale, and present a case study of how Tinycloud, a producti…

  • Nexus: Depth-Adaptive KV-Cache Splicing and Retrieval-Decoupled Tool Routing for Agentic LLMs on Unified Memory

    • Publication Time: 2026-08-24 12:00 Beijing Time
    • Abstract: arXiv:2608.20397v1 Announce Type: new. Abstract: Agentic Large Language Models (LLMs) based on the Model Context Protocol (MCP) re-encode verbose tool schemas in every interaction turn. Consequently, as the tool registry scales, prefill—whose computation is quadratic in sequence length—dominates the Time to First Token (TTFT). Nexus’s primary strategy is to decouple routing from the schema prefill cost: an INT8 semantic lookaside buffer (SLB) with a calibrated cross-encoder margin gate selects tools via retrieval, and parameters are generated based on a compact text signature (median 19 tokens) rather than a concatenated Key-Value (KV) cache. This path is depth-independent: routing accuracy remains near 89% as the registry scales to 250 tools (at which point a baseline that concatenates all schemas completely exceeds the context window limit). Compared to full schema re-prefilling, it generates the first parameter token 1.66x earlier while saving approximately 80% of primary context tokens.
    • EN Key Points:
      • arXiv:2608.20397v1 Announce Type: new
      • Abstract: Agentic large language models (LLMs) on the Model Context Protocol (MCP) re-encode verbose tool schemas every turn, so prefill - quadratic in sequence…
      • Nexus’s primary lever is to decouple routing from the schema-prefill cost: an INT8 semantic lookaside buffer (SLB) with a calibrated cross-encoder margin gate s…
      • This path is depth-independent: routing accuracy stays near 89% as the registry scales to 250 tools - where a concatenate-all-schemas baseline overflows the con…

  • Environmental Slow AI: Design Principles for Generative Systems

    • Publication Time: 2026-08-24 12:00 Beijing Time
    • Abstract: arXiv:2608.20398v1 Announce Type: new. Abstract: Generative AI systems produce cultural products at scale, but their design also reflects the cultural values embedded within them. Once identified, these values can be intentionally reshaped. This position paper examines the maximalist values of current generative AI through the tradition of the environmental humanities and proposes design principles with environmental sustainability as a core value.
    • EN Key Points:
      • arXiv:2608.20398v1 Announce Type: new

  • Abstract: Generative AI (genAI) systems produce cultural artefacts at scale, but they also reflect embedded cultural values through their design

    • Once identified, these values become open to deliberate reshaping
    • This position paper examines the maximalist values of current generative AI through an environmental humanities tradition and proposes design principles in whic…

  • When Retrieval Fails Before It Begins: Structurally Indirect Prerequisite Eviction as a Retention Failure in Agentic Memory

    • Publication Time: 2026-08-24 12:00 Beijing Time
    • Abstract: arXiv:2608.20400v1 Announcement Type: New. Agentic memory under a fixed budget involves two stages: retention and retrieval. Existing retrieval-centered paradigms implicitly assume necessary evidence survives eviction, but we challenge this by isolating a pre-retrieval failure mode: structurally indirect prerequisite eviction, where upstream blocks weakly aligned with the query are discarded under budget pressure. We provide an operational definition of this failure, a reproducible deterministic benchmark, and per-seed trace diagnostics.
    • EN Key Points:
      • arXiv:2608.20400v1 Announce Type: new
      • Abstract: Agentic memory under a fixed budget involves two stages: retention and retrieval
      • Existing retrieval-centered paradigms implicitly assume necessary evidence survives eviction, but we challenge this by isolating a pre-retrieval failure mode: s…
      • We provide an operational definition of this failure, a reproducible deterministic benchmark, and per-seed trace diagnostics

  • World models of environment, agent and joint agent-environment systems

    • Publication Time: 2026-08-24 12:00 Beijing Time
    • Abstract: arXiv:2608.20401v1 Announcement Type: New. World models are a central component of model-based reinforcement learning. They are usually discussed in terms of what variables they predict, such as observations, rewards, states, latent or information states. We argue that there is a prior distinction: which channel they model.
    • EN Key Points:
      • arXiv:2608.20401v1 Announce Type: new
      • Abstract: World models are a central component of model-based reinforcement learning
      • They are usually discussed in terms of what variables they predict, such as observations, rewards, states, latent or information states
      • We argue that there is a prior distinction: which channel they model

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

  • Beyond Raw Transcripts: Structured Persona Extraction for LLM-Based Digital Twins

  • Publication Time: 2026-08-24 12:00 Beijing Time

    • Abstract: arXiv:2608.20344v1 Announcement Type: New Submission. Abstract: “Digital twins” based on large language models aim to simulate an individual’s behavior in new environments or responses to new questions, using representations of their previous answers. Common methods build this representation through summaries of survey records or collected responses. Previous research has shown that compressing long records into shorter summaries generated by large language models does not significantly reduce predictive accuracy, indicating that the amount of information is not the primary bottleneck.
    • EN Highlights:
      • arXiv:2608.20344v1 Announce Type: new
      • Abstract: LLM-based “digital twins” aim to simulate how an individual would behavein new environments or respond to novel questions, given some representation o…
      • A common approach constructs this representation from survey transcripts or summaries responses
      • Prior work shows that compressing long transcripts into shorter LLM-generated summaries does not significantly reduce predictive accuracy, suggesting that infor…

  • When Vocabulary Comprehension Fails Clinical Reasoning: Evaluating Therapy Bots’ Safety Risks for Generation Alpha

    • Publication Time: 2026-08-24 12:00 Beijing Time
    • Abstract: arXiv:2608.20345v1 Announcement Type: New Paper Abstract: Conversational artificial intelligence systems have become an informal mental health support resource for Generation Alpha (Gen Alpha, born 2010-2024). In the United States, 13.1% of adolescents (approximately 5.4 million people) use generative artificial intelligence to obtain mental health advice. Although these systems—from therapy-focused applications to general-purpose chatbots—rely on large language models trained on extensive psychological literature, their safety has not been validated for adolescent communication patterns, which are characterized by exaggerated language, sarcastic positive expressions, rapid semantic drift, and contextual polysemy.
    • EN Highlights:
      • arXiv:2608.20345v1 Announce Type: new
      • Abstract: Conversational AI systems have become informal mental health support resources for Generation Alpha (Gen Alpha, born 2010-2024), with 13.1% of U.S
      • adolescents (5.4 million) using generative AI for mental health advice
      • While these systems, from therapy apps to general chatbots, rely on large language models trained on extensive psychological literature, their safety for youth…

  • Building and Evaluating a Synthetic Bengali Speech Resource for Telecom Customer Care

    • Publication Time: 2026-08-24 12:00 Beijing Time
    • Abstract: arXiv:2608.20346v1 Announcement Type: New Submission. Abstract: Speech systems used in customer-facing applications often require domain-specific language coverage. We present a synthetic Bengali speech dataset for telecom customer service scenarios. The dataset contains 10,000 audio-text pairs, approximately 26.82 hours of 24 kHz speech, and predefined training, validation, and test splits of 9,000, 500, and 500 samples, respectively.
    • EN Highlights:
      • arXiv:2608.20346v1 Announce Type: new

  • Abstract: Speech systems used in customer-facing applications often require domain-specific language coverage

  • We present a synthetic Bengali speech dataset for telecom customer-care scenarios

  • The dataset contains 10,000 audio-text pairs, approximately 26.82 hours of 24 kHz speech, and predefined train, validation, and test splits of 9,000, 500, and 5…

  • Who Do Language Models Think Is Competent? A Mechanistic Analysis of Occupational Bias

    • Published: 2026-08-24 12:00 Beijing Time
    • Abstract: arXiv:2608.20347v1 Announce Type: new. Abstract: Language models (LMs) often pass behavioral bias evaluations, but it remains unclear whether they no longer represent the underlying associations that lead to bias, or have simply learned not to express these biases. In this study, we show that representational biases are often detectable, even when behavioral biases are not visible. We introduce a causal framework that decomposes occupational bias into two measurement points: the model’s internal representation of a user’s competence and its observable output.
    • EN Highlights:
      • arXiv:2608.20347v1 Announce Type: new
      • Abstract: Language models (LMs) often pass behavioral bias evaluations, but it remains unclear whether they no longer represent the underlying associations that…
      • In this study, we show that representational biases are often detectable, even when behavioral biases are not visible
      • We introduce a causal framework that decomposes occupational bias into two measurement points: a model’s internal representation of a user’s competence, and its…

  • Inhibitory Attention for Clinical Long-Context Reasoning: Characterizing and Mitigating Lost-in-the-Middle Effects in EHR Processing

    • Published: 2026-08-24 12:00 Beijing Time
    • Abstract: - arXiv:2608.20348v1 Announce Type: new paper.
      • Abstract: Electronic health records now routinely exceed 100,000 tokens per patient.
      • However, large language models exhibit the “lost-in-the-middle” (LitM) effect: information near the center of a long context is retrieved less reliably than information near the ends.
      • In clinical applications, this issue is not benign: the most critical fact in a patient’s record may be located right in the middle.
    • EN Highlights:
      • arXiv:2608.20348v1 Announce Type: new
      • Abstract: Electronic health records now routinely exceed 100,000 tokens per patient
      • Yet large language models exhibit the lost-in-the-middle (LitM) effect: information near the center of a long context is retrieved less reliably than informatio…

  • In clinical use this is not benign: the single most consequential fact in a note can sit at its center

  • Beyond Prompt Engineering: A Systematic Analysis of Prompt Lexical Sensitivity and Its Impacts on Quality

    • Published: 2026-08-24 12:00 Beijing Time
    • Abstract: arXiv:2608.20349v1 Type: New Release. Abstract: Large language models exhibit extreme sensitivity to superficial prompt changes, where minor lexical alterations can trigger disproportionate performance fluctuations. Moving beyond black-box optimization and coarse-grained templates, we are the first to propose a stability mechanism analysis for prompts based on large-scale, n-gram token levels, using a dataset containing 132,000 prompt variations for our research. Our investigation reveals a fundamental scaling law of prompt performance stability: the higher the average performance of a task, the lower its variance and the stronger its robustness under prompt perturbations.
    • EN Highlights:
      • arXiv:2608.20349v1 Announce Type: new
      • Abstract: Large Language Models (LLMs) exhibit extreme sensitivity to surface-level prompt variations, in which minor lexical changes can trigger disproportiona…
      • Moving beyond black-box optimization and coarse-grained templates, we present the first large-scale, n-gram token-level mechanistic analysis of prompt stability…
      • Our investigation reveals a fundamental Scaling Law of Prompt Performance Stability: higher average task performance is strongly associated with lower variance…

  • How to Train a Real-World Silicon Concierge? Internalizing Complex Business Workflow to Only OneModel

    • Published: 2026-08-24 12:00 Beijing Time
    • Abstract: arXiv:2608.20350v1 Announcement Type: New Release Traditional industrial agents rely on modular pipelines, including components such as routers, retrievers, planners, executors, responders, and reviewers. These systems often fall into a labyrinthine predicament due to the fragmentation of ad-hoc patches, leading to cascading errors and high latency. We propose OneModel, a viable paradigm shift from external workflows to internalized knowledge representation.
    • EN Highlights:
      • arXiv:2608.20350v1 Announce Type: new
      • Abstract: Traditional industrial agents rely on modular pipelines, including Router, Retriever, Planner, Executor, Responder, Reviewer, and other components
      • These systems often fracture into a labyrinth of ad-hoc patches, leading to cascading errors and high latency
      • We propose OneModel, an applicable paradigm shift from external workflows to internalized knowledge representation

  • Exploratory As-Analyzed No-Detection of Culturally-Marked Predicate-Triggered PII Amplification in a Synthetic-English RAG Probe: A Predicate-Resource-Confounded Audit

    • Publication Time: 2026-08-24 12:00 Beijing Time

    • Abstract: arXiv:2608.20351v1 Announce Type: new

      Abstract: We investigate whether stereotype-loaded queries about culturally marked populations leak more personal information from a Retrieval-Augmented Generation (RAG) system than equivalent, neutral queries. We pre-registered a four-culture audit (English-Anglo, Spanish-LATAM, Arabic, Hindi) on a synthetic English Personally Identifiable Information (PII) corpus, comparing five query arms in what we call the Stereotype-Triggered Leakage Differential (STLD). Our locked-in confirmatory estimator was never run, so every test in the paper is either exploratory or a sensitivity analysis, with all plan deviations listed in the appendix.

    • EN Key Points:

      • arXiv:2608.20351v1 Announce Type: new
      • Abstract: We ask whether stereotype-loaded queries about culturally marked people leak more personal information from a retrieval-augmented generation (RAG) sys…
      • We pre-register a four-culture audit (en-Anglo, es-LATAM, Arabic, Hindi) on a synthetic English PII corpus, comparing five query arms we call the Stereotype-Tri…
      • Two caveats up front

  • The Divergence Hypothesis: Unmasking Lexical Interference and Label Bias in Mental Health NLP

    • Publication Time: 2026-08-24 12:00 Beijing Time
    • Abstract: arXiv:2608.20353v1 Announce Type: new. Abstract: Computational mental health (CMH) classifiers often degrade under distribution shift because human annotators and distant-supervision pipelines reward different linguistic signals. We introduce TSS (Triple-Stream Stress probe)—a multi-channel diagnostic framework that decomposes text into: (A) lexical character n-grams; (B) a small, mostly content-free morphosyntactic channel; and (C) a 154-feature psycholinguistic style channel. Across four English datasets (N=12,906), TSS reveals a lexical interference effect: adding lexical features to the style channel reduces Macro-F1 on human-labeled data (average decrease of 0.072, p<10⁻⁴), but has no effect on auto-labeled data.
    • EN Key Points:
      • arXiv:2608.20353v1 Announce Type: new
      • Abstract: Computational mental health (CMH) classifiers often degrade under distribution shift because human annotators and distant-supervision pipelines reward…
      • We introduce TSS (Triple-Stream Stress probe), a multi-channel diagnostic framework that decomposes text into (A) lexical character n-grams, (B) a small, mostly…
      • Across four English datasets (N=12,906), TSS reveals a lexical interference effect: adding lexical features to the style channel reduces Macro-F1 on human-label…

  • ExpertIVS: Sociological Expert Driven Individual Value Simulation in Large Language Models

    • Publication Time: 2026-08-24 12:00 Beijing Time
    • Summary: - arXiv:2608.20355v1 Announcement Type: new.
      • Abstract: Large Language Model (LLM) agents have shown considerable potential in social simulation, but still face difficulties in accurately modeling individual value systems.
      • Most existing methods mechanically stitch survey responses into prompts, which leads to semantic fragmentation and fails to capture the internal coherence of human value systems.
      • The value systems of Large Language Models are typically assessed using static multiple-choice questions, but this method fails to evaluate their value orientation in real-world dialogue interactions.
    • EN Highlights:
      • arXiv:2608.20355v1 Announce Type: new
      • Abstract: Large Language Model (LLM) agents have demonstrated considerable potential for social simulation, yet struggle to accurately model individual value sy…
      • Most existing methods mechanically stitch survey responses into prompts, which suffer from semantic fragmentation, failing to capture the internal coherence of…
      • The value systems of LLMs are typically assessed using static multiple-choice questions, which fail to evaluate the value orientation in real-world dialogue int…

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

  • Bankruptcy Prediction via Hybrid Resampling and Stacking Ensemble Techniques with Explainable Artificial Intelligence (XAI)-Driven Analysis

    • Publication Time: 2026-08-24 12:00 Beijing Time
    • Summary: - arXiv:2608.20343v1 Announcement Type: new paper
      • Abstract: This study develops and evaluates a bankruptcy prediction framework that integrates consensus-based feature selection, hybrid resampling, stacking ensembles, and eXplainable Artificial Intelligence to improve the detection of the minority class in severely imbalanced financial data.
      • Using the Taiwanese Bankruptcy Prediction dataset from the UCI Machine Learning Repository, five feature selection algorithms were first applied, and the input space was reduced to 23 robust variables through a consensus retention rule.
      • Subsequently, balanced training data was generated using SVM-SMOTE, SMOTE-Tomek, and SMOTE-ENN.
    • EN Highlights:
      • arXiv:2608.20343v1 Announce Type: new
      • Abstract: This study develops and evaluates a bankruptcy prediction framework that integrates consensus-based feature selection, hybrid resampling, stacking ens…
      • Using the Taiwanese Bankruptcy Prediction dataset from the UCI Machine Learning Repository, five feature-selection algorithms were first applied, and a consensu…
      • The balanced training data were then generated using SVM-SMOTE, SMOTE-Tomek, and SMOTE-ENN

  • Machine Learning and ARIMA Model Averaging for Adaptive Public Health Forecasting: Comparative Evaluation and an Ontario COVID-19 Case Study

    • Publication Time: 2026-08-24 12:00 Beijing Time
    • Abstract: arXiv:2608.20406v1 Announcement Type: new Abstract: Public health forecasts must be able to respond to abrupt changes in surveillance data while avoiding over-extrapolation of noise, reporting biases, or temporary trends. We evaluated autoregressive integrated moving average (ARIMA), random forest, and extreme gradient boosting (XGBoost) models using 190 weeks of publicly available COVID-19 case data from Ontario between January 2020 and October 2023. Rolling-origin time-series cross-validation preserved the temporal order during model tuning and evaluation.
    • EN Highlights:
      • arXiv:2608.20406v1 Announce Type: new
      • Abstract: Public health forecasts must respond to abrupt changes in surveillance data without over-extrapolating noise, reporting artifacts, or temporary trends
      • We evaluated autoregressive integrated moving average (ARIMA), random forest, and extreme gradient boosting (XGBoost) models using 190 weekly observations of pu…
      • Rolling-origin time-series cross-validation preserved temporal order during model tuning and evaluation

  • From Thermal Preference Prediction to Adaptive Thermal Intervention: A Reinforcement Learning Approach Using Physiological and Environmental Sensing

    • Publication Time: 2026-08-24 12:00 Beijing Time
    • Abstract: arXiv:2608.20423v1 Announcement Type: New submission. Abstract: Personalized thermal comfort is crucial for occupant well-being and for developing more responsive building control strategies. However, traditional heating, ventilation, and air conditioning (HVAC) systems rely on static setpoints and group-level comfort models, failing to capture individual physiological differences. This paper proposes a two-stage personalized thermal comfort approach that integrates multimodal physiological and environmental sensing with reinforcement learning-based decision-making. arXiv:2608.20423v1 Announcement Type: New submission Abstract: Personalized thermal comfort is crucial for occupant well-being and for developing more responsive building control strategies. However, traditional heating, ventilation, and air conditioning (HVAC) systems rely on static setpoints and group-level comfort models, failing to capture individual physiological differences… This paper proposes a two-stage personalized thermal comfort approach that integrates multimodal physiological and environmental sensing with reinforcement learning…
    • EN Highlights:
      • arXiv:2608.20423v1 Announce Type: new
      • Abstract: Personalised thermal comfort is essential for occupant wellbeing and for the development of more responsive building-control strategies, yet conventio…
      • This paper presents a two-stage personalised thermal comfort approach integrating multimodal physiological and environmental sensing with reinforcement learning…

  • BF1: A Causal Dyadic Sparse-Attention Retrofit for Efficient Long-Context Transformers

    • Publication Time: 2026-08-24 12:00 Beijing Time

  • Abstract: arXiv:2608.20427v1 Announcement Type: New Submission Abstract: Even with highly optimized exact kernel implementations, dense causal attention remains expensive in long-context scenarios. We study BF1, a deterministic block-aligned dyadic sparse attention path that combines a small, exact local neighborhood, a global first block, and logarithmically spaced historical blocks. This path is related to prior log-sparse and dilated attention patterns; our contributions include a correctness-gated pretrained model retrofit, a matched topology control study, and a system characterization that links per-layer sparsity to overall model latency.

    • EN Key Points:
      • arXiv:2608.20427v1 Announce Type: new
      • Abstract: Dense causal attention remains expensive at long context even when implemented with highly optimized exact kernels
      • We study BF1, a deterministic block-aligned dyadic sparse-attention route that combines a small exact local neighborhood, a global first block, and logarithmica…
      • The route is related to prior log-sparse and dilated attention patterns; our contribution is a correctness-gated pretrained-model retrofit, a matched topology-c…

  • Approximate Homomorphisms and Convergent Representations in Transducers

    • Publication Time: 2026-08-24 12:00 Beijing Time
    • Abstract: arXiv:2608.20428v1 Announcement Type: New Abstract: We study the stability of minimal representations of controlled stochastic processes (in particular, transducers) under perturbations. This question is motivated by recent experiments finding predictive-state structure in the latent representations of neural networks. We consider standard, linear, and predictive transducers.
    • EN Key Points:
      • arXiv:2608.20428v1 Announce Type: new
      • Abstract: We study the stability of minimal representations of controlled stochastic processes (in particular, transducers) under perturbations
      • This question is motivated by recent experiments finding predictive-state structure in the latent representations of neural networks
      • We consider standard, linear and predictive transducers

  • Wrong-Physics Backdoors in Neural PDE Operators

    • Publication Time: 2026-08-24 12:00 Beijing Time
    • Abstract: arXiv:2608.20439v1 Announcement Type: New Submission. Abstract: Neural PDE operators are increasingly trained on reusable solver archives, yet validation often relies on clean prediction errors and parameter-agnostic sanity checks. We introduce cross-parameter relinking, a data-poisoning primitive that causes a triggered input to select a valid solution from the same PDE family but under incorrect physical parameters. We call this a wrong-physics backdoor: the output is physically plausible but incorrect for the intended parameters.
    • EN Key Points:
      • arXiv:2608.20439v1 Announce Type: new
      • Abstract: Neural PDE operators are increasingly trained on reusable solver archives, yet validation often relies on clean prediction error and parameter-agnosti…

  • We introduce cross-parameter relinking, a data-poisoning primitive that makes a triggered input select a valid solution from the same PDE family under an incorr…

  • We term this a wrong-physics backdoor: the output remains physically plausible but is wrong for the intended parameter

  • Decision Tree and K-Means Analysis of Raman Spectra for Edible Oils: A Physics-Informed AI Approach

    • Published: 2026-08-24 12:00 Beijing Time
    • Abstract: arXiv:2608.20440v1 Announce Type: new. Abstract: The identification of edible oils in processed foods is crucial for food quality, fraud prevention, and regulatory compliance. This study establishes an integrated Raman spectroscopy and machine learning framework that combines intrinsic spectral organization, interpretable classification, and physics-informed AI (PI-AI). Using t-SNE, K-means clustering, decision trees, and spectral decomposition based on non-negative least squares (NNLS), five types of pure edible oils and their forms in a fried potato chip matrix were studied.
    • EN Key Points:
      • arXiv:2608.20440v1 Announce Type: new
      • Abstract: Authentication of edible oils in processed foods is important for food quality, fraud prevention, and regulatory compliance
      • This study establishes an integrated Raman spectroscopy and machine-learning framework that links intrinsic spectral organization, interpretable classification,…
      • Five edible oils were investigated in pure form and within a fried-potato-chip matrix using t-SNE, K-means clustering, Decision Trees, and Non-Negative Least Sq…

  • Shared Physics Responses Recover Hidden Rankings in Neural Operator Libraries

    • Published: 2026-08-24 12:00 Beijing Time

    • Abstract: arXiv:2608.20441v1 Announce Type: new submission

      Abstract: Selecting the optimal neural operator prediction during deployment is challenging when high-fidelity reference solutions are not available. We demonstrate that under a squared Hilbert space loss, the ranking of a finite model library strictly depends on the low-dimensional span of candidate differences, allowing us to use an anchor-based linearized response of the governing equations to score all models simultaneously. This shared physics diagnostic accurately recovers over 99.6% of pairwise preferences and 99.0% of optimal checkpoints across various Fourier and convolutional operator libraries in fluid, reaction-diffusion, and wave dynamics.

    • EN Key Points:

      • arXiv:2608.20441v1 Announce Type: new
      • Abstract: Selecting the optimal neural-operator prediction during deployment is challenging when high-fidelity reference solutions are unavailable
      • We demonstrate that under a squared Hilbert-space loss, ranking a finite model library depends strictly on the low-dimensional span of candidate differences, al…

  • This shared physical diagnostic accurately recovered over 99.6% of pairwise preferences and 99.0% of optimal checkpoints across diverse Fourier and convolutio…

  • Stored in Optimizer State, Valued by Later Training: A Causal Account of Subliminal Trait Transfer

    • Publication Time: 2026-08-24 12:00 Beijing Time
    • Abstract: arXiv:2608.20442v1 Announcement Type: New Paper. Abstract: Subliminal trait transfer allows a student model to acquire behavioral dispositions from teacher-generated data in which the trait is not semantically expressed. Recent work explains how such signals enter gradients, but not how they survive source removal or acquire different signs under later training. We treat parameters and optimizer moments as a single trainer state and derive an exact transport-valuation identity separating observer-independent propagation of source perturbations from the valuation of future continuations and behavioral readouts.
    • EN Highlights:
      • arXiv:2608.20442v1 Announce Type: new
      • Abstract: Subliminal trait transfer allows a student model to acquire behavioral dispositions from teacher-generated data in which the trait is not semantically…
      • Recent work explains how such signals enter gradients, but not how they survive source removal or acquire different signs under later training
      • We treat parameters and optimizer moments as a single trainer state and derive an exact transport-valuation identity separating observer-independent propagation…

  • Amortized Bandwidth Learning for Kernel Density Estimation under Logarithmic Score

    • Publication Time: 2026-08-24 12:00 Beijing Time
    • Abstract: arXiv:2608.20445v1 Announcement Type: New Submission. Abstract: Kernel density estimation converts finite samples into probability densities, but its performance depends critically on bandwidth selection. Classical selectors prescribe the sample-to-bandwidth rule analytically or asymptotically, or solve a new optimization for each sample. An amortized framework is proposed that instead learns this mapping across a distribution of density-estimation tasks by optimizing the logarithmic score.
    • EN Highlights:
      • arXiv:2608.20445v1 Announce Type: new
      • Abstract: Kernel density estimation converts finite samples into probability densities, but its performance depends critically on bandwidth selection
      • Classical selectors prescribe the sample-to-bandwidth rule analytically or asymptotically, or solve a new optimization for each sample
      • An amortized framework is proposed that instead learns this mapping across a distribution of density-estimation tasks by optimizing the logarithmic score