System translated (Gemini)

🤖 AI 速览

Today’s focus shifts from model parameters to two ends: on one side is Stripe’s acquisition of OpenRouter, with AI routing, billing, and access continuing to move upstream; on the other side, OpenAI incorporates national security oversight and youth AI education into its product …
📋 文章元数据
发布时间
2026-08-19
类型
ai-daily
字数
7870
阅读时长
37 min

2026-08-19 AI Daily | AI Infrastructure and Governance Heat Up: Stripe Acquires OpenRouter, OpenAI Launches Oversight and Education Collaborations Link to heading

Today’s focus shifts from model parameters to two key areas: on one hand, Stripe’s acquisition of OpenRouter is moving AI routing, billing, and gateways further up the stack; on the other, OpenAI is expanding its product scope to include national security oversight and AI education for youth. Meanwhile, the engineering community is concentrating on efficiency improvements, evaluation, and cost control.

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

There are three main themes worth a deep dive today. First, OpenAI continues to update its safety and governance agenda: from democratic oversight in national security and the pace of model development for critical cyber capabilities, to educational collaborations with ChatGPT and CodeAI for young people. The main thread is clear: as AI capabilities expand, institutions, education, and product safeguards must be upgraded in tandem.

The second theme is engineering implementation. Asana’s use of Codex to clear five years of technical debt in two weeks is a case worth close examination by engineering teams today. Meanwhile, a paper on whether Language Server Protocols (LSP) can save tokens for coding agents brings the cost issue of “AI agents writing code” to a measurable level.

The third theme is research into efficiency and reliability: token-inflation routing, long-context BCMT, prompt compression, latent reasoning interpretability, and multilingual GRPO all address the same question. The next phase of model competition is not just about being more powerful, but also more economical, controllable, and verifiable.

🌐 AI Hot Topics on X Link to heading

Topic 1: Stripe Acquires OpenRouter for Over $7 Billion in Major AI Deal Link to heading

  • Category: AI · News
  • Overview: Trending for: 2 days ago, Related posts: 16,000
  • What happened: Stripe has reportedly acquired OpenRouter, an AI model routing and gateway platform, for over $7 billion, making it a high-profile AI infrastructure deal this year.
  • Why it matters: This deal reflects that the AI ecosystem competition is shifting from models themselves to the layers of API calls, routing, billing, and infrastructure. Stripe is also deepening its strategic position in AI payments and developer access.
  • Discussion summary: The discussion on X is mainly focused on three points: whether the $7 billion valuation is too high, the value of OpenRouter as a middle layer connecting over 400 models, and whether this signals Stripe’s transition from a payment company to a more comprehensive AI infrastructure platform.

Topic 2: Tech Teams Build Custom Harnesses to Scale Claude AI Agents Link to heading

  • Category: AI · News
  • Overview: Trending for: 22 hours ago, Related posts: 409
  • What happened: Tech teams are building custom runtime frameworks and testing/scheduling tools for AI agents like Claude to support larger-scale automated task execution.
  • Why it matters: This indicates that the focus of AI agent implementation is shifting from model capabilities to engineering infrastructure, including data pipelines, evaluation, monitoring, and cost control. This could affect the scaling speed of enterprise-level AI applications.
  • Discussion summary: The discussion on X centers on whether custom harnesses will be key to scaling AI agents and whether a large-scale “AI data factory” model can deliver reliable output. Opinions are divided: some are optimistic about its potential to boost automation efficiency, while others worry about complexity, cost, and controllability issues.

Topic 3: Anthropic CEO Predicts AI Will Cure Most Diseases in 5-10 Years Link to heading

  • Category: AI · Other
  • Overview: Trending for: 2 days ago, Related posts: 51,000
  • What happened: The CEO of Anthropic predicts that AI could help cure most diseases within the next 5 to 10 years.
  • Why it matters: This prediction reflects high expectations for AI in biomedicine, drug discovery, and clinical research. It also raises the question of whether AI can truly accelerate scientific translation and impact the healthcare industry’s landscape.
  • Discussion summary: The discussion on X is split between optimists and skeptics. Supporters highlight AI’s progress in protein structure prediction, drug screening, and medical data analysis. Skeptics argue that disease mechanisms are complex, clinical trials are lengthy, and regulatory, safety, and liability issues cannot be quickly solved by AI.

Summary of AI Public Opinion on X Today Link to heading

Today’s public discourse clearly points to a consensus: the AI competition is shifting from “who has the stronger model” to “who controls the more critical infrastructure, orchestration capabilities, and scenario entry points.” This is reflected in both Stripe’s acquisition of OpenRouter and the custom frameworks being built around Claude Agents, highlighting the industry’s focus on invocation, routing, evaluation, and cost control. The main points of divergence are centered on valuation and the speed of implementation. Some believe that middleware layers like OpenRouter and Agent harnesses will be key to scaling, while others worry that these platforms are overvalued, too complex from an engineering perspective, and will ultimately struggle to generate stable returns. Anthropic CEO’s optimistic prediction of “curing most diseases within 5 to 10 years” elevates this disagreement to a higher dimension. Supporters believe AI will significantly accelerate the translation of scientific research into practical applications. In contrast, skeptics argue that the hurdles of clinical validation, regulation, and the chain of responsibility in biomedicine are far more challenging to overcome than model capabilities. The potential risk is that capital and expectations may outpace actual implementation. If investments in infrastructure, Agent automation, and the healthcare narrative are all priced with excessive optimism, it could easily lead to subsequent valuation corrections, shortfalls in technological delivery, and disputes over safety and compliance.

💡 Influencer Insights Link to heading

AI Influencer Watch: Daily Briefing Link to heading

1. Agent-based Operating Systems and Coding Platforms Become the Main Battlefield

  • Explosion of the DeepSeek Harness (DSH) Ecosystem: @vista8 and @dotey both extensively discussed DSH’s plugin ecosystem, GUI clients, and the anime-style avatar developer community, stating, “DSH was just open-sourced a few days ago, and its plugin ecosystem is already so prosperous; the future is promising.” @Pluvio9yte even shared a community-packaged, out-of-the-box desktop client.
  • Doubao Becomes Fully Agent-based: @dotey noted that Doubao is “becoming more and more like Codex.” Its Windows version achieves GUI operations through a virtual desktop, and the mobile app can remotely control the desktop Agent. @vista8 also commented that the experience “is very similar to Codex” and “somewhat exceeded expectations.”
  • Cursor’s Code Hosting Platform, Origin, Goes Live: @dotey explained that Origin is designed for AI Agents as primary users, supporting 22.6 commits per second and featuring built-in AI to automatically resolve merge conflicts, creating a closed loop of “editor -> cloud intelligence -> code review -> code hosting.”
  • Omarchy (Agent-first Linux system) Gains Popularity: @vista8 jumped on the bandwagon and installed it, quoting DHH who called it his most satisfying work.

2. On-device Models and Local Execution Continue to Heat Up

  • @zhixianio tested the DeepSeek V4 Flash 4-bit quantized version on a Mac Studio, comparing it with Gemma 4 12B Coder and Qwen 3.6-35B-A3B MoE. The conclusion was that a 12B model size is insufficient for complex programming tasks that require “long, stateful, single-pass generation,” with 35B remaining the sweet spot.
  • @zhixianio experienced the full-duplex audio and video interaction of MiniCPM-o 4.5, remarking, “It’s hard to imagine that a 9B model can achieve this level of performance.”
  • @zixianio announced the development of a “simp module” (a proactive memory system) and a workload scheduler for multiple on-device models on consumer-grade hardware.
  • @ruanyf compared local AI hardware: the RTX 5090 versus AMD’s Strix Halo onboard solution, pointing out that “often, the onboard chipset is the better solution for local AI.”

3. AI Video Generation is Becoming More Practical and Refined

  • @Pluvio9yte argued that video generation can’t rely solely on prompt lottery. Instead, one should iteratively feed existing clips into “video-to-video” models, modifying segments repeatedly. They also showed a side-by-side comparison of MiniMax H3 and Seedance 2.5 with the same prompt, noting the latter’s superior micro-expression control.
  • @vist8 shared a dinosaur educational video quickly generated using AI video tools.
  • @Pluvio9yte discovered the counter-intuitive result that increasing sampling steps from 4 to 8 actually caused facial expressions to become more symmetrical.

4. Multi-Agent Collaboration and New Workflow Platforms

  • @dotey introduced Cumora (open-sourced by yetone), which turns AI Agents into official members of a chat group. They have personas, can speak proactively, support both cloud and local BYOA (Bring Your Own Agent) modes, and have a built-in coordination mechanism to prevent conflicts.
  • @dotey shared complaints about Claude Code’s feature that automatically sends messages across sessions, pointing out that it’s enabled by default in the latest version and is difficult to disable.

5. Large Model Memory, Context Accumulation, and Data Flywheels

  • Several points indicate that memory and context are becoming a product’s competitive moat: @zhixianio’s proactive “simp module” for memory; @vista8 paying for Obsidian sync and emphasizing that “in the AI era, you need to take context accumulation seriously”; and @dotey mentioning that products like ZCode could create a flywheel of “product usage -> data -> model upgrade.”
  • @ruanyf wrote an article explaining that for large models, the input price for a cache hit is 1/50th of a cache miss, urging users to fully leverage caching to reduce costs.

II. Noteworthy Unique Perspectives and Industry Foresight Link to heading

  • Powerful Rebuttal to the “Small Model + Tools” Approach ( @dotey citing @_jasonwei)：Jason Wei believes that relying solely on a 1B cognitive core plus tool retrieval cannot replace the internalized knowledge of large models; speed, depth of understanding, and reliability are all insufficient. Pursuing the highest quality always requires larger models, which aligns with the “bitter lesson.”
  • Code is Truth, Bash is Enough ( @dotey citing Pi authors)：The two authors of Pi believe that code does not require memory systems/RAG; Bash can be combined arbitrarily, and most scenarios do not require MCP; skill + scripts are sufficient.
  • Meituan Reflects on “All-Staff Shrimp Farming” ( @dotey quoting)：Meituan executives stated that the company-wide use of imported AI tools in February-March led to daily bills of tens of millions and generated errors that interfered with operations.
  • AI Programming Might Be More Expensive Than Human Programmers ( @ruanyf)：OpenClaw founder’s monthly token consumption is estimated at $1.3 million, and the cost for enterprises using flagship models without limits is astonishing.
  • Xiaohongshu Becomes a Skill Publishing Platform ( @ruanyf)：Xiaohongshu launched the new RedSkill feature, where notes can include Skill files for one-click copy and installation, aiming to be the “GitHub for Skills.”
  • Anthropic: AI Open Source is a False Proposition ( @ruanyf)：The founder stated that only making weights public doesn’t reveal internal operations, making it impossible to participate in development, and thus it shouldn’t be called open source.
  • Free Costs and Moral Responsibilities in the AI Era ( @ruanyf)：The SQLite author rejected external PRs, comparing it to “taking in a free puppy,” with maintenance responsibility lasting for twenty-five years.
  • Should We Take Holidays After AI Improves Efficiency? ( @ruanyf)：The article suggests that since AI can complete a week’s work in a few hours, taking a day or two off is logical.
  • Claude’s Hesitant Approach to Quota Increases ( @dotey)：Anthropic repeatedly postponed and reversed quota increases, drawing criticism for being “not straightforward,” preferring to maintain a 50% increase indefinitely.
  • Deep Seak Harness Plugin Ecosystem ( @vista8)：Recommends modlens (image recognition), dsh-at-file (@cite file), dsh-paste-input (paste file), and dsh-cc-tui (terminal-style interface). Aggregation sites and GUI clients have both emerged.
  • Cumora ( @yetone/@dotey)：Open-source multi-agent collaborative chat room, website cumora.com, open-sourced on GitHub, can be deployed locally or used in the cloud.
  • Omarchy OS ( @vista8)：An agent-first Linux distribution, created by DHH, images available for official download.
  • Doubao PC Task Management ( @dotey/@vista8)：Supports GUI operations, remote phone connection to PC, domestic users can directly experience advanced Agent features.
  • OpenConnector ( @ruanyf)：Open-source password connection gateway to prevent AI Agents from leaking credentials, supports Cloudflare Workers deployment.
  • fireworks-tech-gaph skill ( @vista8)：A skill for generating technical images, has gained tens of thousands of stars, supports 12 styles and SVG/PNG/GIF.
  • Niulai.skill ( @Pluvio9yte)：An open-source Skill for Xiaohongshu/Douyin account creation and abstract content, with material sourced from Dragon Restaurant.
  • Kimi K3 as a Powerful Large Model Without Magic ( @vista8)：Configures networks and Agents through it to complete tasks.
  • Method to Enable 1M Context in Codex ( @dotey)：Simply modify model_context_window and model_auto_compact_token_limit in config.toml.
  • Mail Agent and Codex for Scheduled Email Summarization ( @Pluvio9yte)：Uses GPT-5.4/5.6-luna to summarize emails every night, efficiently handling multiple business mailboxes.
  • Starryblu Singapore Bank Card ( @AI_Jasonyu)：Some bloggers are promoting it during a limited-time, fee-free card opening period.
  • Giffgaff to Lebara Porting Tutorial for Number Retention ( @AI_Jasonyu)：A practical self-help guide for account blocking and roaming service suspension.
  • CapWords ( @nishuang)：An AI foreign language learning app with fun live-action animations, suitable for Xiaohongshu users and loyal vocabulary memorizers.

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

Time Window: Last 3 days; Covering 22 sources; Total 37 updates

All-In Podcast (A_full) Link to heading

  • Flock CEO Garrett Langley on Controversy, “Surveillance State” Claims, and Privacy vs Safety
  • Publication Time: 2026-08-18 08:47 Beijing Time
    • Summary: - AppLovin Ads is AppLovin’s AI advertising platform, reaching over 1 billion daily active users in the mobile gaming sector.
      • Full-screen video ads with a median viewing time of 35 seconds.
      • Advertisers spend hundreds of thousands of dollars daily to generate profits.
      • Over 3,000 businesses worldwide use it for sales tax, VAT, and GST.
      • They handle registration, filing, and tax rates, so you can stay ahead of risks.
    • EN Key Points:
      • (0:00) The most controversial company in privacy right now, Flock CEO joins the show
      • (7:23) License plate data retention: 7 days solves 90% of crimes
      • (13:00) Camera vandalism, felony charges, and privacy concerns
      • (18:15) Dirty cops exposed: Flock’s audit tool got 9 Georgia officers fired

Stratechery by Ben Thompson (A_full) Link to heading

  • Nvidia Backs OpenAI Data Center, Anthropic News, Google Buys Spirit Airlines Data
    • Publication Time: 2026-08-18 18:00 Beijing Time
    • Summary: - Nvidia has struck another deal, this time with a frontier lab; Anthropic’s revenue continues to amaze; and perhaps data is finally oil.
      • $15/month* or *$150/year.
      • Substantive analysis of the day’s news via three weekly emails or podcasts.
      • Strategy Interviews.
      • Interviews with leading public company CEOs, private company founders, and discussions with fellow analysts.
    • EN Key Points:
      • Nvidia makes another deal, this time with a frontier lab; Anthropic’s revenue continues to amaze; and maybe data finally is oil.

OpenAI Blog (A_full) Link to heading

  • Strengthening democratic oversight in national security

    • Publication Time: 2026-08-19 03:00 Beijing Time
    • Summary: - Artificial intelligence is transforming how democratic governments protect their people.
      • It can help stop cyberattacks, protect critical infrastructure, detect threats earlier, and give public servants a clearer picture in a crisis.
      • Used properly, these tools can enhance national security.
      • Democratic oversight helps ensure that public power is used accountably to the people it serves.
      • As AI makes national security work faster and more capable, the institutions responsible for overseeing that work need to keep pace.
    • EN Key Points:
      • OpenAI launches an initiative to strengthen democratic oversight of AI in national security, supporting government institutions with tools, training, and expert…

  • Partnering with CodeAI to prepare the first AI generation

    • Publication Time: 2026-08-18 19:00 Beijing Time
    • Summary: - Today’s students will be the first generation to grow up with AI.
      • For parents and educators, the question is not just whether young people will use AI, but whether they will learn to critically evaluate its outputs, understand its limitations, and use it responsibly.
      • Currently, there is a gap between use and understanding.
      • Young people need to understand how AI works, think critically about its outputs, and develop the skills to shape the future by using tools that prioritize their safety and development.
      • Therefore, OpenAI and CodeAI are announcing a landmark partnership to provide students and educators with the tools and resources to learn how to use and benefit from AI.
    • EN Key Points:

  • OpenAI and CodeAI are partnering to help students build AI literacy, think critically about AI, and develop the skills to use and shape it responsibly.

  • Pacing model development in an era of cyber-critical capabilities

    • Publication Time: 2026-08-18 19:00 Beijing Time
    • Abstract: - OpenAI is strengthening the monitoring, coordination, and security of frontier AI models.
      • Learn how new safeguards are guiding the pace of model development.
      • OpenAI is enhancing the monitoring, alignment, and security of frontier AI models. Learn how new safeguards are guiding the pace of model development.
      • The pace of model development in an era of cyber-critical capabilities.
    • EN Highlights:
      • OpenAI is strengthening monitoring, alignment, and security for frontier AI models
      • See how new safeguards are guiding the pace of model development.

  • Introducing ChatGPT for Teens: Built for learning, backed by protections

    • Publication Time: 2026-08-18 19:00 Beijing Time
    • Abstract: - ChatGPT for Teens helps teenagers learn, think critically, and use AI with confidence, offering stronger built-in protections, healthy-use features, and additional parental controls.
      • ChatGPT for Teens helps teenagers learn, think critically, and use AI with stronger built-in protections, healthy-use features, and additional content…
      • Introducing ChatGPT for Teens: Designed for learning and backed by protections.
    • EN Highlights:
      • ChatGPT for Teens helps teens learn, think critically, and use AI with confidence, with stronger built-in protections, healthy-use features, and additional cont…

  • Asana cleared 5 years of engineering work in 2 weeks with Codex

    • Publication Time: 2026-08-18 15:00 Beijing Time
    • Abstract: - Asana used OpenAI Codex to replace an outdated testing system in two weeks, completing work expected to take five years for about $12,000.
      • This article from the OpenAI blog explains how Asana completed 5 years of engineering work in 2 weeks and utilized Codex to shape the broader AI and infrastructure landscape.
      • Asana completed 5 years of engineering work in 2 weeks with Codex, which also brings practical implications for founders, operators, and investors.
    • EN Highlights:
      • Asana used OpenAI Codex to replace an outdated testing system in two weeks, completing work expected to take five years for about $12K.

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

  • FLOPs vs Real Work: The Importance of Replication in AI Efficiency Assessment

    • Publication Time: 2026-08-18 12:00 Beijing Time
    • Abstract: - arXiv:2608.14550v1 Announcement Type: New.
      • Abstract: Due to the large scale of models, high energy demands, and environmental costs, AI efficiency has recently garnered attention from both academia and industry.
      • While reporting floating-point operations (FLOPs) is the traditional method for evaluating computational cost, the relationship between FLOPs and execution time is not straightforward. Layers with the same number of FLOPs may not have the same execution time, as some operations are more easily parallelized than others.

  • This paper sets out to replicate the original experiments from a study that proposed the $\alpha-FLOPs$ estimation formula to verify whether the results still apply to newer, more powerful hardware.

    • EN Highlights:
      • arXiv:2608.14550v1 Announce Type: new
      • Abstract: AI efficiency has recently taken the spotlight in both academy and industry due to massive model scales, high energy demands, and environmental costs
      • While reporting Floating Point Operations (FLOPs) is a traditional approach for assessing computational costs, the relationship between FLOPs and execution time…
      • This paper sets out to replicate the original experiments from a study that proposed the $\alpha-FLOPs$ estimation formula to verify whether the results remain…

  • Large Language Models Show Metacognitive Sensitivity in Medical Reasoning

    • Publication Time: 2026-08-18 12:00 Beijing Time
    • Abstract: - arXiv:2608.14552v1 Announce Type: new.
      • Abstract: Large language models (LLMs) are increasingly evaluated and used in medicine, but clinical usefulness depends on answer accuracy and whether confidence tracks evidence quality and uncertainty.
      • We developed a controlled, psychophysics-inspired clinical benchmark to test diagnostic choice and confidence behavior in medical LLMs.
      • The benchmark focused on probable Alzheimer-type neurocognitive disorder (AT-NCD) versus depression-related cognitive impairment (DRCI).
    • EN Highlights:
      • arXiv:2608.14552v1 Announce Type: new
      • Abstract: Large language models (LLMs) are increasingly evaluated and used in medicine, but clinical usefulness depends on answer accuracy and whether confidenc…
      • We developed a controlled, psychophysics-inspired clinical benchmark to test diagnostic choice and confidence behavior in a medical LLM
      • The benchmark focused on probable Alzheimer-type neurocognitive disorder (AT-NCD) versus depression-related cognitive impairment (DRCI)

  • The Unwritten Benchmark: A New Challenge for Multimodal Machine Learning in Abstract Perceptual Reasoning

    • Publication Time: 2026-08-18 12:00 Beijing Time
    • Abstract: - arXiv:2608.14558v1 Announce Type: new.
      • Abstract: Current multimodal models have demonstrated exceptional capabilities in recognizing static visual and auditory content.
      • However, their abstract perceptual reasoning ability—the capacity to infer unseen information from dynamic generation processes—remains a critical and underexplored frontier.
      • In this paper, we introduce the Unwritten Benchmark, a new challenge designed to explore this abstract perceptual and cognitive capability.
    • EN Highlights:
      • arXiv:2608.14558v1 Announce Type: new

  • Abstract: Current multimodal models have demonstrated remarkable proficiency in recognizing static visual and auditory content

  • However, their capacity for abstract perceptual reasoning, inferring unseen information from dynamic, generative processes, remains a critical and underexplored…

  • In this paper, we introduce The Unwritten Benchmark, a new challenge designed to probe this abstract perceptual and cognitive ability

  • When to Communicate: Belief Distributions and KL Divergence for Principled Gating in Multi-Agent RL

    • Publish Time: 2026-08-18 12:00 Beijing Time
    • Abstract: - arXiv:2608.14559v1 Announce Type: new.
      • Abstract: Effective communication in multi-agent reinforcement learning requires agents to decide not only \textit{what} to communicate, but when to communicate?
      • Existing approaches either communicate at every timestep or learn a binary gate through REINFORCE policy gradients \cite{singh2019}, a high-variance signal that yields unstable and uninterpretable gating behaviors.
      • I propose a principled alternative: agents communicate only when the KL divergence between their learned belief distributions exceeds a fixed threshold.
    • EN 要点:
      • arXiv:2608.14559v1 Announce Type: new
      • Abstract: Effective communication in multi-agent reinforcement learning requires agents to decide not only \textit{what} to communicate, but when
      • Existing approaches either communicate at every timestep or learn a binary gate through REINFORCE policy gradients \cite{singh2019}, a high-variance signal that…
      • I propose a principled alternative: agents communicate only when the KL divergence between their learned belief distributions exceeds a fixed threshold

  • Global AI Regulations for FAIR and Ethics in High-Risk Use Cases: A Comparative Review

    • Publish Time: 2026-08-18 12:00 Beijing Time
    • Abstract: - arXiv:2608.14562v1 Announce Type: new.
      • Abstract: AI governance is shifting from voluntary ethics to enforceable, risk-based regulation, but cross-jurisdictional divergence creates compliance uncertainty for high-risk AI operators.
      • We present a comparative matrix for the EU, US, and China, mapping (i) risk classification triggers, (ii) binding obligations, (iii) enforcement and accountability mechanisms, and (iv) the extent to which FAIR principles are implemented in practice.
      • We stress-test the matrix across three high-impact domains: EEG-guided rehabilitation robotics, AI-enabled debt collection in future Central Bank Digital Currency (CBDC) ecosystems, and AI-driven allocation of scarce Graphics Processing Unit (GPU) resources in emerging AI factory infrastructure.
    • EN 要点:
      • arXiv:2608.14562v1 Announce Type: new

  • Abstract: AI governance is shifting from voluntary ethics to enforceable, risk-based regulation, yet cross-jurisdictional divergence creates compliance uncertai…

  • We present a comparative matrix for the EU, US, and China that maps (i) risk classification triggers, (ii) binding obligations, (iii) enforcement and accountabi…

  • We stress-test the matrix on three high-impact domains: Electroencephalography (EEG)-guided rehabilitation robotics, AI-enabled debt collection in prospective C…

  • Position: AI Lock-In Is in Progress, and We Must Be Prepared

    • Published: 2026-08-18 12:00 Beijing Time
    • Abstract: - arXiv:2608.14565v1 Announce Type: new.
      • Abstract: AI safety research has mainly focused on two areas: technical alignment (ensuring AI systems produce human-aligned outputs) and the regulation of generative AI’s societal impacts (including risks of unemployment and labor market disruption).
      • However, an equally important dimension remains underexplored: the risk inherent in dependence on AI systems themselves.
      • In this position paper, we argue that AI safety research should address AI Lock-In, the phenomenon whereby excessive reliance on AI systems leads to human deskilling, erodes the capacity for independent human operation, and creates systemic vulnerabilities when AI systems are unavailable or compromised.
    • EN Highlights:
      • arXiv:2608.14565v1 Announce Type: new
      • Abstract: AI safety research has mainly focused on two areas: technical alignment (ensuring AI systems produce human-aligned outputs) and the regulation of gene…
      • However, an equally important dimension remains underexplored: the risk inherent in dependence on AI systems themselves
      • In this position paper, we argue that AI safety research should address AI Lock-In, the phenomenon whereby excessive reliance on AI systems leads to human deski…

  • Position: Evaluations of AI Moral Reasoning Still Miss Half of the Picture

    • Published: 2026-08-18 12:00 Beijing Time
    • Abstract: - arXiv:2608.14566v1 Announce Type: new.
      • Abstract: Recent work on evaluating the moral competence of large language models (LLMs) has focused primarily on what we call the moral value problem, i.e., whether model outputs align with human moral values.
      • In contrast, the moral norm problem—whether models can identify and correctly apply context-relevant moral norms—remains underexplored.
      • We argue this imbalance stems from the field’s reliance on descriptive ethics frameworks, such as Moral Foundations Theory and Kohlberg’s stages of moral development, which emphasize value representation over norm application.
    • EN Highlights:
      • arXiv:2608.14566v1 Announce Type: new
      • Abstract: Recent work on evaluating the moral competence of large language models (LLMs) has focused primarily on what we call the moral value problem, i.e., wh…

  • In contrast, the moral norm problem, i.e., whether models can identify and correctly apply context-sensitive moral norms, remains underexplored

  • We posit that this imbalance stems from the field’s reliance on descriptive ethics frameworks, such as Moral Foundations Theory and Kohlberg’s stages of moral d…

  • From Doyle to AGM: A Survey and an Implementation Roadmap for Belief Change

    • Publication Time: 2026-08-18 12:00 Beijing Time
    • Abstract: - arXiv:2608.14567v1 Announce Type: new.
      • Abstract: This paper presents a targeted narrative review establishing the historical and theoretical foundations for computational belief change implementation
      • Seeded by Doyle and London’s foundational 1980 taxonomy, we trace the evolution of belief revision from computational origins through the theoretical transforma…
      • Our analysis demonstrates how pre-AGM computational pragmatism relates to AGM theoretical constructs, revealing both continuities and transformations across thi…

  • Position: AI Governance Needs ISO-like Interoperability Protocols, Not Just Laws

    • Publication Time: 2026-08-18 12:00 Beijing Time
    • Abstract: - arXiv:2608.14568v1 Announce Type: new.
      • Abstract: As Artificial Intelligence (AI) systems become deeply integrated into critical global infrastructure, the urgency to establish robust governance frameworks has grown significantly.
      • However, current approaches, dominated by jurisdiction-specific laws, policies, and voluntary frameworks (e.g., the EU AI Act, China’s algorithmic governance, and the U.S. NIST AI Risk Management Framework), create a fragmented regulatory landscape.
      • In this position paper, we argue that \textbf{\textit{AI governance must be built not just on laws, but on ISO-like interoperability protocols to enable standardized, machine-readable, cross-border risk communication}}.

  • However, current approaches, led by jurisdiction-specific laws, policies, and voluntary frameworks such as the EU AI Act, China’s algorithm governance, and the…

  • In this position paper, we argue that \textbf{\textit{AI governance must be built not on laws alone, but on ISO-like interoperability protocols that enable stan…

  • Position: Certified Correctness in Neural Constraint Reasoning Requires Symbolic Integration

    • Posted: 2026-08-18 12:00 Beijing Time
    • Abstract: - arXiv:2608.14569v1 Announce Type: new.
      • Abstract: Neural solvers for constraint satisfaction problems have achieved remarkable in-distribution accuracy, yet they suffer from a fundamental limitation: persistent constraint violations occur under distributional shifts, even when the model reports high confidence.
      • This position paper argues that when hard constraints exist and verification costs are relatively low, neural constraint reasoning must prioritize symbolic integration over pure learning.
      • We justify our focus on Sudoku as a representative NP-complete testbed because it exhibits a sharp asymmetry between easy verification and hard solving: checking a candidate solution requires only polynomial time $O(n^{2})$, while finding a solution can demand an exponential search.
    • EN Highlights:
      • arXiv:2608.14569v1 Announce Type: new
      • Abstract: Neural solvers for constraint satisfaction problems have achieved remarkable in-distribution accuracy, yet they suffer from a fundamental limitation p…
      • This position paper argues that when hard constraints exist and the cost of verification is relatively low, neural constraint reasoning must prioritize symbolic…
      • We justify our focus on Sudoku as a representative NP-complete testbed because it exhibits a sharp asymmetry between easy verification and hard solving: checkin…

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

  • Does a Language Server Save Tokens for Coding Agents? A Measurement Methodology and Preliminary Study

    • Posted: 2026-08-18 12:00 Beijing Time
    • Abstract: - arXiv:2608.13568v1 Announce Type: new.
      • Abstract: Coding agents spend most of their context budget on retrieval.
      • Lexical retrieval (grep) is universal, instant, and zero-setup, but noisy: it cannot distinguish between definitions, calls, and comments.
      • Semantic retrieval via the Language Server Protocol (LSP) is precise and typed, but requires a running index server and incurs a round-trip cost per symbol.
    • EN Highlights:
      • arXiv:2608.13568v1 Announce Type: new
      • Abstract: Coding agents spend most of their context budget on retrieval
      • Lexical retrieval (grep) is universal, instant, and zero-setup, but noisy: it cannot tell a definition from a call from a comment

  • Semantic retrieval via the Language Server Protocol (LSP) is precise and typed, but needs a running, indexed server and pays a per-symbol round-trip

  • Think in Latent, Explain in Language: Self-Explainable Latent Reasoning

    • Publication Time: 2026-08-18 12:00 Beijing Time
    • Abstract:
      • arXiv:2608.13570v1 Announce Type: new.
      • Latent reasoning has emerged as a powerful alternative to text-based Chain-of-Thought (CoT), significantly improving computational efficiency by compressing lengthy reasoning into compact embeddings.
      • However, compressing reasoning into the latent space renders the thinking opaque, hindering its interpretability.
      • Current methods present a stark trade-off: they either function as unexplainable “black boxes” (e.g., Coconut), where the latent reasoning is not human-readable, or rely on separate post-hoc decoders for interpretability (e.g., Heima), introducing architectural overhead and decoupling the explanation from the actual reasoning process.
    • EN Highlights:
      • arXiv:2608.13570v1 Announce Type: new
      • Abstract: Latent reasoning has emerged as a powerful alternative to text-based Chain-of-Thought (CoT), offering significant gains in computational efficiency by…
      • However, compressing reasoning into the latent space renders the thinking opaque, hindering its interpretability
      • Current methods present a stark trade-off: they either function as unexplainable ‘‘black boxes’’ (e.g., Coconut), where the latent reasoning is not human-readab…

  • Not All Tokens Are Equal: Inflation-Aware Routing for Agentic LLM Systems

    • Publication Time: 2026-08-18 12:00 Beijing Time
    • Abstract:
      • arXiv:2608.13571v1 Announce Type: new.
      • When a language model fails to answer a query on the first attempt, agentic systems retry, consuming additional tokens with each attempt.
      • This retry overhead creates a gap between the price implied by the model’s per-token cost and the actual cost of the full workflow.
      • We call this gap \emph{token inflation} and define it as the ratio of the true workflow cost to the single-call cost.
    • EN Highlights:
      • arXiv:2608.13571v1 Announce Type: new
      • Abstract: When a language model fails to answer a query on the first attempt, an agentic system retries, consuming additional tokens each time
      • This retry overhead creates a gap between what a model’s per-token price implies and what a full workflow actually costs
      • We call this gap \emph{token inflation} and define it as the ratio of true workflow cost to single-call cost

  • BCMT: Blockwise Causal Memory Transformer

    • Publication Time: 2026-08-18 12:00 Beijing Time
    • Abstract:
      • arXiv:2608.13578v1 Announce Type: new.

  • Abstract: Transformer architectures rely on dense self-attention to model long-range dependencies, but this mechanism exhibits quadratic complexity with respect to sequence length.

  • We introduce BCMT (Blockwise Causal Memory Transformer), an architecture for long-context language modeling that decouples local token interactions from global context propagation.

  • Dense causal self-attention is applied independently within local blocks, while each block produces an adaptive summary aggregated through an exponential causal memory.

  • EN Highlights:

    • arXiv:2608.13578v1 Announce Type: new
    • Abstract: Transformer architectures rely on dense self-attention to model long-range dependencies, but this mechanism exhibits quadratic complexity with respect…
    • We introduce BCMT (Blockwise Causal Memory Transformer), an architecture for long-context language modeling that decouples local token interactions from global…
    • Dense causal self-attention is applied independently within local blocks, while each block produces an adaptive summary aggregated through an exponential causal…

  • Jais 2: A Family of Arabic-Centric Open Large Language Models

    • Publication Time: 2026-08-18 12:00 Beijing Time
    • Abstract: - arXiv:2608.13580v1 Announce Type: new.
      • Jais 2 is a family of Arabic-centric large language models developed jointly by MBZUAI, Cerebras, and Inception, designed to advance Arabic-centric language modeling with strong performance on the Arabic and cultural benchmarks evaluated in this report.
      • To our knowledge, the family includes the largest open Arabic-centric LLM trained from scratch at 70B parameters, and a competitive 8B-parameter variant among evaluated open models.
      • A custom Arabic-centric vocabulary enables efficient training and inference.
    • EN Highlights:
      • arXiv:2608.13580v1 Announce Type: new
      • Abstract: Jais 2 is a family of Arabic-centric large language models developed jointly by MBZUAI, Cerebras, and Inception, designed to advance Arabic-centric la…
      • The family includes, to our knowledge, the largest open Arabic-centric LLM trained from scratch at 70B parameters, and a competitive 8B-parameter variant among…
      • A custom Arabic-centric vocabulary enables efficient training and inference

  • IterCOMP: Reasoning-aware Adaptive Prompt Compression for Multi-hop Question Answering

    • Publication Time: 2026-08-18 12:00 Beijing Time
    • Abstract: - arXiv:2608.13588v1 Announce Type: new.
      • Multi-hop question answering requires complex reasoning across multiple evidence snippets, which often overwhelms retrieval-augmented generation systems with lengthy and noisy contexts, harming both efficiency and accuracy.
      • While existing prompt compression methods attempt to address this issue, they are typically designed for single-turn queries and fail to capture interdependent reasoning steps.
      • We propose IterCOMP, a unified, training-free prompt compression framework that incorporates multi-hop reasoning into an iterative compression loop.
    • EN Highlights:

  • Measuring Fairness in Large Audio Language Models via Semantic-Aware Bias Estimation

    • Release Time: 2026-08-18 12:00 Beijing Time
    • Abstract: - arXiv:2608.13624v1 Announce Type: new.
      • Abstract: The increasing use of Large Audio Language Models (LALMs) in audio understanding tasks such as speech recognition and audio question answering has raised concerns about fairness across demographic groups.
      • Fairness evaluation in spoken-input settings is challenging due to confounding factors, including semantic variations in spoken content and speaker-specific characteristics.
      • Ignoring these factors can lead to misleading conclusions about model bias.
    • EN Highlights:
      • arXiv:2608.13624v1 Announce Type: new
      • Abstract: Large Audio Language Models (LALMs) have seen increasing use for audio understanding tasks such as speech recognition and audio question answering, ra…
      • Fairness evaluation in spoken-input settings is challenging due to confounding factors, including semantic variation in spoken content and speaker-specific char…
      • Ignoring these factors can result in misleading conclusions about model bias

  • GRPO Beyond English: A Large-Scale Study of GRPO in Non-English and Multilingual Settings

    • Release Time: 2026-08-18 12:00 Beijing Time
    • Abstract: - arXiv:2608.13698v1 Announce Type: new.
      • Abstract: Reinforcement Learning with Verifiable Rewards (RLVR), often optimized through Group Relative Policy Optimization (GRPO), has become a core method for enhancing the reasoning capabilities of pretrained language models, yet current research remains predominantly English-centric.
      • We conducted a large-scale empirical study on multilingual and non-English GRPO, involving a wide range of base models, training languages, and different reasoning language rewards.
      • We found that native-language reasoning training often shows a small gap compared to English reasoning training.
    • EN Highlights:
      • arXiv:2608.13698v1 Announce Type: new
      • Abstract: Reinforcement Learning with Verifiable Rewards (RLVR), often optimized with Group Relative Policy Optimization (GRPO), has become a central recipe for…

  • We conduct a large-scale empirical study of multilingual and non-English GRPO across a wide range of base models, training languages, and different reasoning la…

    • We find that training to reason in the native language often leaves only a small gap to training for English reasoning

  • CLAIR-Fin: An Adversarial Multi-Agent Framework for Claim-Level Verification and Adaptive Debate in Cross-Modal Financial QA

    • Publication Time: 2026-08-18 12:00 Beijing Time
    • Abstract: - arXiv:2608.13706v1 Announcement Type: New.
      • Abstract: In retrieval-augmented and multi-agent pipelines, existing defenses against hallucination remain one-sided: evidence is trusted despite modal disagreements, debates verify the overall report rather than individual claims, and this verification only occurs post-drafting, leaving inter-agent errors undiscovered before the final text.
      • To bridge this gap, we introduce CLAIR-Fin, a nine-agent framework that breaks down each question into atomic claims maintained in a typed financial claim ledger.
      • Each claim is resolved through an Asymmetric Evidence Authority, which determines evidence trust based on claim type rather than treating all modalities as equally reliable; Chain of Custody Verification, which checks for grounding at the handoff between drafting and adversarial review, not just at the pipeline exit; an Adaptive Rebuttal Loop, which directs contentious claims through adversarial debate with a depth proportional to what the debate uncovers; and a final Inevitability Audit combined with a continuous hallucination risk index, which distinguishes claims that have passed review from those that were never challenged.
    • EN Highlights:
      • arXiv:2608.13706v1 Announce Type: new
      • Abstract: Existing defenses against hallucination in retrieval-augmented and multi-agent pipelines remain partial: evidence is trusted despite modality disagree…
      • To close this gap, we present CLAIR-Fin, a nine-agent framework that decomposes each question into atomic claims maintained in a typed Financial Claim Ledger
      • Each claim is resolved through Asymmetric Evidence Authority, which conditions evidence trust on claim type rather than treating all modalities as equally relia…

  • TeachMateGPT: A Multi-Agent Knowledge-Grounded Framework for Pedagogical Assessment Generation from Science Curriculum Materials

    • Publication Time: 2026-08-18 12:00 Beijing Time
    • Abstract: - arXiv:2608.13708v1 Announcement Type: New.
      • Abstract: Automatically generating textbook-based assessment items can reduce the workload of science teachers, but existing Retrieval-Augmented Generation (RAG) systems rely on flat retrieval, support only single-question generation, lack safeguards against weak evidence, and are ill-suited for curricula with scarce resources and exam structures.
      • We address these limitations with TeachMateGPT, a multi-agent system that introduces four advancements for curriculum-based science assessment authoring.
      • (i) COPE, a hierarchical knowledge base that replaces token-window chunking with a multi-resolution index that segments documents along the syllabus structure and links them at three granularities via a traversable graph-based lineage, matching evidence to the instructional level of each topic.
    • EN Highlights:
      • arXiv:2608.13708v1 Announce Type: new

  • Abstract: Automatically generating textbook-grounded assessment items can reduce science teachers’ workload, but existing retrieval-augmented generation (RAG) s…

    • We address these limitations with TeachMateGPT, a multi-agent system contributing four advances to curriculum-grounded science-assessment authoring
    • (i) COPE, a hierarchical knowledge base replacing token-window chunking with a multi-resolution index that segments documents along syllabus structure and links…

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

  • Learning Discrete Riemannian Metrics for Physical Fields with Cochain-Frame Equivarianc

    • Release Time: 2026-08-18 12:00 Beijing Time
    • Abstract: - arXiv:2608.14556v1 Announce Type: new.
      • Physical fields on meshes require a separation of topology and geometry: conservation laws are topological and should be exact, while geometry, material response, and anisotropic coupling must be learned from data.
      • Existing neural agents often mix these roles in unconstrained message passing.
      • We introduce Riemannian Hodge Message Passing (RHMP), which turns this separation into an architectural principle.
    • EN Key Points:
      • arXiv:2608.14556v1 Announce Type: new
      • Abstract: Physical fields on meshes require a separation between topology and geometry: conservation laws are topological and should be exact, while geometry, m…
      • Existing neural surrogates often mix these roles inside unconstrained message passing
      • We introduce Riemannian Hodge Message Passing (RHMP), which turns this separation into an architectural principle

  • Forward Pass Domain Adaptation (Without Cross-Layer Backpropagation)

    • Release Time: 2026-08-18 12:00 Beijing Time
    • Abstract: - arXiv:2608.14563v1 Announce Type: new.
      • Forward-Pass-Only MLP training (FPO) adapts large language models without a backward pass through the model body, achieving 2.7–3.2x the throughput of standard fine-tuning with approximately 40% less peak training memory, while preserving out-of-domain benchmarks within the seed noise of the baseline, a property that full-network fine-tuning cannot reliably reproduce.
      • FPO relies on a single empirical observation: in the later layers of a Transformer, the output layer prediction error approximates the true gradient with a cosine similarity of 0.47–0.59 across the six public models we investigated.
      • We introduce a two-minute diagnostic that can quantify the approximation per-layer for any model, thus determining the feasibility of late-layer adaptation.
    • EN Key Points:
      • arXiv:2608.14563v1 Announce Type: new
      • Abstract: Forward-Pass-Only MLP training (FPO) adapts large language models without a backward pass through the model body, achieving 2.7–3.2x the throughput o…

  • FPO rests on a single empirical observation: at late layers of a transformer, the output-layer prediction error approximates the true gradient with cosine simil…

  • We introduce a two-minute diagnostic that quantifies this approximation per layer for any model, identifying where late-layer adaptation is viable

  • Coarse-to-Fine Multi-Resolution Diffusion Models for Trajectory Generation in Urban Systems

    • Publication Time: 2026-08-18 12:00 Beijing Time
    • Abstract: - arXiv:2608.14570v1 Announcement Type: New.
      • Abstract: Understanding human mobility is crucial for a wide range of urban applications, including traffic management, epidemic control, and urban planning.
      • However, due to privacy concerns, the availability of large-scale public trajectory data remains limited, which poses challenges for downstream mobility analysis.
      • Existing methods for synthetic trajectory generation primarily focus on matching global distribution similarity, while often overlooking movement patterns at different spatial and temporal resolutions, which are crucial for practical applications.
    • EN Key Points:
      • arXiv:2608.14570v1 Announce Type: new
      • Abstract: Understanding human mobility is critical for a wide range of urban applications, including traffic management, epidemic control, and urban planning
      • However, due to privacy concerns, the availability of large-scale public trajectory data remains limited, posing challenges for downstream mobility analysis
      • Existing methods for synthetic trajectory generation primarily focus on matching global distribution similarity, while often overlooking mobility patterns acros…

  • Geometry Is Not Robustness: A Trajectory-Level Study of PGD Evaluation

    • Publication Time: 2026-08-18 12:00 Beijing Time
    • Abstract: - arXiv:2608.14594v1 Announcement Type: New.
      • Abstract: Projected Gradient Descent (PGD) is widely used for evaluating adversarial robustness, typically assessed by final adversarial accuracy, but it does not capture the model’s behavior throughout the attack process.
      • Recent work has proposed trajectory-level diagnostics, such as loss evolution, gradient alignment, and steps-to-failure, to gain deeper insights into adversarial optimization dynamics.
      • However, it remains unclear whether these diagnostics reliably indicate the strength of robustness.
    • EN Key Points:
      • arXiv:2608.14594v1 Announce Type: new
      • Abstract: Projected Gradient Descent (PGD) is widely used to evaluate adversarial robustness, typically via final adversarial accuracy, which does not capture m…
      • Recent work proposes trajectory-level diagnostics, such as loss evolution, gradient alignment, and steps-to-failure, for deeper insight into adversarial optimis…

  • However, whether these diagnostics reliably indicate robustness strength remains unclear

  • DumpsterCluster: From Dumpster Diving to Serving LLaMA-70B on $60 GPUs

    • Publication Time: 2026-08-18 12:00 Beijing Time
    • Abstract: - arXiv:2608.14614v1 Announcement Type: New.
      • Abstract: As AI datacenters retire functional GPUs, vast quantities of still-capable accelerators are entering secondary markets.
      • This paper investigates whether these retired GPUs can find a productive afterlife to form a “DumpsterCluster” that can serve modern LLM inference, and under what conditions this reuse is economically feasible and environmentally sustainable.
      • We physically built a 128-GPU DumpsterCluster from scratch using only second-hand components and ran it for one year.
    • EN Highlights:
      • arXiv:2608.14614v1 Announce Type: new
      • Abstract: As AI datacenters retire functional GPUs, vast quantities of still capable accelerators enter secondary markets
      • This paper investigates whether these retired GPUs can find a productive afterlife to form a DumpsterCluster that can serve modern LLM inference, and under what…
      • We physically built a 128-GPU DumpsterCluster from scratch using only second-hand components and ran it for one year

  • Calibrated Trust, Not Sharper Prediction: An Empirical Test of Uncertainty Fusion

    • Publication Time: 2026-08-18 12:00 Beijing Time
    • Abstract: - arXiv:2608.14617v1 Announcement Type: New.
      • Abstract: A recurring proposal in legal AI is to improve case-outcome prediction by fusing uncertainty tools (evidence graphs with belief propagation, sequential Bayesian odds updating, Dempster-Shafer combination, and conformal prediction) into a pipeline.
      • We test this on 1,000 real European Court of Human Rights cases from LexGLUE and FairLex, predicting whether the Court found a Convention violation from the case’s fact paragraphs.
      • We compare three families across two frontier LLMs (Claude Opus 4.8 and GPT-5.5) as per-fact evidence estimators: (A) the raw LLM, (B) the LLM routed through the fusion pipeline, and (C) a term-frequency baseline through the same pipeline.
    • EN Highlights:
      • arXiv:2608.14617v1 Announce Type: new
      • Abstract: A recurring proposal in legal AI is to improve case-outcome prediction by fusing uncertainty tools (evidence graphs with belief propagation, sequentia…
      • We test this on 1,000 real European Court of Human Rights cases from LexGLUE and FairLex, predicting whether the Court found a Convention violation from the cas…
      • We compare three families across two frontier LLMs (Claude Opus 4.8 and GPT-5.5) as per-fact evidence estimators: (A) the raw LLM, (B) the LLM routed through th…

  • PIKFNO: An Interpretable Neural Operator Based on Physics Informed Kernel Function

    • Publication Time: 2026-08-18 12:00 Beijing Time
    • Abstract: - arXiv:2608.14619v1 Announcement Type: new.
      • Abstract: This work proposes a new interpretable neural operator framework, called the Physics-Informed Kernel Function Neural Operator (PIKFNO), which explicitly incorporates physics-informed kernel functions derived from governing equations into the neural operator architecture.
      • Unlike traditional neural operators such as DeepONet, which rely on deep networks to implicitly learn basis functions, PIKFNO constrains the trunk network with physics-informed kernel functions, aligning its operator structure with the kernel expansions used in mesh-free methods.
      • Two construction strategies are introduced: one learns kernel functions directly from data, where the learned kernel can be regarded as a non-singular fundamental solution, while the other constructs them through transformations of analytical fundamental solutions.
    • EN Highlights:
      • arXiv:2608.14619v1 Announce Type: new
      • Abstract: This work proposes a new interpretable neural operator framework, termed the Physics Informed Kernel Function Neural Operator (PIKFNO), which explicit…
      • Unlike traditional neural operators such as DeepONet, which rely on deep networks to implicitly learn basis functions, PIKFNO constrains the trunk network throu…
      • Two construction strategies are introduced: one learns kernel functions directly from data, where the learned kernel can be regarded as a nonsingular fundamenta…

  • Explaining Reinforcement Learning Decisions in Self-adaptive Systems

    • Publication Time: 2026-08-18 12:00 Beijing Time
    • Abstract: - arXiv:2608.14620v1 Announcement Type: new.
      • Abstract: Reinforcement Learning (RL) has been widely applied in autonomous and self-* systems, but RL policies, especially Deep RL policies that rely on neural networks, lack transparency and are difficult to understand.
      • This can lead to reduced user trust and make system verification more challenging.
      • To address this challenge, this paper introduces Explanations using Alternative Realities for Reinforcement Learning (EARL), a Python library for generating counterfactual explanations in an RL setting.
    • EN Highlights:
      • arXiv:2608.14620v1 Announce Type: new
      • Abstract: Reinforcement Learning (RL) has been extensively used in autonomous and self-* systems, but RL policies, especially deep RL ones relying on neural net…
      • This can lead to diminished user trust, and makes for a more challenging verification of systems
      • To address this challenge, this paper introduces Explanations using Alternative Realities for Reinforcement Learning (EARL), a Python library to produce counter…

  • Metaplasticity as adaptive gradient preconditioning for incremental learning

    • Publication Time: 2026-08-18 12:00 Beijing Time
    • Abstract: - arXiv:2608.14634v1 Announcement Type: new.

  • Abstract: Biological intelligence naturally prevents catastrophic forgetting through Complementary Learning Systems (CLS) theory, a macroscopic consolidation process driven by synaptic plasticity at the local level: the continuous, history-dependent neuromodulation of individual synapses.

  • While artificial neural networks struggle with the stability-plasticity dilemma in non-stationary environments, existing solutions often require task labels or incur significant memory overhead, contradicting biological reality.

  • Re-framing this localized neuromodulation as an optimization-driven process, we introduce $\textbf{SynGAP}$: $\textbf{Syn}$aptic $\textbf{G}$eometric $\textbf{A}$daptive $\textbf{P}$reconditioning.

    • EN Highlights:
      • arXiv:2608.14634v1 Announce Type: new
      • Abstract: Biological intelligence naturally prevents catastrophic forgetting through Complementary Learning Systems (CLS) theory, a macroscopic consolidation pr…
      • While artificial neural networks struggle with the stability-plasticity dilemma in non-stationary environments, existing solutions often require task labels or…
      • Re-framing this localized neuromodulation as an optimization-driven process, we introduce $\textbf{SynGAP}$: $\textbf{Syn}$aptic $\textbf{G}$eometric $\textbf{A…

  • Fractional Optimizers Meet Fractal Activation Functions: An Empirical Study of Multi-Scale Optimization in Neural Network

    • Publication Time: 2026-08-18 12:00 Beijing Time
    • Abstract: - arXiv:2608.14636v1 Announce Type: new.
      • Abstract: Fractional optimization methods and fractal activation functions are two independent directions for improving neural network training.
      • Fractional optimizers extend first-order optimization through fractional derivatives and memory effects, whereas fractal activations introduce multi-scale nonlinear representations based on self-similar Weierstrass and Blancmange-type functions.
      • Here, we investigate their interaction within a unified experimental framework.
    • EN Highlights:
      • arXiv:2608.14636v1 Announce Type: new
      • Abstract: Fractional optimization methods and fractal activation functions are two independent directions for improving neural network training
      • Fractional optimizers extend first-order optimization through fractional derivatives and memory effects, whereas fractal activations introduce multi-scale nonli…
      • Here, we investigate their interaction within a unified experimental framework