2026-06-10 AI Daily | From Gemma 4 to Claude Fable, AI Deployment Enters a Dual-Track Era of Safety Tiering and On-Device Explosion Link to heading

On the same day, Google DeepMind released the Gemma 4 12B unified multimodal model and Gemini 3.5 Live real-time translation, with on-device AI capabilities far exceeding expectations. Anthropic launched Claude Fable 5, opening higher capability tiers to the public through safety classifiers, signaling a new paradigm for model deployment centered around “capability licensing boundaries.” Meanwhile, the developer community is shifting to “results engineering” to drive AI programming, as human-agent collaboration evolves from prompt debugging to goal definition.

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

Today, Google DeepMind released three major announcements in a row, from the near-real-time natural speech translation of Gemini 3.5 Live Translate, to the unified, encoder-free multimodal model Gemma 4 12B, and its forward-looking strategy for European robotics. These developments clearly outline the next piece of the puzzle for AI infrastructure and global implementation. It is recommended that all ecosystem participants read each announcement in detail.

Meanwhile, Codex is reshaping the fundamental logic of software engineering. The engineering teams at Nextdoor and Notion have shared highly insightful practical reviews: they are no longer focusing on repeatedly debugging prompts but are shifting to “results engineering.” Engineers now directly define their desired outcomes and collaborate with agents to design implementation paths. Even managers who haven’t written production code in years are returning to the codebase. This is a strong signal of an upward shift in the developer’s role, and it’s a trend all engineering teams should watch.

Finally, a reflection on the iPhone’s “last fortress” is worth the attention of product professionals. It astutely points out that the endgame for Agents isn’t helping people operate computers, but making the entire process from request to result invisible to the user. When interactions become “invisible,” the rules of growth and experience will be rewritten. This is a point worth pondering, especially when combined with Ernst & Young’s insights on the shifting investment landscape for Agentic AI.

🌐 Breaking AI News on X Link to heading

Topic 1: Anthropic Launches Claude Fable 5 as Most Capable Public AI Model Link to heading

• Category: AI · News
• Overview: Trending for: 1 day ago, Related posts: 107,000
• What it is: Anthropic has reportedly launched Claude Fable 5, positioned as the first Mythos-class model available to the public. It boasts a capability level higher than Opus and includes built-in safety guardrails against dual-use cyber threats.
• Why it matters: This marks a shift where frontier AI models are no longer just competing on intelligence but are also being tiered based on “capability licensing boundaries.” Previously restricted top-tier abilities are now being opened to the public through safety distillation and policy controls, potentially defining a new paradigm for model deployment.
• Discussion summary: The discussion centers on the “guardrail tax”—how much of the original Mythos-level capability in long-range tasks, coding, and autonomy the public Fable version sacrifices for safety. Users are also questioning if it can block malicious cyber abuse while preserving legitimate security research functions, and how its relationship with Opus and the restricted Mythos will reshape Claude’s product hierarchy.

Topic 2: Kimi AI Swarm Predicts Spain and France to Win 2026 World Cup Link to heading

• Category: AI · Sports
• Overview: Trending for: 7 hours ago, Related posts: 161
• What it is: The Kimi AI Swarm system predicted that Spain and France will win the 2026 World Cup.
• Why it matters: This prediction showcases the potential of AI in forecasting complex sporting events, promoting the application of data analysis and swarm intelligence in non-traditional domains. It could influence future betting, team strategies, and the formation of public opinion.
• Discussion summary: The debate focuses on the prediction’s accuracy, its potential over-reliance on historical data, the actual strengths of the Spanish and French teams, and whether AI swarm decisions are superior to judgments made by human experts.

Topic 3: Developers Debate Codex vs Claude Code AI Agents Link to heading

• Category: AI · News
• Overview: Trending for: 9 hours ago, Related posts: 877
• What it is: The developer community is in a heated debate over whether OpenAI’s Codex or Anthropic’s Claude Code is the superior AI programming agent, with widespread discussion about rumored upgrades for both.
• Why it matters: AI programming agents are one of the most mature application areas for AI. The competition between these two giants directly shapes developer workflows and tool standards, marking an evolution from simple code completion to deep engineering collaboration.
• Discussion summary: The core points of contention are: 1) Whether Claude Code’s agent-like capabilities in handling complex engineering tasks are superior to Codex/Copilot; 2) How to weigh the cost-effectiveness of Codex against the performance advantages of Claude Code in high-end configurations; 3) Whether to choose a single tool or a combined stack (e.g., Cursor for orchestration, Claude Code for deep execution, and Codex for workflow reviews).

Topic 4: Tesla’s FSD Supervised Gains Approval in Denmark, Fourth European Country Link to heading

• Category: AI · News
• Overview: Trending Time: 19 hours ago, Related Posts: 34,000
• What it is: Tesla’s Full Self-Driving (Supervised) feature has received regulatory approval in Denmark, becoming the fourth European market where it is approved, following the Netherlands, Norway, and Sweden.
• Why it matters: This indicates that assisted driving solutions based on pure vision and end-to-end neural networks are breaking through Europe’s notoriously strict regulatory barriers. It provides valuable compliance experience for deploying AI in safety-critical scenarios and serves as a bellwether for the global commercialization of autonomous driving.
• Discussion Summary: The discussion focuses on whether the safety validation for the approval is sufficient, the controversy over the misleading nature of the name “Full Self-Driving” versus its actual capabilities, and whether Denmark’s decision will accelerate the approval process in major EU markets like Germany. There are also expressions of both anticipation and concern about its performance and adaptation to local road conditions.

Topic 5: HeyGen Brings Video Creation to Claude AI with Hyperframes Link to heading

• Category: AI · News
• Overview: Trending Time: , Related Posts: 985
• What it is: HeyGen has launched its Hyperframes technology, integrating AI video creation capabilities into Anthropic’s Claude AI platform.
• Why it matters: This move signifies a deeper integration between large language models and AI video generation tools, expanding Claude’s multimodal content creation capabilities and potentially reshaping creative workflows.
• Discussion Summary: Users on X are focused on the technical implementation of Hyperframes, its differences from competitors like Sora, and its impact on video creators. There are varied opinions regarding its actual effectiveness and commercialization potential.

Topic 6: The Sandbox Launches AI Engine for Rapid Game Creation Link to heading

• Category: AI · News
• Overview: Trending Time: , Related Posts: 102
• What it is: The metaverse platform The Sandbox has released an AI engine that allows users to rapidly generate game content.
• Why it matters: This demonstrates the practical application of generative AI in creating interactive 3D content, which is expected to significantly lower the barrier to entry for game development and accelerate the construction of virtual worlds.
• Discussion Summary: The community is focused on the actual generation quality of the engine, whether it preserves creative freedom, and how it will integrate with the existing creator economy. There is some disagreement about its true value within the NFT and metaverse ecosystem.

Topic 7: World Labs Launches ICARE: Browser-Based 3D Adventure with Genius Icons Link to heading

• Category: AI · News
• Overview: Trending Time: 5 hours ago, Related Posts: 75
• What it is: World Labs has released a browser-based 3D adventure experience called ICARE, where users can interact with avatars of various historical “geniuses.”
• Why it matters: This showcases new possibilities for combining spatial intelligence and generative AI in real-time 3D environments, providing a cutting-edge example for immersive education, AI-driven narratives, and lightweight browser-based 3D applications.
• Discussion Summary: Discussions on X primarily revolve around whether its technical implementation truly relies on generative 3D, the entertainment and educational value of the experience, and the differentiated path of World Labs (Fei-Fei Li’s team) in the spatial intelligence race. Some users are divided on the depth of interaction with the “genius icons” and the performance in the browser.

AI Public Opinion Summary on X Today Link to heading

Today’s main theme in public opinion is the AI industry’s shift from competing on pure intelligence to establishing “capability license boundaries.” Events like Anthropic releasing higher-capability models to the public through safety distillation and Tesla’s FSD breaking through another strict European regulation demonstrate a new paradigm where cutting-edge models must be bundled with sophisticated safety guardrails for mass adoption. The community consensus is that the practical application of AI programming agents, multimodal generation, and predictive analytics has significantly improved. However, disagreements on technical paths—such as whether Codex or Claude Code is better for complex engineering, or if the Kimi World Cup predictions rely too heavily on historical data—reveal that tool standardization is far from established. Deeper divisions focus on how much core capability is sacrificed for the “guardrail tax,” whether the “Full Self-Driving” name is misleading, and the trade-off between the true quality of generative 3D/video content and creative freedom. Underlying these debates are multidimensional risks: excessive safety restrictions could weaken the practical value of models, inadequately validated naming and low-quality generated content could exacerbate the trust deficit, and the fragmented competition among programming tools, along with improper reliance on predictive models, could lead to misjudgments and regulatory backlash in safety-critical scenarios and public discourse.

💡 Influencer Insights Link to heading

AI Industry Dynamics Daily (2026-06-09) Link to heading

I. Today’s Key Technology Trends and Product Hotspots Link to heading

1. Claude Fable 5 Released: Anthropic’s “Safe Version of Mythos” Link to heading

@dotey provided a detailed analysis of the two new models released by Anthropic on the same day:

• Claude Fable 5: For general users, features a built-in safety classifier and automatically downgrades to Opus 4.8 for processing when sensitive content is detected.
• Claude Mythos 5: Removes some safety restrictions and is available only to Project Glasswing cybersecurity partners.

Key Data: Pricing is 60% lower than Mythos Preview ($10/million tokens for input, $50/million for output), but still twice the price of Opus 4.8. Pro/Max/Team/Enterprise users can use it for free until June 22, after which they will need to purchase usage credits.

Policy Change: Traffic for Mythos-level models is now mandatorily retained for 30 days for security monitoring, breaking the previous zero-retention promise.

2. On-device Model Explosion: The Maturation of Gemma 4 and the Local AI Ecosystem Link to heading

@zhixianio intensively tested Google’s on-device layout:

• Gemma 4 12B: A unified multimodal model that runs on an M5 Max 128G via mlx-vlm. English/Japanese speech recognition is “instantaneous,” while Chinese performance is subpar.
• Gemma 4 E4B + MTP: Excellent performance on Japanese email parsing and classification tasks.
• QAT (Quantization-Aware Training): @zhixianio notes this is an optimization approach that “assumes quantization will occur during training.” Google’s focus on on-device models suggests that “Android will soon have native models.”

Real-world Feedback: @zhixianio ran Qwen3.6-35B-A3B-oQ6-fp16-mtp on oMLX, noting that the “response speed is faster than remote LLMs, and its intelligence is on point,” and that the capability of on-device models “far exceeds expectations.”

3. AI Agent Browsers and Automation Tools Link to heading

@vista8 conducted an in-depth review of the Aye Browser developed by @okasupportgroup:

• Based on Chromium, it fully simulates human operations with AI (not a CLI/plugin, which helps avoid account anomaly detection).
• Features built-in Skill recording and scheduled execution, supporting tasks like auto-replying to Little Red Book comments and transcribing articles across multiple platforms.
• Integrates an RSS reader, ad blocker, and video translation/download features.

@vista8 suggests: It needs to support Chrome account migration and the plugin ecosystem, and should clarify its paid plans soon.

II. Unique Perspectives and Industry Foresight Link to heading

1. “Contract First”—The Best Practice for Vibe Coding Link to heading

@Pluvio9yte’s summary from transitioning from a security professional to a full-stack developer:

“The best practice for Vibe Coding isn’t actually Requirement First or Code First, it’s Contract First. Without a well-defined contract, everything else is just empty talk.”

He developed a framework based on a customized version of OpenSpec that “externalizes easily shifting context into a contract,” giving both humans and AI a stable reference point.

2. The “Innovator’s Dilemma” of WeChat’s AI Ecosystem Link to heading

@dotey has repeatedly criticized WeChat’s AI strategy:

“WeChat always thinks it’s an OS, but it’s not. It’s just a behemoth living on top of the phone’s operating system… In the future, WeChat’s role as an entry point will diminish. The younger generation won’t open WeChat; they’ll just ask their Agent.”

He suggests that WeChat should “use a team that is physically and financially separate from the current one to create an almost completely independent AI app.”

3. The Cost Paradox of AI Programming Link to heading

@ruanyf cites token consumption data from the founder of OpenClaw:

“603 billion tokens in one month, worth $1.3 million… Even if we switch to cheaper models, it would still cost 2-3 million RMB per year. Companies will find that with unlimited use, AI programming is much more expensive than human programmers.”

4. Exploring the Packaging of Skills Link to heading

@lijigang proposes two development paths for LLMs:

• Go downwards: Atomization, breaking them down into skill packages for specific tasks.
• Go upwards: Componentization, encapsulating best practices for scenarios (workflows, node optimization, skill packages).

And asks: “Could the browser extension mechanism be a possible answer?”

5. The “Shadow Book” Reading Method Link to heading

@lijigang proposes a reading method unique to the AI era:

“In the age of print, we could only read the book the author wrote. The unique action in the AI era is to read the shadow books—whenever you encounter an assertion, immediately use AI to analyze the three schools of thought that oppose it, the premises the author skipped, from whom the ideas were inherited…”

III. Recommended Tools and Resources Link to heading

Development Tools Link to heading

Productivity Tools Link to heading

Content Creation Link to heading

Models & APIs Link to heading

IV. Noteworthy Data Points Link to heading

• GitHub Commits: 14 times higher in the first three months of this year compared to the same period last year (@ruanyf)
• Zhipu’s Market Cap: Equal to Xiaomi, about two JD.coms, “the world’s most valuable open-source software company” (@ruanyf)
• SEO Value: A certain site gets 11,000 daily organic search visits, saving 300,000 in monthly marketing costs (@gefei55)
• Domain Investment: A .com domain registered for about $10 was sold for $1000 half a month later (@gefei55)

Report compiled from public posts on the X platform within 24 hours around 2026-06-09

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

Timeframe: Last 3 days; covers 22 sources; 37 updates in total

All-In Podcast (A_full) Link to heading

• Bill Maris: How Google Could Crush AI Competitors, Why Small Funds Win, and AI’s Atari Stage
  • Release Time: 2026-06-09 23:07 Beijing Time
  • Abstract: - EY - Agentic AI is introducing new rules for investment.
    • As AI shifts to consumption-based models, EY is linking spending to enterprise value.
    • NYSE - Thanks to our partner the New York Stock Exchange - a modern marketplace and exchange committed to building the future.
    • Plaud, our official wearable AI note-taking partner at the All-In Liquidity Summit, captured every insight.
    • Bill Maris: How Google Could Crush AI Competitors, Why Small Funds Win, and AI’s Atari Stage.
  • EN Highlights:
    • (0:00) Bill Maris joins the Besties
• (0:33) Four critical lessons from a career in technology
• (5:58) Building Google Ventures with data and machine learning
• (9:51) Why small VC funds beat big ones on average

Stratechery by Ben Thompson (A_full) Link to heading

• The iPhone’s Last Stand
  • Published: 2026-06-09 18:00 Beijing time
  • Abstract: - Listen to this post**:**.
    • For years, Apple fans would sneer at Microsoft’s tendency to talk about products that might or might not be released, mocking them as vaporware. -> This becomes even clearer when you consider the next wave of artificial intelligence: agents.
    • The purpose of an agent is not to use the computer for you; it’s to accomplish a specific task.
    • At least in theory, everything between the request and the result should be invisible to the user.
  • EN Key points:
    • Listen to this post :
    • Log in to listen
    • Apple fans would, for years and years, sneer at Microsoft’s penchant for talking about products that may or may not ship, deriding them as vaporware
    • After Apple’s bungled 2024 launch of Apple Intelligence and new Siri , however, vaporware is fair game, and just in time for this Article

OpenAI Blog (A_full) Link to heading

• How engineers at Nextdoor use Codex to build without limits

  • Published: 2026-06-09 20:00 Beijing time
  • Abstract: - A product like Nextdoor, which serves over 110 million users in 11 countries, places many demands on the platform team.
    • For engineering lead Cory Dolphin, Codex represents a significant shift: “From iteratively prompting an agent to results engineering, where an engineer starts to think about the outcome they want to see and works with the agent to design that result.”
    • This means individual engineers move up the stack - no longer locked in as an expert in a certain system or framework, they are able to more or less own an end-to-end product experience, even across multiple platforms.
    • The acceleration in productivity has been so fast that the bottleneck is no longer engineering, but the hard strategic questions about what to build next. -> “Codex has fundamentally changed how we think about engineering, to the point where we can’t even imagine it without it.”
  • EN Key points:
    • How engineers at Nextdoor use Codex with GPT-5.5 to investigate hard-to-reproduce issues, build across platforms, and focus on product outcomes.

• What Codex unlocks for Notion

  • Published: 2026-06-09 18:00 Beijing time
  • Abstract: - At Notion, Codex is changing how engineers build.
    • The company is rethinking the software primitives and abstractions it builds so that agents can use them.
    • When bringing new engineers onto the team, they are hired for curiosity and open-mindedness because the years of experience typically required for the field don’t exist yet.
    • Managers who haven’t written production code in years are getting back into the codebase, shipping alongside their teams.
    • Ryan Nystrom leads AI product engineering at Notion.
  • EN Key points:
    • How Notion uses Codex to one-shot specs, build AI Voice Input for the web, and multiply engineering power across small teams.

Google DeepMind Blog (A_full) Link to heading

• Fluid, natural voice translation with Gemini 3.5 Live Translate

  • Posted: 2026-06-09 23:16 Beijing Time
  • Abstract: - Gemini 3.5 Live Translate brings near real-time, natural voice translation to Google AI Studio, Google Translate and Google Meet.
    • This article from the Google DeepMind blog explains how the fluid, natural voice translation of Gemini 3.5 Live Translate is shaping the broader artificial intelligence and infrastructure landscape.
    • The introduction of fluid, natural voice translation with Gemini 3.5 Live Translate also has practical implications for founders, operators, and investors.
  • EN Highlights:
    • Gemini 3.5 Live Translate brings near real-time, natural speech translation to Google AI Studio, Google Translate and Google Meet.

• Introducing Gemma 4 12B: a unified, encoder-free multimodal model

  • Posted: 2026-06-09 22:10 Beijing Time
  • Abstract: - Introducing Gemma 4 12B: a unified, encoder-free multimodal model.
    • This article from the Google DeepMind blog explains how the introduction of Gemma 4 12B, a unified, encoder-free multimodal model, is shaping the broader artificial intelligence and infrastructure landscape.
    • Following the introduction of Gemma 4 12B, a unified, encoder-free multimodal model, it also has practical implications for founders, operators, and investors.
  • EN Highlights:
    • Introducing Gemma 4 12B: a unified, encoder-free multimodal model

• Powering the future of robotics in Europe

  • Posted: 2026-06-09 22:02 Beijing Time
  • Abstract: - Powering the future of robotics in Europe.
    • This article from the Google DeepMind blog explains how “Powering the future of robotics in Europe” is shaping the broader artificial intelligence and infrastructure landscape.
    • Following “Powering the future of robotics in Europe,” it also has practical implications for founders, operators, and investors.
  • EN Highlights:
    • Powering the future of robotics in Europe

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

• PathoSage: Towards Multi-Source Evidence Adjudication in Pathology via Experience-Aware Agentic Workflow

  • Posted: 2026-06-09 12:00 Beijing Time
  • Abstract: - arXiv:2606.07549v1 Announce Type: new.
    • Abstract: Recent advances in Multimodal Large Language Models (MLLMs) and agent workflows have shown strong promise for computational pathology, yet reliable patch-level reasoning remains challenging.
    • End-to-end pathology MLLMs often hallucinate morphological features, while recent agentic systems typically consolidate tool outputs and retrieved knowledge into a shared context, making decisions susceptible to contradictory evidence and context contamination.
    • We propose PathoSage, a three-stage framework that explicitly distinguishes between knowledge retrieval, evidence collection, and evidence adjudication for patch-level pathology multimodal reasoning.
  • EN Highlights:
    • arXiv:2606.07549v1 Announce Type: new
    • Abstract: Recent advances in Multimodal Large Language Models (MLLMs) and agent workflows have shown strong promise for computational pathology, yet reliable pa…

• End-to-end pathology MLLMs often hallucinate morphological features, while recent agentic systems usually merge tool outputs and retrieved knowledge into a shar…

• We propose PathoSage, a three-stage framework that explicitly separates knowledge retrieval, evidence collection, and evidence adjudication for patch-level path…

• OmniMem: Perturbation-aware Memory Compression for Streaming Audio-Visual LLMs

  • Release Time: 2026-06-09 12:00 Beijing Time
  • Abstract: - arXiv:2606.07577v1 Announce Type: new.
    • Abstract: Audio-visual large language models (LLMs) hold strong promise for long-form video understanding, yet their long-video inference is fundamentally limited by the linear growth of video tokens and key-value (KV) cache.
    • We introduce OmniMem, a memory-efficient streaming framework designed specifically for audio-visual LLMs.
    • Unlike existing compression methods that treat all tokens uniformly, OmniMem introduces a modality-aware memory allocation strategy that separately manages visual and audio contexts, addressing the severe token imbalance between the two modalities.
  • EN Highlights:
    • arXiv:2606.07577v1 Announce Type: new
    • Abstract: Audio-visual large language models (LLMs) hold strong promise for long-form video understanding, yet their long-video inference is fundamentally limit…
    • We present OmniMem, a memory-efficient streaming framework designed specifically for audio-visual LLMs
    • Unlike existing compression methods that treat all tokens uniformly, OmniMem introduces a modality-aware memory allocation strategy that separately manages visu…

• Syll: Open-Source Personal Automation with Cross-Surface Execution

  • Release Time: 2026-06-09 12:00 Beijing Time
  • Abstract: - arXiv:2606.07594v1 Announce Type: new.
    • Abstract: Personal AI agents must increasingly operate across APIs, shells, Web interfaces, and desktop GUIs, yet many systems remain tuned to a single interface and provide limited support for user instruction and auditability.
    • We introduce Syll, an open-source, self-hosted, multi-modal agent tool that unifies MCP/API tools, CLI execution, and visual GUI control in a modular runtime, enabling agents to coordinate computer use across heterogeneous interfaces while simplifying how users and agents exchange information.
    • At Syll’s core is a bidirectional user-agent interaction layer: users teach procedures through direct demonstration, which Syll compiles into reusable skills; agent executions are converted back into multi-modal evidence (logs, keyframes, and approval checkpoints) for inspection and control.
  • EN Highlights:
    • arXiv:2606.07594v1 Announce Type: new
    • Abstract: Personal AI agents must increasingly operate across APIs, shells, web surfaces, and desktop GUIs, yet many systems remain tuned to a single interface…

• We present Syll, an open-source, self-hosted multimodal agent harness that unifies MCP/API tools, CLI execution, and visual GUI control in a modular runtime, en…

• At the core of Syll is a bidirectional user-agent interaction layer: users teach procedures through direct demonstration, which Syll compiles into reusable skil…

• A case study of evaluating AI agents on a neuroscience data-to-discovery pipeline

  • Publication Time: 2026-06-09 12:00 Beijing Time
  • Abstract: - arXiv:2606.07718v1 Announcement Type: New.
    • Abstract: Agentic AI tools offer a promising path to automating software development bottlenecks in scientific research pipelines, particularly for stages that require days to months for domain experts to build, where scientists are concerned with correctness and robustness rather than implementation details.
    • We present an empirical study of general-purpose coding agents on a fly optogenetics data-to-discovery pipeline.
    • We evaluate agents on tasks substantially larger than existing benchmarks, with datasets orders of magnitude larger, and evaluation criteria based on domain expert standards.
  • EN Highlights:
    • arXiv:2606.07718v1 Announce Type: new
    • Abstract: Agentic AI tools offer a promising path to automating software development bottlenecks in scientific research pipelines, particularly for stages that…
    • We present an empirical study of general-purpose coding agents on a fly optogenetics data-to-discovery pipeline
    • We assess agents on tasks substantially larger than existing benchmarks, datasets orders of magnitude bigger, and evaluation criteria grounded in domain expert…

• Why Limit the Residual Stream to Layers and Not Tokens? Persistent Memory for Continuous Latent Reasoning

  • Publication Time: 2026-06-09 12:00 Beijing Time
  • Abstract: - arXiv:2606.07720v1 Announcement Type: New.
    • Abstract: Large language models (LLMs) have demonstrated remarkable reasoning abilities on mathematical and multi-hop planning tasks.
    • The CoCoNuT (Continuous Chain of Thought) paradigm~\cite{hao2024coconut} extends this by enabling the model to reason in latent space while exploring multiple reasoning paths, rather than committing to a single chain early on.
    • However, we have discovered a limitation, which we call the \textbf{conceptual bottleneck}.
  • EN Highlights:
    • arXiv:2606.07720v1 Announce Type: new
    • Abstract: Large language models (LLMs) have demonstrated remarkable reasoning abilities on mathematical and multi-hop planning tasks
    • The CoCoNuT (Chain of Continuous Thought) paradigm~\cite{hao2024coconut} extends this by enabling models to reason in latent space, exploring multiple reasoning…

• However, we identify a limitation we term the \textbf{concept bottleneck}

• Automatic Extraction of Structured Information from Brain MRI Reports Using an Open-Weight Large Language Model

  • Posted: 2026-06-09 12:00 Beijing Time
  • Abstract: - arXiv:2606.07721v1 Announcement Type: New.
    • Abstract: Objective: Automatically extracting data from free-text radiology reports can enable large-scale research, but few studies have evaluated the performance of Large Language Models (LLMs) on Dutch neuroradiology reports.
    • Methods: We analyzed 947 brain MRI reports from a tertiary memory clinic (2016-2021), written by consultant neuroradiologists.
    • Trained medical students annotated thirty variables; 100 reports were dual-annotated to assess inter-rater reliability.
  • EN Highlights:
    • arXiv:2606.07721v1 Announce Type: new
    • Abstract: Objectives: Automatic data extraction from free-text radiology reports enables large-scale research, but few studies assessed the performance of large…
    • Methods: We analyzed 947 brain MRI reports from a tertiary memory clinic (2016-2021), authored by consultant neuroradiologists
    • Trained medical students annotated thirty variables; 100 reports were double-annotated to assess inter-rater reliability

• Some hypotheses on how chatbots work in problem-solving-driven conversations. Large Language Models as confirmation of the Innovation Illusion

  • Posted: 2026-06-09 12:00 Beijing Time
  • Abstract: - arXiv:2606.07722v1 Announcement Type: New.
    • Abstract: This article provides a perspective on the nature of chatbots as true conversational partners when discussing problems related to their solutions.
    • What can chatbots do, what can’t they do, and how can this be explained?
    • Our argument draws on aggregation dynamics, cognitive linguistics, neuropsychology, and psychology.
  • EN Highlights:
    • arXiv:2606.07722v1 Announce Type: new
    • Abstract: This article offers a perspective on the nature of chatbots as genuine conversation partners when discussing problems in relation to their solutions
    • What can chatbots do and what can’t they do, and how can this be explained
    • Our argument draws on Aggregation Dynamics, Cognitive Linguistics, Neuropsychology and Psychology

• Land cover and flood type govern the detection limits of satellite-based flood mapping across diverse global flood events

  • Posted: 2026-06-09 12:00 Beijing Time
  • Abstract: - arXiv:2606.07780v1 Announcement Type: New.

• Abstract: Floods are one of the most destructive natural disasters. Climate change has led to an increasing frequency of floods, making satellite-based flood mapping crucial for disaster response.

• Geospatial foundation models pre-trained on satellite archives offer geographic transferability, but their operational reliability across various unseen events remains uncharacterized.

• Here, we deploy Prithvi-EO-2.0 across 19 out-of-distribution flood events (2017-2025) spanning six continents, eight climate zones, and six flood mechanisms, and validate it against two independent reference products.

  • EN Highlights:
    • arXiv:2606.07780v1 Announce Type: new
    • Abstract: Floods are among the most destructive natural hazards, and their increasing frequency under climate change makes satellite-based inundation mapping es…
    • Geospatial foundation models pretrained on satellite archives offer geographic transferability, but their operational reliability across diverse, unseen events…
    • Here we deploy Prithvi-EO-2.0 across 19 out-of-distribution flood events (2017-2025) spanning six continents, eight climate zones, and six flood mechanisms, val…

• Reconstructing and forecasting disease trajectories of patients with Alzheimer’s disease using routine data in resource-constrained settings

  • Release Time: 2026-06-09 12:00 Beijing Time
  • Abstract: - arXiv:2606.07798v1 Announce Type: new.
    • Abstract: Alzheimer’s disease is a progressive neurodegenerative disorder, and its progression varies substantially across patients.
    • Existing work aims to forecast patients’ future cognitive states, with minimal focus on reconstructing the state from past visits.
    • Furthermore, in current research, quantifying predictive uncertainty remains underexplored and relies on costly modalities such as MRI, PET, and CSF, limiting their deployment in resource-limited environments.
  • EN Highlights:
    • arXiv:2606.07798v1 Announce Type: new
    • Abstract: Alzheimer’s disease is a progressive neurodegenerative disorder, and its progression varies substantially across patients
    • Existing work aims to forecast patients’ future cognitive state, with minimal focus on reconstructing the state from past visits
    • Furthermore, in current research, quantifying predictive uncertainty remains underexplored and relies on costly modalities such as MRI, PET, and CSF, limiting t…

• Improving Multimodal Reasoning via Worst Dimension Optimization

  • Release Time: 2026-06-09 12:00 Beijing Time
  • Abstract: - arXiv:2606.07801v1 Announce Type: new.
    • Abstract: Multimodal reasoning requires a path that maintains integrity under various constraints, from visual grounding to logical consistency.
    • However, current process reward models focus on heuristically defined rewards that weigh these factors equally, which can lead to dominant factors masking failures in individual dimensions without guaranteeing the overall effectiveness of the reasoning process.

• arXiv:2606.07801v1 Announce Type: new Abstract: Multimodal reasoning requires a path that retains integrity under a wide range of constraints, from visual grounding to logical consistency. However, current Process Reward Models focus on heuristically defined rewards that weigh these factors equally, which may lead to the concealment of individ….

  • EN Highlights:
    • arXiv:2606.07801v1 Announce Type: new
    • Abstract: Multimodal reasoning requires a path that retains integrity over a wide range of constraints, from visual grounding to logic consistency
    • However, the current Process Reward Models focus on heuristically defined rewards that equally weigh these factors, which may lead to the concealment of individ…

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

• Bidirectional Small-Granularity Search between Code and Text

  • Release Time: 2026-06-09 12:00 Beijing Time
  • Abstract: - arXiv:2606.07519v1 Announce Type: new.
    • Abstract: We introduce the novel task of bidirectional small-granularity search between code and text, where the query is a small snippet of text or code, and the result is also a small snippet of the opposite modality (i.e., code or text).
    • This task establishes direct links between text in scientific publications and corresponding code segments to support a better and faster understanding of scientific methods.
    • We introduce a large dataset for the proposed task, which includes a training partition with textual descriptions of code automatically generated using GPT-4, and three test partitions—one in-domain and two out-of-domain (OOD)—containing manually annotated data and materials from other domains.
  • EN Highlights:
    • arXiv:2606.07519v1 Announce Type: new
    • Abstract: We introduce the novel task of bidirectional small-granularity search between code and text, where the queries are small snippets of text or code and…
    • This task establishes direct links between text in scientific publications and corresponding code segments, in support of better and faster understanding of sci…
    • We introduce a large dataset for the proposed task that includes a training partition with textual descriptions of code generated automatically using GPT-4, and…

• TinyJudge: Unverifiable Constraint Alignment via Lightweight Specialist Ensembles

  • Release Time: 2026-06-09 12:00 Beijing Time
  • Abstract: - arXiv:2606.07520v1 Announce Type: new.
    • Abstract: Instruction Following (IF) is a core capability of LLMs, requiring strict adherence to various constraints, from verifiable ones (e.g., output length) to unverifiable ones (e.g., tone).
    • Reinforcement learning with verifiable rewards has become a paradigm for IF tasks, utilizing an LLM as a judge to evaluate unverifiable constraints.
    • However, we empirically find that this approach remains a significant bottleneck, suffering from severe reward hacking and higher computational overhead.
  • EN Highlights:
    • arXiv:2606.07520v1 Announce Type: new

• Abstract: Instruction Following (IF) is a core capability of LLMs, requiring strict adherence to diverse constraints, ranging from verifiable ones (e.g., output…

• Reinforcement learning with verifiable rewards has emerged as a paradigm for IF tasks, leveraging LLM-as-a-judge to assess unverifiable constraints

• However, we empirically find that this approach remains a significant bottleneck, suffering from severe reward hacking and higher computational overhead

• Evaluating Hallucinations in Domain-Adapted Large Language Models

  • Publication Time: 2026-06-09 12:00 Beijing Time
  • Abstract: - arXiv:2606.07521v1 Announcement Type: New.
    • Abstract: This study investigates the phenomenon of hallucinations in domain-adapted Large Language Models (LLMs), focusing on the fine-tuning of the Llama-2 model using the Lamini dataset.
    • Hallucinations, or the generation of nonsensical or unfaithful content by LLMs, pose a significant challenge, especially when these models are fine-tuned with domain-specific data.
    • Our methodology involves a series of experiments testing the memory, recall, and reasoning abilities of the fine-tuned LLM, comparing its performance on novel question-answering pairs and domain-specific information.
  • EN Highlights:
    • arXiv:2606.07521v1 Announce Type: new
    • Abstract: This study investigates the phenomenon of hallucinations in domain-adapted Large Language Models (LLMs), focusing on the fine-tuning of the Llama-2 mo…
    • Hallucinations, or the generation of nonsensical or unfaithful content by LLMs, pose a significant challenge, especially when these models are fine-tuned with d…
    • Our methodology involves a series of experiments testing memorization, recall, and reasoning capabilities of the fine-tuned LLM, comparing its performance on no…

• Community-Specific Slang and Entity Detection via Semantic Shift in Fine-Tuned Language Models

  • Publication Time: 2026-06-09 12:00 Beijing Time
  • Abstract: - arXiv:2606.07522v1 Announcement Type: New.
    • Abstract: We propose an unsupervised method for parsing slang, unique entities, and folklore in online communities by isolating words in the lexicon with the highest degree of semantic shift.
    • Semantic shift is defined as the evolution of a word’s encoded representation, resulting from fine-tuning a pre-trained Large Language Model (LLM) on a community-specific text corpus.
    • This value is inversely proportional to the cosine similarity between the word’s encoded representation from the base model and its encoded representation from the fine-tuned model.
  • EN Highlights:
    • arXiv:2606.07522v1 Announce Type: new
    • Abstract: We propose an unsupervised method of resolving slang, unique entities, and folklore from online communities by isolating words in the lexicon that hav…

• Semantic shift is defined as the evolution of a word’s encoded representation as a result of fine-tuning a pretrained Large Language Model (LLM) on a community-…

• This value is inversely proportional to the cosine similarity between the base model’s encoded representation of a word, and a fine-tuned model’s encoded repres…

• Retrieval Augmented Generation Framework for the Nepali Legal Domain Question Answering

  • Published time:2026-06-09 12:00 Beijing Time
  • Abstract:- arXiv:2606.07523v1 Announcement Type: new.
    • Abstract: Legal domains in high-resource languages like English have widely adopted artificial intelligence for legal question answering.
    • However, data scarcity in low-resource languages such as Nepali has limited the training of large language models on Nepali legal texts.
    • This study presents the first application of a Retrieval Augmented Generation based model for Nepali legal question answering, using case laws extracted from the Nepali Kanun Patrika digital archives.
  • EN Highlights:
    • arXiv:2606.07523v1 Announce Type: new
    • Abstract: Legal domains in high-resource languages like English have widely adopted artificial intelligence for legal question answering
    • However, data scarcity in low resource languages such as Nepali has limited the training of large language models on Nepali legal texts
    • This study presents the first application of a Retrieval Augmented Generation based model for Nepali legal question answering using case laws extracted from the…

• ABLE: Representing and Mapping LLMs via Attribution-Based Large-model Embedding

  • Published time:2026-06-09 12:00 Beijing Time
  • Abstract:- arXiv:2606.07524v1 Announcement Type: new.
    • Abstract: The explosive growth of Large Language Models (LLMs) has created a heterogeneous and poorly documented ecosystem, making systematic model comparison increasingly important for provenance auditing, security analysis, and model selection.
    • Existing representation methods struggle to efficiently address this problem.
    • While methods that analyze internal parameters are powerful when architectures are compatible, they face scalability barriers under structural heterogeneity, and methods relying on external outputs may conflate models with similar behavior and struggle to align across richer output spaces between different tokenizers.
  • EN Highlights:
    • arXiv:2606.07524v1 Announce Type: new
    • Abstract: The explosive growth of large language models (LLMs) has created a heterogeneous and poorly documented ecosystem, making systematic model comparison i…
    • Existing representation methods struggle to address this setting efficiently

• Approaches analyzing internal parameters are powerful when architectures are compatible, but face scalability barriers under structural heterogeneity, while met…

• Implicit Causal Graph Construction in Text via Chain Discovery

  • Published: 2026-06-09 12:00 Beijing Time
  • Abstract: - arXiv:2606.07525v1 Announcement Type: New.
    • Abstract: Causal graphs in text are typically populated by observable, predefined events.
    • In contrast, we study implicit causal graph construction from text by treating each described cause-effect pair as the beginning and end point of a potential underlying causal graph, and using large language models (LLMs) to infer intermediate causal events.
    • We compare end-to-end graph construction with methods that frame the task as causal chain discovery.
  • EN Key Points:
    • arXiv:2606.07525v1 Announce Type: new
    • Abstract: Causal graphs in text are typically populated by observable, predefined events
    • In contrast, we study implicit causal graph construction from text by treating each described cause-effect pair as the begin- and endpoint of an underlying late…
    • We compare end-to-end graph construction with methods that frame the task as causal chain discovery

• GraphLoRA: Structure-Aware Low-Rank Adaptation for Large Language Model Recommendation

  • Published: 2026-06-09 12:00 Beijing Time
  • Abstract: - arXiv:2606.07526v1 Announcement Type: New.
    • Abstract: Large Language Models (LLMs) have shown strong potential for recommendation (LLMRec) due to their powerful reasoning and generalization abilities.
    • However, effectively aligning the textual semantics modeled by LLMs with collaborative signals remains a key challenge.
    • Existing methods either translate collaborative information into textual prompts or inject pre-trained embeddings into the LLM, both of which treat structural information as static input and cannot capture high-order relational dependencies.
  • EN Key Points:
    • arXiv:2606.07526v1 Announce Type: new
    • Abstract: Large Language Models (LLMs) have shown strong potential for recommendation (LLMRec) due to their powerful reasoning and generalization abilities
    • However, effectively aligning the textual semantics modeled by LLMs with the collaborative signals remains a key challenge
    • Existing methods either translate collaborative information into textual prompts or inject pre-trained embeddings into the LLM, both of which treat structural i…

• Post-training is (Massive) Supervised Learning

  • Published: 2026-06-09 12:00 Beijing Time
  • Abstract: - arXiv:2606.07527v1 Announcement Type: New.
    • Abstract: The mainstream paradigm for LLM training has evolved to rely on a large-scale post-training stage composed of SFT and RL.

• In this position paper, we argue that this methodology effectively marks a reversion to the “pre-train then fine-tune” approach of the BERT era, explicitly tailoring models to desired behaviors and the specific benchmarks on which they are evaluated.

• We begin with a historical overview of LLMs, describing the different phases of the LLM evolution.

  • EN Highlights:
    • arXiv:2606.07527v1 Announce Type: new
    • Abstract: The prevailing paradigm for training LLMs has evolved to rely on a massive post-training phase consisting of SFT and RL
    • In this position paper, we argue that this methodology effectively marks a reversion to the ``pre-train then fine-tune’’ approach of the BERT era, explicitly ta…
    • We begin with a historical overview of LLMs, describing the different phases of the LLM evolution

• BEACON: Behavioral Entropy Aggregation for Cross-Model Hallucination Detection in Large Language Models

  • Publication Time: 2026-06-09 12:00 Beijing Time
  • Abstract: - arXiv:2606.07528v1 Announce Type: new.
    • Abstract: Hallucination in large language models (LLMs), defined as the generation of factually incorrect or unsupported content, remains a critical barrier to reliable deployment.
    • We present BEACON (Behavioral Entropy Aggregation for Cross-model hallucinatiON), a black-box hallucination detection framework that operates purely on model outputs, without requiring access to internal representations or external knowledge bases.
    • BEACON extracts a 31-dimensional feature vector from structured multi-pass generation, integrating NLI-based semantic entropy, embedding geometry, chain-of-thought consistency, and paraphrase stability signals.
  • EN Highlights:
    • arXiv:2606.07528v1 Announce Type: new
    • Abstract: Hallucination in large language models (LLMs), defined as the generation of factually incorrect or unsupported content, remains a critical barrier to…
    • We present BEACON (Behavioral Entropy Aggregation for Cross-model hallucination detectiON), a black-box hallucination detection framework that operates purely o…
    • BEACON extracts a 31-dimensional feature vector from structured multi-pass generation, integrating NLI-based semantic entropy, embedding geometry, chain-of-thou…

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

• Offline Reinforcement Learning for Plasma Control in Nuclear Fusion: Codebase and Benchmark

  • Publication Time: 2026-06-09 12:00 Beijing Time
  • Abstract: - arXiv:2606.07550v1 Announce Type: new.
    • Abstract: Offline reinforcement learning (RL) offers a promising avenue for developing plasma controllers from historical tokamak data, as online trial-and-error on real devices is costly and risky.
    • However, progress in this direction remains difficult to measure due to the lack of a standardized offline RL benchmark for the practical multi-actuator, long-horizon plasma control problem in nuclear fusion.
    • We introduce RL4F, an offline RL benchmark for plasma control in nuclear fusion, providing a closed-loop evaluation environment and baseline comparisons for four full-profile tracking tasks: rotation, density, temperature, and pressure.
  • EN Highlights:

• arXiv:2606.07550v1 Announce Type: new

• Abstract: Offline reinforcement learning (RL) offers a promising route for developing plasma controllers from historical tokamak data, since online trial-and-er…

• However, progress in this direction remains difficult to measure due to the lack of a standardized offline RL benchmark for realistic multi-actuator, long-horiz…

• We introduce RL4F, an Offline Reinforcement Learning Benchmark for Plasma Control in Nuclear Fusion, providing closed-loop evaluation environments and baseline…

• MedicalRec: Medical recommender system for image classification without retraining

  • Publication Time: 2026-06-09 12:00 Beijing Time
  • Abstract: - arXiv:2606.07553v1 Announce Type: new.
    • Abstract: The emergence of machine learning and deep learning has revolutionized the efficiency of diagnostic, therapeutic, and administrative systems in healthcare.
    • However, this rapid adoption has come at the cost of requiring significant computing power and energy consumption, as well as e-waste disposal and carbon emissions.
    • One of the challenges of these models is choosing the right model for classification tasks.
  • EN Key Points:
    • arXiv:2606.07553v1 Announce Type: new
    • Abstract: The emergence of machine learning and deep learning has revolutionized the efficiency of diagnostic, therapeutic, and administrative systems in health…
    • However, this rapid adoption has come at the cost of requiring significant computing power and energy consumption, as well as e-waste disposal and carbon emissi…
    • One of the challenges of these models is choosing the right model for classification tasks

• SPIN: Decentralized Swarm Control via Tensorized Policy Coordination

  • Publication Time: 2026-06-09 12:00 Beijing Time
  • Abstract: - arXiv:2606.07557v1 Announce Type: new.
    • Abstract: Decentralized multi-agent swarm coordination on resource-constrained edge platforms remains fundamentally bottlenecked by the exponential scaling of the joint action space and high-latency communication overhead.
    • This paper introduces the Swarm Policy Interference Network (SPIN) framework, an architectural paradigm that bypasses these limitations by modeling the swarm topology as a compressed tensor network.
    • We decompose the joint policy tensor of local multi-agent factions into Matrix Product State (MPS) chains, reducing the computational complexity of evaluation from an exponential $O(n^m)$ barrier to a strictly linear $O(m \cdot n \cdot \chi^2)$ constraint.
  • EN Key Points:
    • arXiv:2606.07557v1 Announce Type: new
    • Abstract: Decentralized multi-agent swarm coordination on resource-constrained edge platforms remains fundamentally bottlenecked by the exponential scaling of j…

• This paper introduces the Swarm Policy Interference Network (SPIN) framework, an architectural paradigm that bypasses these limitations by modeling swarm topolo…

• We factorize the joint policy tensors of local multi-agent cliques into Matrix Product State (MPS) chains, reducing the computational complexity of evaluation f…

• Boundary Variance Inflation Causes Acquisition Bias in Gaussian Processes

  • Publication Time: 2026-06-09 12:00 Beijing Time
  • Abstract: - arXiv:2606.07561v1 Announcement Type: New.
    • Abstract: Gaussian processes with stationary kernels on bounded domains exhibit inflated posterior variance near the boundary.
    • Although boundary-induced acquisition bias is a long-recognized artifact in geostatistics and a source of over-exploration in Bayesian optimization, its causes and effects have not been fully studied.
    • We trace the root cause to a simple geometric mechanism: the truncation of the kernel’s correlation neighborhood at the domain boundary creates an observation-independent distortion that worsens with increasing dimensionality.
  • EN Highlights:
    • arXiv:2606.07561v1 Announce Type: new
    • Abstract: Gaussian processes with stationary kernels on bounded domains exhibit inflated posterior variance near the boundary
    • Despite being a long-recognized artifact in geostatistics and a source of over-exploration in Bayesian optimization, the causes and effects of boundary-induced…
    • We trace the root cause to a simple geometric mechanism: the truncation of the kernel correlation neighborhood at the domain boundary creates an observation-ind…

• Emergence via Phase Transitions: Mechanism Landscapes and Universal Convergence Across Complex Systems

  • Publication Time: 2026-06-09 12:00 Beijing Time
  • Abstract: - arXiv:2606.07563v1 Announcement Type: New.
    • Abstract: In machine learning, biology, and physics, independently evolving systems often converge toward strikingly similar high-level structures despite vastly different microscopic details.
    • Grokking circuits converge across random seeds, evolutionary lineages rediscover similar metabolic solutions, and renormalization flows approach common fixed points.
    • We propose the Hierarchical Emergence Framework (HEF) as a candidate universality framework for such convergent phenomena.
  • EN Highlights:
    • arXiv:2606.07563v1 Announce Type: new
    • Abstract: Across machine learning, biology, and physics, independently evolving systems often converge toward strikingly similar high-level structures despite r…
    • Grokking circuits converge across random seeds, evolutionary lineages rediscover similar metabolic solutions, and renormalization flows approach common fixed po…

• We propose the Hierarchical Emergence Framework (HEF) as a candidate universality framework for such convergence phenomena

• STARIXNet: Multivariate and Multi-attribute Deep Learning Approach to Real-Time Resource Allocation in Cloud Platforms

  • Publication Time: 2026-06-09 12:00 Beijing Time
  • Abstract: - arXiv:2606.07565v1 Announcement Type: New.
    • Abstract: Intelligent scaling of microservices in cloud platforms is crucial for mitigating escalating computational costs while avoiding service disruptions.
    • Current solutions are limited to the univariate space, typically focusing only on CPU usage to drive scaling decisions.
    • Moreover, they address the problem as a purely forecasting task, focusing on prediction precision while neglecting the greater risks of underestimation and system response delays.
  • EN Key Points:
    • arXiv:2606.07565v1 Announce Type: new
    • Abstract: Intelligent scaling of microservices in cloud platforms is crucial for mitigating escalating compute costs while avoiding service disruptions
    • Current solutions are limited to the univariate space, typically focusing on CPU usage alone to drive scaling decisions
    • Moreover, they address the problem as a purely forecasting task, focusing on prediction precision while neglecting the greater risks of underestimation and dela…

• TriHead-GAN: A Generative Adversarial Network with Triple-Head Discriminator for Carbon Emission Time Series Generation

  • Publication Time: 2026-06-09 12:00 Beijing Time
  • Abstract: - arXiv:2606.07569v1 Announcement Type: New.
    • Abstract: Accurate carbon emission monitoring is crucial for climate policy and emerging regulatory mechanisms such as the EU Carbon Border Adjustment Mechanism. However, city-level high-frequency monitoring data remains extremely scarce, severely limiting deep learning models that require large amounts of data.
    • Time series generation is a natural remedy, but existing GAN and diffusion-based generators often provide limited explicit supervision for the domain structure of carbon emission data: they may match marginal distribution statistics while failing to preserve the cross-variable correlations between CO$_2$, co-emitted pollutants, and meteorological factors. They also tend to corrupt the first-order difference statistics of atmospheric measurements, producing sequences that are smooth on average but lack the realistic step-by-step variations of the underlying signal.
    • We propose TriHead-GAN, a Transformer-based adversarial framework whose three-head discriminator jointly supervises three complementary aspects of the joint distribution: distributional realism via a Wasserstein critic, cross-variable dependencies via no-leak regression of the target variable, and step-wise temporal smoothness via adjacent difference prediction.
  • EN Key Points:
    • arXiv:2606.07569v1 Announce Type: new
    • Abstract: Accurate carbon emission monitoring is critical for climate policy and emerging regulatory mechanisms such as the EU Carbon Border Adjustment Mechanis…
    • Time series generation is a natural remedy, but existing GAN and diffusion-based generators often provide limited explicit supervision for the domain structure…

• We propose TriHead-GAN, a Transformer-based adversarial framework whose triple-head discriminator jointly supervises three complementary aspects of the joint di…

• Enabling KV Caching of Shared Prefix for Diffusion Language Models

  • Publication Time: 2026-06-09 12:00 Beijing Time
  • Abstract: - arXiv:2606.07571v1 Announcement Type: New.
    • Abstract: Key-value (KV) caching for shared prefixes is crucial for high-throughput large language model (LLM) serving, but it faces severe challenges in emerging diffusion language models (DLMs).
    • In DLMs, bidirectional attention means that updating any token dynamically alters the entire context and its corresponding KVs.
    • Therefore, existing caching techniques developed for LLMs, which assume KVs remain unchanged once computed, corrupt the shared prefix KVs.
  • EN Key Points:
    • arXiv:2606.07571v1 Announce Type: new
    • Abstract: Key-value (KV) caching for shared prefixes is essential for high-throughput large language model (LLM) serving, but it faces critical challenges in em…
    • In DLMs, bidirectional attention means that updating any token dynamically alters the entire context and its corresponding KVs
    • Thus, existing caching techniques developed for LLMs, which assume that KVs remain invariant once computed, corrupt the shared prefix KVs

• When Should an AI Scientist Stop? Verifiable Experiment Steering and Refusal for Autonomous Discovery

  • Publication Time: 2026-06-09 12:00 Beijing Time
  • Abstract: - arXiv:2606.07576v1 Announcement Type: New.
    • Abstract: We present CARTOGRAPH, a verification layer for AI scientists that combines unresolved subspace experiment steering (selection), explicit ambiguity closure (resolution), and residual-based library insufficiency detection (refusal).
    • Under a local linear-Gaussian bridge, the original unresolved projection is the isotropic unresolved Fisher information trace, while CARTOGRAPH-A is the exact unresolved A-optimal rule; closed-form EIG and Box-Hill emerge as local rather than global comparators.
    • On five test platforms, CARTOGRAPH-A defeated the original projection 129W/0T/15L at d = 8 (p < 10^-21) in replicated structured cascades.
  • EN Key Points:
    • arXiv:2606.07576v1 Announce Type: new
    • Abstract: We present CARTOGRAPH, a verification layer for AI scientists that couples unresolved-subspace experiment steering (select), explicit ambiguity closur…
    • Under a local linear-Gaussian bridge, raw unresolved projection is the isotropic unresolved Fisher-information trace, while CARTOGRAPH-A is the exact unresolved…

• Across five testbeds, CARTOGRAPH-A beats raw projection 129W/0T/15L at d = 8 (p < 10^-21) in a replicated structured cascade

• MST-Direct at Scale: Multivariate and Conditional Geostatistical Simulation via Sinkhorn Optimal Transport

  • Published: 2026-06-09 12:00 Beijing Time
  • Abstract: - arXiv:2606.07578v1 Announce Type: New.
    • Abstract: This paper extends MST-Direct, a Matching-via-Sinkhorn-Transport approach for multivariate geostatistical simulation, from the original bivariate, unconditional, small-grid formulation to multivariate, conditional, and large-grid settings.
    • We address the three main limitations identified in the original work: (i) scalability beyond a few thousand nodes through a sparse, candidate-restricted Sinkhorn matcher with O(nC) memory complexity; (ii) extension to multiple variables by matching target value tuples to an independent FFT-MA Gaussian backbone that reproduces the specified variogram; and (iii) hard data conditioning by fixing observed data tuples at their spatial locations while conditioning the backbone via kriging.
    • Because the transport plan remains a permutation of the target tuples, the multivariate joint distribution is preserved exactly.
  • EN Key Points:
    • arXiv:2606.07578v1 Announce Type: new
    • Abstract: This paper extends MST-Direct, a Matching-via-Sinkhorn-Transport approach for multivariate geostatistical simulation, from the original bivariate, unc…
    • We address the three main limitations identified in the original work: (i) scalability beyond a few thousand nodes through a sparse, candidate-restricted Sinkho…
    • Because the transport plan remains a permutation of the target tuples, the multivariate joint distribution is preserved exactly