System translated (Gemini)

🤖 AI 速览

Google DeepMind launched Gemini 3.8 Live and Live Avatar, pushing real-time multimodal interaction to the infrastructure layer; agents are moving from planning to collaboration and observability, with Bot Screen and Claude Code cloud conversation response transparency and scalability needs; …
📋 文章元数据
发布时间
2026-09-25
类型
ai-daily
字数
7134
阅读时长
34 min

2026-09-25 AI Daily | Gemini 3.8 Live Launches, with a New Focus on Real-time Multimodality and Agent Observability Link to heading

Google DeepMind launches Gemini 3.8 Live and Live Avatar, pushing real-time multimodal interaction to the infrastructure level. Agents are evolving from planning to collaboration and observability, with tools like Bot Screen and Claude Code’s cloud sessions addressing the need for transparency and scalability. Meanwhile, the negative results of using synthetic personas to predict real audience outcomes and the Bitget hacking incident serve as reminders that effectiveness and risk control validation are more critical than model capabilities alone.

📖 This Issue’s Watch List: A Deep Dive Link to heading

There are three main themes worth watching today. First, multimodality and real-time interaction: Google DeepMind’s launch of Gemini 3.8 Live and Live Avatar brings real-time conversation and interactive avatars into the infrastructure discussion; Ovis-Embedding uses a unified backbone to integrate text, images, video, and audio, and discussions around Claude Opus 5.5’s capabilities are also worth noting. Second, the shift of agents from planning to collaboration: X-Planner adds event structure planning for embodied intelligence, ReAdapt explores social relationship reasoning in LLMs, and COMED discusses the middle ground of multi-model routing and collaboration, making it a deep read for teams focused on agent architecture. Third, validation issues in vertical-specific scenarios: Research such as MedGate-Fusion, explainable GBM for survival prediction, and the negative results of using synthetic personas to predict real audience outcomes highlights that “effectiveness validation” is more crucial than pure model capability when implementing solutions in healthcare and marketing.

🌐 AI Hot Topics on X Link to heading

Topic 1: Nous Research Launches Bot Screen for Live AI Agent Monitoring Link to heading

  • Category: AI · News
  • Overview: Trending for: 1 day ago, Related posts: 842
  • What it is: Nous Research has launched Bot Screen, a tool for real-time monitoring of AI agent operational status.
  • Why it’s important: As the number of autonomous AI agents grows, real-time observability has become critical for secure deployment and debugging. This tool addresses the industry’s demand for agent transparency.
  • Discussion summary: Discussions on X are focused on whether the tool can effectively reduce the risk of AI agents going rogue, how it differs from existing monitoring solutions, and the impact of real-time monitoring on agent privacy and performance.

Topic 2: Bitget Crypto Exchange Hit by $351.6 Million Hot Wallet Breach Link to heading

  • Category: AI · News
  • Overview: Trending for: 3 hours ago, Related posts: 16,000
  • What it is: The crypto exchange Bitget suffered an attack on its hot wallet, resulting in a loss of approximately $351.6 million.
  • Why it’s important: This incident highlights the vulnerabilities of digital asset platforms in real-time risk control and anomaly detection. It may accelerate the adoption of AI-driven on-chain monitoring, automated threat response, and security audit technologies.
  • Discussion summary: The main topics of discussion on X are the attack methods and the platform’s responsibility. The focus is on whether user funds were affected, the platform’s ability to pay, and whether AI can proactively identify abnormal transactions or counter increasingly sophisticated attacks.

Topic 3: House Report Accuses DOJ of Secretly Targeting Republican Lawmakers Link to heading

  • Category: AI · Other
  • Overview: Trending for: , Related posts: 125
  • What it is: Republicans on the U.S. House Judiciary Committee released declassified files from the FBI’s “Arctic Frost” investigation, accusing the Department of Justice (DOJ) of secretly monitoring Republican lawmakers and allies of Donald Trump.
  • Why it’s important: The incident exposes large-scale data collection (financial, social media, and communications) by law enforcement agencies. This could intensify debates around AI-assisted surveillance, data privacy, and algorithmic accountability, potentially affecting the regulatory boundaries for government use of AI tools.
  • Discussion summary: The focus on X is whether the DOJ has been weaponized. Supporters claim the evidence shows the FBI overstepped its authority in collaboration with the J6 Committee to target conservatives. Opponents argue it is a standard accountability measure related to interference in the 2020 election and the Capitol riot.

Topic 4: Alexandr Wang’s Camo Crocs Outfit Sparks Luxury Memes at Meta Connect Link to heading

  • Category: AI · News
  • Overview: Trending for: 17 hours ago, Related posts: 4,700
  • What it is: Alexandr Wang’s outfit at Meta Connect, featuring camouflage Crocs paired with formal wear, sparked a wave of luxury-brand-style memes created by netizens.
  • Why it’s important: This event reflects how the personal image and public profile of star founders in the AI field can transcend technical discussions, becoming part of the industry’s cultural dissemination and brand identity.
  • Discussion summary: Discussions centered on joking about the clash between his mix-and-match style and Silicon Valley’s elite dress code. People also debated whether these lighthearted memes help bridge the gap between AI leaders and the public or distract from important technical topics.

Topic 5: CLM Model Speeds Up AI Agent Decisions with Contrastive Learning Link to heading

  • Category: AI · News
  • Overview: Trending for: 20 hours ago, Related posts: 1400
  • What it is: A model named CLM has significantly increased the decision-making speed of AI agents through contrastive learning.
  • Why it matters: This advancement is expected to reduce the inference latency of AI agents, improving efficiency in real-time interactions and automated tasks, and has positive implications for deployment in resource-constrained environments.
  • Discussion summary: Discussions on X are mainly focused on the practical advantages of this model compared to traditional reinforcement learning methods, its generalization ability and stability in complex environments, and whether contrastive learning is genuinely effective for decision-making tasks.

Topic 6: Anthropic Launches Cloud Sessions for Claude Code with Subscriber Credits Link to heading

  • Category: AI · News
  • Overview: Trending for: 1 day ago, Related posts: 8000
  • What it is: Anthropic has launched a cloud sessions feature for Claude Code, allowing subscribers to use their account credits to run coding agents in the cloud.
  • Why it matters: This marks the expansion of AI coding assistants from local command-line tools to managed cloud services, which could lower the barrier to entry for developers and promote the adoption of agent-based programming.
  • Discussion summary: Discussions on X are primarily focused on the credit consumption rate and cost, the difference in experience between cloud sessions and local execution, security and latency concerns, and its positioning relative to competitors like OpenAI Codex and Cursor.

Topic 7: AI Agent Trims 400,000 Lines of Redundant Tests in OpenClaw Project Link to heading

  • Category: AI · News
  • Overview: Trending for: 6 hours ago, Related posts: 367
  • What it is: An AI agent cleaned up approximately 400,000 lines of redundant test code in the OpenClaw project.
  • Why it matters: This demonstrates the capability of AI agents in large-scale code cleanup and automated maintenance, potentially reducing technical debt and improving development efficiency significantly.
  • Discussion summary: The discussion mainly revolves around the accuracy of the AI cleanup, whether it might mistakenly delete valid tests, and if such automated operations should undergo more rigorous manual review.

Topic 8: Paper Rex Beats Team Liquid 2-1 to Top VALORANT Champions Odds Link to heading

  • Category: AI · Sports
  • Overview: Trending for:, Related posts: 164
  • What it is: Paper Rex defeated Team Liquid 2-1, rising to become the top favorite in the VALORANT Champions betting odds.
  • Why it matters: This event has garnered attention as a topic at the intersection of AI and sports, reflecting how the influence and accuracy of AI-driven match predictions and betting models in esports are being put to a real-world test.
  • Discussion summary: Discussions on X mainly center on whether AI predictions are overestimating Paper Rex, the reasonableness of the odds changes, and the credibility of data models in esports forecasting.

Topic 9: Israeli Forward Scores Then Sent Off for Flag Celebration Link to heading

  • Category: AI · Other
  • Overview: Trending for: 3 hours ago, Related posts: 17000
  • What it is: An Israeli football forward was sent off with a red card by the referee for celebrating a goal by displaying a flag.
  • Why it matters: While the event itself is not directly related to AI technology, its automatic categorization under “AI · Other” and its high trending status reflect the challenges of automated content classification, politically sensitive content identification, and recommendation algorithms in scenarios where sports and politics intersect. This serves as a case study for AI content moderation.
  • Discussion summary: Discussions on X are mainly focused on whether the red card was justified, whether the flag constitutes a political statement, and the boundaries of political symbolism in sporting events. Users are divided on refereeing standards and the platform’s handling of content labeling.

Topic 10: Dagestani Fighters’ Relentless Discipline Powers UFC Dominance Link to heading

  • Category: AI · Other
  • Overview: Trending for: 8 hours ago, Related posts: 328
  • What it is: Dagestani fighters have achieved a dominant position in the UFC through their rigorous discipline.
  • Why it matters: Although this topic is from the realm of mixed martial arts, it has drawn attention from the AI community to the methodology of “discipline-driven iteration” due to the analogy between their training discipline and the principle of continuous optimization in reinforcement learning.
  • Discussion summary: Discussions on X are centered on whether the training system of Dagestani fighters can be analogized to the design of reward functions in AI, and whether extreme discipline from non-technical fields can offer insights for the long-term, robust training of AI models.

Topic 11: Van Dijk Commits to Netherlands as Gakpo’s Late Goal Earns Draw vs. Klopp’s Germany Link to heading

  • Category: AI · Other
  • Overview: Trending for: 1 day ago, Related posts: 57000
  • What it is: The Netherlands drew with Klopp’s Germany thanks to a last-minute goal from Gakpo, and Van Dijk has confirmed he will continue to play for the Dutch national team.
  • Why it matters: This topic is a sports hotspot with no direct technical or business impact on the AI field. Its inclusion in the “AI · Other” trend reflects the vague boundaries of social media topic categorization for non-AI content, which may interfere with the observation of AI trends.
  • Discussion summary: Discussions on X primarily revolved around Van Dijk’s future with the national team, Gakpo’s crucial goal, and Klopp’s tactical adjustments after becoming Germany’s coach. There was no significant discussion or disagreement on this topic within the AI community.

Topic 12: Lewis Hamilton Shares AI Video of F1 Drivers in Funny Car Ride Link to heading

  • Category: AI · Sports
  • Overview: Trending for: 8 hours ago, Related posts: 10000
  • What happened: Lewis Hamilton shared a humorous AI-generated video on X featuring F1 drivers in a car.
  • Why it matters: This event demonstrates the viral potential of generative AI in sports entertainment content and raises concerns about the authenticity and boundaries of AI-generated videos.
  • Discussion summary: The discussion focused on the video’s realism and entertainment value, as well as issues such as whether AI-synthesized content could mislead viewers and potential violations of drivers’ image rights.

Topic 13: Chile’s Golden Retriever Puppies Nap After Military Parade Link to heading

  • Category: AI · Entertainment
  • Overview: Trending for: , Related posts: 146
  • What happened: After a military parade in Chile, videos and photos of several Golden Retriever puppies napping on the roadside went viral on X and were categorized by the platform under the “AI · Entertainment” trend.
  • Why it matters: This topic reflects how AI recommendation systems automatically cluster and amplify entertaining pet content. It also sparked discussions about the difficulty of distinguishing between real animal videos and AI-generated content.
  • Discussion summary: Users on X primarily debated whether the puppy footage was real or AI-generated, and why the platform’s algorithm placed natural pet content into an AI category. Others focused on the impact of AI recommendation mechanisms on the spread of entertainment topics.

Topic 14: Birdman and Toni Braxton Spotted Together After Rocky Marriage Timeline Link to heading

  • Category: AI · Entertainment
  • Overview: Trending for: , Related posts: 48
  • What happened: Birdman and Toni Braxton were spotted together again after multiple breakups and reconciliations, a period the media has dubbed a “rocky marriage timeline.”
  • Why it matters: This topic’s entry into the AI · Entertainment trend list shows that social media AI recommendation and auto-classification systems are including high-engagement entertainment content under AI-related tags. This exposes issues with cross-domain content understanding and tagging accuracy, serving as a classic case for evaluating bias in AI content moderation and classification.
  • Discussion summary: Users on X mainly debated whether the two had gotten back together or secretly married, with fans reviewing their relationship timeline. Some also questioned the rationality of classifying this entertainment gossip as AI-related, discussing why AI tags failed to distinguish between entertainment and technology topics.

Topic 15: OnlyFans Faces Claims of Fake Subscribers and Hidden Motives Link to heading

  • Category: AI · Entertainment
  • Overview: Trending for: , Related posts: 50
  • What happened: OnlyFans has been accused of having a large number of fake subscribers and allegedly hiding the platform’s operational motives.
  • Why it matters: The incident highlights how generative AI and automated bots can be used to forge users and interactions, exacerbating the platform’s trust crisis and impacting the creator economy.
  • Discussion summary: Discussions on X focused on whether the fake accounts were AI-driven, whether the platform condoned or participated in this activity, and the potential impact on creators’ income and the content ecosystem.

Summary of AI Public Opinion on X Today Link to heading

Today’s main discussion revolved around “trust and boundaries” as AI agents and generative AI accelerate their implementation. The industry generally agrees that real-time observability, low-latency decision-making, cloud-based agent coding, and automated code cleanup are key to enhancing the security and efficiency of AI deployment. However, the Bitget hack and the OnlyFans fake subscriber incident reveal that AI can be used not only for risk control but also for attacks or faking interactions, leading to significant disagreement on whether risks can be effectively prevented. In terms of content governance, multiple sports and entertainment topics being algorithmically classified as AI, combined with the authenticity debates over the F1 parody video and cute pet videos, reflect the unreliability of recommendation and moderation systems in cross-domain understanding and identifying AI-generated content. Furthermore, the FBI surveillance files have intensified the political opposition regarding government use of AI-assisted monitoring and privacy boundaries. Potential risks are concentrated in autonomous agents losing control or making erroneous deletions, automated fraud in finance and the creator economy, misinformation from AI-synthesized content, and the potential for inaccurate platform algorithm classification to amplify governance chaos.

💡 Influencer Insights Link to heading

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

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

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

Stratechery by Ben Thompson (A_full) Link to heading

  • An Interview with Colossus EIC Jeremy Stern About Profiling Mark Zuckerberg
    • Publication Time: 2026-09-24 18:00 Beijing Time
    • Abstract: - An interview with Colossus EIC Jeremy Stern about profiling Mark Zuckerberg and other prominent tech figures.
      • $15/month* or *$150/year.
      • Substantive analysis of the day’s news through three weekly emails or a podcast.
      • Strategy interviews.
      • Interviews with leading public company CEOs, private company founders, and discussions with fellow analysts.
    • EN Highlights:
      • An interview with Colossus EIC Jeremy Stern about profiling Mark Zuckerberg and other prominent tech figures.

Google DeepMind Blog (A_full) Link to heading

  • Introducing Gemini 3.8 Live with Live Avatar
    • Publication Time: 2026-09-25 00:20 Beijing Time
    • Abstract: - Introducing Gemini 3.8 Live and Live Avatar.
      • This article from the Google DeepMind Blog explains how the introduction of Gemini 3.8 Live with Live Avatar is shaping the broader AI and infrastructure landscape.
      • It also lays out the practical implications for founders, operators, and investors following the launch of Gemini 3.8 Live with Live Avatar.
    • EN Highlights:
      • Introducing Gemini 3.8 Live with Live Avatar

Two Minute Papers (B_intro+search) Link to heading

  • Claude Opus 5.5 AI: An Incredible Leap Forward
    • Publication Time: 2026-09-24 16:40 Beijing Time
    • Abstract: - ❤️ Check out Lambda here and sign up for their GPU Cloud:.
      • Note: In the walking creatures experiment, Astra used a simplified model and was unable to implement the correct one.
      • Things did not improve after simulating it for more generations.
      • The walking creatures video and paper:.
    • EN Highlights:
      • ❤️ Check out Lambda here and sign up for their GPU Cloud:
      • Note: in the walking creatures experiment, Astra used a simplified model and was unable to implement the correct one
      • Things did not improve after simulating it for more generations
      • Claude Opus 5.5:

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

  • Do Synthetic Personas Predict Real Audience Response? A Sim-to-Real Study Where a No-Persona Baseline Beats Persona-Based Copy Simulation

    • Publication Time: 2026-09-24 12:00 Beijing Time
    • Abstract: - arXiv:2609.25010v1 Announcement Type: new.
      • Abstract: Marketers are increasingly using large language models (LLMs) as “synthetic personas” to predict audience response to a piece of copy before its release, encouraged by evidence of LLMs mimicking human samples on profile-based conditioning.
      • But does this prediction actually work for real behavior? Does the persona mechanism help?
      • We present a sim-to-real validity study using the Upworthy Research Archive – thousands of headline A/B tests on live traffic with click-through rates measured – as held-out ground truth.
    • EN Highlights:
      • arXiv:2609.25010v1 Announce Type: new

  • Abstract: Marketers increasingly use large language models (LLMs) as “synthetic personas” to predict how an audience will react to a piece of copy before it shi…

  • But is that prediction actually valid against real behaviour - and does the persona machinery help

  • We present a sim-to-real validity study using the Upworthy Research Archive - thousands of headline A/B tests on shared real traffic, with measured click-throug…

  • Do Existing Preconditioners Improve Biomedical Tabular Foundation Learning? An Empirical Study on TabPFN Optimization

    • Publication Time: 2026-09-24 12:00 Beijing Time
    • Abstract: - arXiv:2609.25013v1 Announce Type: new.
      • Abstract: Tabular foundation models have recently shown strong potential for structured biomedical data analysis.
      • Among them, TabPFN has emerged as an effective approach for low-data tabular classification tasks.
      • However, the impact of optimization and preconditioning strategies on biomedical fine-tuning remains largely unexplored.
    • EN 要点:
      • arXiv:2609.25013v1 Announce Type: new
      • Abstract: Tabular foundation models have recently shown strong potential for structured biomedical data analysis
      • Among them, TabPFN has emerged as an effective approach for low-data tabular classification tasks
      • However, the impact of optimization and preconditioning strategies on biomedical fine-tuning remains largely unexplored

  • 4DGS-JEPA: Temporally Compositional Joint-Embedding Prediction for Dynamic Gaussian Splatting

    • Publication Time: 2026-09-24 12:00 Beijing Time
    • Abstract: - arXiv:2609.25036v1 Announce Type: new.
      • Abstract: Dynamic Gaussian Splatting provides an explicit representation of evolving 3D scenes, but existing approaches are primarily optimized for reconstruction, future state generation, or rendering, rather than for learning reusable predictive dynamics.
      • We propose 4DGS-JEPA, a Gaussian-native joint-embedding predictive architecture for causal multi-horizon prediction over dynamic Gaussian scenes.
      • The model uses hierarchical scene, motion group, and Gaussian-level representations along with horizon-conditioned transformation operators that support direct prediction and recursive rollout.
    • EN 要点:
      • arXiv:2609.25036v1 Announce Type: new
      • Abstract: Dynamic Gaussian Splatting provides an explicit representation of evolving 3D scenes, but existing approaches are primarily optimized for reconstructi…
      • We propose 4DGS-JEPA, a Gaussian-native joint-embedding predictive architecture for causal multi-horizon prediction over dynamic Gaussian scenes

  • The model uses a hierarchical scene-, motion-group-, and Gaussian-level representation together with a horizon-conditioned transition operator that supports bot…

  • An Accurate and Interpretable Hyper Graph Neural Network for GBM Survival Prediction

    • Published: 2026-09-24 12:00 Beijing Time
    • Abstract: - arXiv:2609.25088v1 Announcement Type: new.
      • Abstract: Survival prediction for glioblastoma multiforme (GBM) demands models that are both accurate and interpretable, yet existing approaches treat these objectives as competing, where high-performance models sacrifice transparency, and interpretable models accept diminished predictive power.
      • We argue that this trade-off is not inherent.
      • Graph neural networks offer a structural foundation for extracting interpretable, explainable representations without compromising discriminative ability.
    • EN Key Points:
      • arXiv:2609.25088v1 Announce Type: new
      • Abstract: Survival prediction for glioblastoma multiforme (GBM) demands models that are both accurate and interpretable, yet existing approaches treat these obj…
      • We argue that this trade-off is not inherent
      • Graph neural net- works offer a structural foundation for extracting interpretable, explainable representations without compromising discriminative ability

  • Ovis-Embedding: Pushing the Frontiers of Universal Omni-Modal Embeddings

    • Published: 2026-09-24 12:00 Beijing Time
    • Abstract: - arXiv:2609.25165v1 Announcement Type: new.
      • Abstract: In this report, we introduce \textbf{Ovis-Embedding}, a state-of-the-art omni-modal embedding series built on the native integration of text, image, video and audio.
      • Instead of assembling separate modality towers, Ovis-Embedding uses a shared multimodal backbone to encode different modalities in a common representation space.
      • Specifically, we have made \textbf{three key advances}: (1) \textbf{Native omni-modal initialization}: We adopt a pre-trained Qwen-omni model as the embedding backbone and adapt it through contrastive training with low-rank initialization; (2) \textbf{Data-centric omni-modal training}: We built an extensive, high-quality corpus covering text, images, video, audio, and interleaved multimodal data.
    • EN Key Points:
      • arXiv:2609.25165v1 Announce Type: new
      • Abstract: In this report, we introduce \textbf{Ovis-Embedding}, a state-of-the-art omni-modal embedding family built on native integration of text, image, video…
      • Instead of assembling separate modality towers, Ovis-Embedding uses a shared multimodal backbone to encode different modalities in a common representation space
      • Specifically, we make \textbf{three key advances}: (1) \textbf{native omni-modal initialization}: we adopt a pretrained Qwen-omni model as the embedding backbon…

  • X-Planner: Event-Structured Task Planning for Embodied Intelligence

    • Publication Time: 2026-09-24 12:00 Beijing Time
    • Abstract: - arXiv:2609.25187v1 Announce Type: new.
      • Abstract: Task planning connects high-level instructions and executable behaviors in long-horizon operations, but modern Vision-Language-Action (VLA) systems often implicitly convey this intermediate structure.
      • Existing Chain-of-Thought (CoT) planners also tend to rely on coarse task-level annotations or serialize long reasoning traces token by token.
      • We introduce X-Planner, a planning front-end that addresses the supervision and representation problems of embodied reasoning.
    • EN Highlights:
      • arXiv:2609.25187v1 Announce Type: new
      • Abstract: Task planning bridges high-level instructions and executable behavior in long-horizon manipulation, yet modern Vision-Language-Action (VLA) systems of…
      • Existing chain-of-thought (CoT) planners also tend to rely on coarse task-level annotations or serialize long reasoning traces token by token
      • We present X-Planner, a planning front-end that addresses both the supervision and representation of embodied reasoning

  • Lean Pool: An AI-Maintained Archive of Formalized Mathematics

    • Publication Time: 2026-09-24 12:00 Beijing Time
    • Abstract: - arXiv:2609.25199v1 Announce Type: new.
      • Abstract: Lean Pool is a repository of formalized mathematics.
      • It is grown, maintained, and optimized by AI agents.
      • arXiv:2609.25199v1 Announce Type: new Abstract: Lean Pool is a repository of formalized mathematics, grown, maintained, and optimized by AI agents.
    • EN Highlights:
      • arXiv:2609.25199v1 Announce Type: new
      • Abstract: Lean Pool is a repository of formalized mathematics
      • It is grown, maintained and optimized by AI agents.

  • The AI Neuroscientist: An Interactive Agentic Interface for Neuroimaging Analysis

    • Publication Time: 2026-09-24 12:00 Beijing Time
    • Abstract: - arXiv:2609.25254v1 Announce Type: new.
      • Abstract: Analyzing neuroimaging data requires specialized coding and statistical expertise, which limits accessibility for researchers without a computational background.
      • We present the AI Neuroscientist, a language agent for interactive data exploration.
      • The system integrates large language models (LLMs) with neuroimaging toolkits to perform quality control, modeling, and visualization.
    • EN Highlights:
      • arXiv:2609.25254v1 Announce Type: new
      • Abstract: Analyzing neuroimaging data requires specialized coding and statistical expertise, which limits accessibility for researchers without computational ba…
      • We present the AI Neuroscientist, a language agent for interactive data exploration

  • The system integrates a large language model (LLM) with a neuroimaging toolset to perform quality control, modeling, and visualization

  • MedGate-Fusion: Integrating First-Encounter Semantic Narratives and Physiological Biomarkers for Prospective Stroke Risk Stratification

    • Publication Time: 2026-09-24 12:00 Beijing Time
    • Abstract: - arXiv:2609.25272v1 Announcement Type: New.
      • Abstract: Prospective stroke risk stratification in primary care is challenging because early risk signals are distributed across routine biomarkers and unstructured clinical narratives.
      • We propose MedGate-Fusion, a multi-modal gated architecture that integrates transformer-based embeddings of first-encounter narratives with ten routinely recorded risk markers.
      • We used electronic medical record data from the Canadian Primary Care Sentinel Surveillance Network (CPCSSN).
    • EN Highlights:
      • arXiv:2609.25272v1 Announce Type: new
      • Abstract: Prospective stroke risk stratification in primary care is challenging because early risk signals are distributed across routine biomarkers and unstruc…
      • We propose MedGate-Fusion, a multi-modal gated architecture that integrates transformer-based embeddings of first-encounter narratives with ten routinely record…
      • We used electronic medical record data from the Canadian Primary Care Sentinel Surveillance Network (CPCSSN)

  • When LLM Agents Fail to Read the Room: ReAdapt for Relational Social Reasoning

    • Publication Time: 2026-09-24 12:00 Beijing Time
    • Abstract: - arXiv:2609.25284v1 Announcement Type: New.
      • Abstract: A social agent’s most basic decisions (Should I react to this post?
      • Who should I contact?) are not purely a matter of content.
      • The correct action often depends on the underlying relationships between people—strength of connection, reciprocity, interconnectedness—rather than which content is most prominent.
    • EN Highlights:
      • arXiv:2609.25284v1 Announce Type: new
      • Abstract: A social agent’s most basic decisions (should I react to this post
      • who should I reach out to
      • ) are not purely content problems

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

  • COMED: The Missing Middle Between Routing and Collaboration in Multi-LLM Inference

    • Publication Time: 2026-09-24 12:00 Beijing Time
    • Abstract: - arXiv:2609.26913v1 Announcement Type: New.
      • Abstract: No single large language model (LLM) is consistently reliable across all queries, which has motivated multi-model inference systems to either route between models or combine their outputs.
      • However, routing stops after selecting the initial model, while dense collaboration invokes peers on every query.
      • We demonstrate that collaboration is non-monotonic: peers can recover from failures that a model would have made alone, but they can also corrupt an initially correct answer.
    • EN Highlights:
      • arXiv:2609.26913v1 Announce Type: new

  • Abstract: No single Large Language Model (LLM) is uniformly reliable across queries, motivating multi-model inference systems that either route among models or…

    • However, routing stops after selecting an initial model, while dense collaboration invokes peers on every query
    • We show that collaboration is non-monotonic: peers can recover failures that no model solves alone, but can also corrupt initially correct answers

  • Experts Rise Where LLMs Disagree: Using Cross-Model Disagreement to Target Expert Effort in LLM Codebook Revision for Large-Scale Annotation

    • Publication Time: 2026-09-24 12:00 Beijing Time
    • Abstract: - arXiv:2609.26926v1 Announcement Type: new.
      • Abstract: Large-scale text annotation often brings expert insights to millions of documents through a codebook followed by AI annotators.
      • However, developing a robust codebook takes several months.
      • Large Language Models (LLMs) can accelerate this process by applying an early codebook to the data, surfacing cases with strong LLM disagreement, and soliciting expert feedback to resolve them.
    • EN Highlights:
      • arXiv:2609.26926v1 Announce Type: new
      • Abstract: Large-scale text annotation brings expert insight to millions of documents, often through a codebook that AI annotators follow
      • Developing a robust codebook, however, takes months
      • Large language models (LLMs) could speed this process by applying an early codebook to the data, surfacing cases with strong LLM disagreement, and eliciting exp…

  • Recognized but Not Produced: A Generation Benchmark for Culturally Specific Kinship Terms

    • Publication Time: 2026-09-24 12:00 Beijing Time
    • Abstract: - arXiv:2609.26942v1 Announcement Type: new.
      • Abstract: Current literature evaluates Large Language Models (LLMs) on multilingual kinship understanding using multiple-choice benchmarks, treating it as a recognition problem.
      • In contrast, we prompt five open-weight LLMs to generate kinship terms in two communicative tasks across three non-Western languages (Hindi, Tamil, and Korean), and pair this with a matched-options-supported choice baseline.
      • On the same relational-linguistic units, GPT OSS120B selected the correct term in 90.67% of 75 valid units but generated an acceptable term in 36.00% of corresponding attempts; Llama 3.370B showed the same pattern (77.92% vs 24.24%).
    • EN Highlights:
      • arXiv:2609.26942v1 Announce Type: new
      • Abstract: Current literature evaluates large language models (LLMs) on multilingual kinship understanding using multiple choice benchmarks, treating it as a rec…

  • We instead prompt five open weight LLMs to generate kinship terms in three non Western languages (Hindi, Tamil, and Korean) across two communicative tasks and p…

  • On identical relation language cells, GPT OSS120B selects the correct term in 90.67% of 75 valid cells but produces an accepted term in 36.00% of the correspond…

  • Classifying Interpretive Canons at the Sentence Level: A Benchmark from the German Federal Constitutional Court

    • Posted: 2026-09-24 12:00 Beijing Time
    • Abstract: - arXiv:2609.26945v1 Announcement Type: new.
      • Abstract: Judicial reasoning remains challenging for large language models (LLMs) to analyze.
      • This paper provides a sentence-level benchmark for evaluating the ability of LLMs to classify interpretive canons, as articulated by Larenz in the tradition of Savigny.
      • Our contributions are threefold.
    • EN Key Points:
      • arXiv:2609.26945v1 Announce Type: new
      • Abstract: Judicial reasoning remains challenging for large language models (LLMs) to analyze
      • This paper contributes a sentence-level benchmark for evaluating the ability of LLMs to classify interpretive canons as articulated by Larenz in the tradition o…
      • Our contributions are threefold

  • When Learned Context Planning Fails to Beat Strong Retrieval: A Controlled Study of Planning, Routing, and Reranking for Long-Context QA

    • Posted: 2026-09-24 12:00 Beijing Time
    • Abstract: - arXiv:2609.26976v1 Announcement Type: new.
      • Abstract: Learned context planning selects evidence atoms before an answer model reasons over them.
      • We test whether this learned selection improves long-context multiple-choice QA after strong retrieval, routing, budgeted-selector, and reranking controls.
      • Our primary diagnostic uses all 503 LongBench-v2 MCQ questions with Qwen2.5-7B-Instruct.
    • EN Key Points:
      • arXiv:2609.26976v1 Announce Type: new
      • Abstract: Learned context planning selects evidence atoms before an answer model reasons over them
      • We test whether this learned selection improves long-context multiple-choice QA after strong retrieval, routing, budgeted-selector, and reranking controls
      • Our primary diagnostic uses all 503 LongBench-v2 MCQ questions with Qwen2.5-7B-Instruct

  • LEGO: Synergizing Expert GraphRAG and Expert Chain-of-Thought for Legal Reasoning

    • Posted: 2026-09-24 12:00 Beijing Time

  • Abstract: - arXiv:2609.27009v1 Announce Type: new.

    • Abstract: Large language models are increasingly applied to high-risk domains such as law, yet complex legal reasoning remains limited by two structural challenges.
    • First, existing RAG and GraphRAG methods emphasize lexical or semantic similarity while overlooking normative relations among legal provisions.
    • Second, vanilla Chain-of-Thought prompting may generate plausible rationales without enforcing the normative structure of legal reasoning.

  • LexLattice: Multilingual Extractive Summarization via Neural Cellular Automata on Document Hierarchies

    • Release Time: 2026-09-24 12:00 Beijing Time
    • Abstract: - arXiv:2609.27032v1 Announce Type: new.
      • Abstract: Faithfulness is a central concern in legal text summarization, which motivates extractive approaches that select verbatim content traceable to its source.
      • Such methods typically rank paragraphs or other structural units in isolation, yet give little attention to consolidating evidence that is distributed across, and shares significance between, the document’s more distant parts.
      • We introduce LexLattice, an extractive summarizer that reifies a legal act’s hierarchy as a two-dimensional semantic lattice and consolidates over it with a masked two-dimensional neural cellular automaton before selection.

  • EduBehaviors: Assertion-based Schemas for Auditable Coding of Educational Dialogues

    • Release Time: 2026-09-24 12:00 Beijing Time
    • Abstract: - arXiv:2609.27043v1 Announce Type: new.
      • Abstract: Large language models allow for the rapid deployment of instructional annotations corresponding to structures of interest, allowing natural language interfaces to generate classifications on conversational datasets.
      • However, due to the opacity of LLM reasoning, we are unable to have verifiable, mechanistic insight into the reasons a model chooses a label for an utterance.

  • We introduce the EduBehaviors framework, an interpretable, scalable approach to annotating educational data that uses LLMs to measure repeated observable behaviors related to many constructs of interest, and then learns classifiers for those constructs based on these observable behaviors.

    • EN Highlights:
      • arXiv:2609.27043v1 Announce Type: new
      • Abstract: Large language models have allowed the rapid deployment of pedagogical annotations corresponding to constructs of interest, allowing a natural languag…
      • However due to the opaque nature of LLM reasoning, we have no verifiable, mechanistic insight into why a model chose a label for an utterance
      • We introduce the EduBehaviors framework, an interpretable, scalable approach to annotating educational data that uses LLMs to measure repeated observable behavi…

  • The Illinois Social Attitudes Aggregate Corpus (ISAAC): An Open Tool and Reproducible Pipeline for Analyzing Social Group Discourse at Scale

    • Published: 2026-09-24 12:00 Beijing Time
    • Abstract: - arXiv:2609.27059v1 Announce Type: new.
      • Abstract: We introduce the Illinois Social Attitudes Aggregate Corpus (ISAAC), an open, modular, and accessible corpus of over 527 million English Reddit posts related to six key social group characteristics based on factors such as race, gender, age, ability, weight, and skin color, covering a 17-year period from 2007 to 2023.
      • A multi-step, human-audited filtering pipeline was used to keep irrelevant content in the curated dataset below 10%, both overall and for each distinct social group.
      • Each post was then algorithmically annotated with the user’s estimated home region, along with a suite of validated off-the-shelf and custom semantic labels, including morality, emotion, sentiment, and linguistic generalizations.
    • EN Highlights:
      • arXiv:2609.27059v1 Announce Type: new
      • Abstract: We introduce the Illinois Social Attitudes Aggregate Corpus (ISAAC), an open, modular, and accessible corpus of 527 million+ English-language Reddit p…
      • A multi-step, human-audited filtering pipeline was used to keep irrelevant content in the curated dataset below 10%, both overall and for each social group dist…
      • Each post was then algorithmically annotated with the user’s estimated home region, along with a suite of validated off-the-shelf and custom semantic labels inc…

  • What Changes When Fact-Verification Scores Improve? Evidence and Answer Accounting Across Trained Verifiers and LLMs

    • Published: 2026-09-24 12:00 Beijing Time
    • Abstract: - arXiv:2609.27064v1 Announce Type: new.
      • Abstract: Joint fact-verification scores evaluate answers and submitted evidence together.
      • When scores improve, if the answer remains unchanged, how much gain is still left?

  • On FEVEROUS, the strict score is the percentage of claims with a correct answer and a complete annotated evidence group in the submitted evidence.

    • EN Highlights:
      • arXiv:2609.27064v1 Announce Type: new
      • Abstract: A joint fact-verification score assesses answers and submitted evidence together
      • When the score improves, how much of the gain remains if the answers are held fixed
      • On FEVEROUS, strict score is the percentage of claims with a correct answer and a complete annotated evidence group in the submitted evidence

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

  • The Drift Contract: Spectral Updates for Depth-Robust Local Learning

    • Publication Time: 2026-09-24 12:00 Beijing Time
    • Abstract: - arXiv:2609.26811v1 Announce Type: new.
      • Abstract: Local learning trains each layer with its own auxiliary loss and no global backward pass, which makes layer updates structurally parallel.
      • Two problems have kept it marginal: accuracy degrades as depth grows, and hyperparameters are fragile.
      • We apply Muon-style spectral update geometry (momentum orthogonalization with spectral step scaling) to per-layer local updates, an intersection not previously studied.
    • EN Highlights:
      • arXiv:2609.26811v1 Announce Type: new
      • Abstract: Local learning trains each layer with its own auxiliary loss and no global backward pass, which makes layer updates structurally parallel
      • Two problems have kept it marginal: accuracy degrades as depth grows, and hyperparameters are fragile
      • We apply Muon-style spectral update geometry (momentum orthogonalization with spectral step scaling) to per-layer local updates, an intersection not previously…

  • Signal2Symbol: Neuro-Symbolic Temporal Reasoning for Explainable Physiological Time-Series Anomaly Detection

    • Publication Time: 2026-09-24 12:00 Beijing Time
    • Abstract: - arXiv:2609.26820v1 Announce Type: new.
      • Abstract: Physiological time series such as electrocardiograms (ECG) and electroencephalograms (EEG) exhibit complex temporal structure, substantial acquisition variability, and a strong need for transparent decision-making.
      • Although deep models can achieve high detection performance, they often provide only limited insight into why a segment is anomalous, how local anomalies evolve over time, and whether a detection is part of a broader, recurring pattern.
      • We propose Signal2Symbol, a neuro-symbolic framework for explainable biosignal anomaly detection.
    • EN Highlights:
      • arXiv:2609.26820v1 Announce Type: new
      • Abstract: Physiological time series such as electrocardiograms (ECG) and electroencephalograms (EEG) exhibit complex temporal structure, substantial acquisition…

  • Although deep models can achieve high detection performance, they often provide limited insight into why a segment is anomalous, how local anomalies relate over…

    • We propose Signal2Symbol, a neuro-symbolic framework for explainable biosignal anomaly detection

  • HARN: Hierarchical Associative Resonance Network for Event-Driven Multi-Timeframe Forecasting

    • Published: 2026-09-24 12:00 Beijing Time
    • Abstract: - arXiv:2609.26822v1 Announcement Type: new.
      • Abstract: Financial time series evolve across multiple temporal resolutions, posing a challenge for forecasting systems to incorporate newly available information without repetitive recalculation of unchanged representations.
      • We introduce HARN, a Hierarchical Associative Resonance Network for event-driven multi-timeframe forecasting.
      • HARN maintains persistent representations across temporal levels and updates each level only when its corresponding complete bar becomes available.
    • EN Highlights:
      • arXiv:2609.26822v1 Announce Type: new
      • Abstract: Financial time series evolve across multiple temporal resolutions, challenging forecasting systems to incorporate newly available information without…
      • We introduce HARN, a Hierarchical Associative Resonance Network for event-driven multi-timeframe forecasting
      • HARN maintains persistent representations across temporal levels and updates each level only when its corresponding completed bar becomes available

  • What Makes a Terminal-Bench Task Hard? Separating Genuine Hardness from Fake-Hardness on an Adjudicated Agentic Corpus

    • Published: 2026-09-24 12:00 Beijing Time
    • Abstract: - arXiv:2609.26826v1 Announcement Type: new.
      • Abstract: Frontier benchmarks require tasks that current models cannot solve.
      • However, a task that no model can solve is not necessarily a hard task.
      • The same zero pass rate can stem from an actual capability gap, but it can also come from missing context, a broken reference solution, infrastructure failure, or a bypassable validator.
    • EN Highlights:
      • arXiv:2609.26826v1 Announce Type: new
      • Abstract: Frontier benchmarks need tasks that current models cannot solve
      • But a task that no model solves is not automatically a hard task
      • The same zero pass rate can come from a real capability gap, but it can also come from missing context, a broken reference solution, infrastructure failure, or…

  • LWCal: Loss-Weighted Calibration for Tabular Classifiers with Noisy Calibration Labels

    • Published: 2026-09-24 12:00 Beijing Time
    • Abstract: - arXiv:2609.26839v1 Announcement Type: new.

  • Abstract: Post-hoc probability calibration is usually evaluated under an optimistic assumption: the held-out calibration labels are clean.

  • In many AI deployment settings, however, labels come from weak annotators, historical decisions, heuristics, or distant supervision, so the same label noise that corrupts training also corrupts calibration.

  • We study this overlooked failure mode for tabular classifiers and propose LWCal, a CPU-only post-hoc calibrator that down-weights calibration examples whose noisy labels contradict the base model’s held-out probabilities.

  • EN Highlights:

    • arXiv:2609.26839v1 Announce Type: new
    • Abstract: Post-hoc probability calibration is usually evaluated under an optimistic assumption: the held-out calibration labels are clean
    • In many AI deployment settings, however, labels come from weak annotators, historical decisions, heuristics, or distant supervision, so the same label noise tha…
    • We study this overlooked failure mode for tabular classifiers and propose LWCal, a CPU-only post-hoc calibrator that down-weights calibration examples whose noi…

  • A Leakage-Aware Multimodal Evaluation Framework for Early Intraoperative Acute Kidney Injury Prediction

    • Publication Time: 2026-09-24 12:00 Beijing Time
    • Abstract: - arXiv:2609.26848v1 Announce Type: new.
      • Abstract: The incidence of postoperative acute kidney injury (AKI) after major non-cardiac surgery is high, but early intraoperative risk stratification remains difficult.
      • In this retrospective cohort study, we propose SynerT, a waveform-only hybrid temporal backbone that combines a causal dilated TCN with a hierarchy of dilated recurrent layers to encode early intraoperative physiological trajectories for AKI risk prediction.
      • Building on SynerT, we further design two model variants that extend the backbone with structured clinical context: SynerT-MM, a late-fusion multimodal extension integrating hemodynamic load summaries and preoperative covariates; and SynerTStack, a leakage-safe stacked ensemble that combines SynerT-MM’s cross-validated predictions with a strong tabular baseline in a meta-learning phase.
    • EN Highlights:
      • arXiv:2609.26848v1 Announce Type: new
      • Abstract: Postoperative acute kidney injury (AKI) after major non-cardiac surgery carries substantial morbidity, yet early intraoperative risk stratification re…
      • In this retrospective cohort study, we propose SynerT, a waveform-only hybrid temporal backbone that combines a causal dilated TCN with a hierarchy of dilated r…
      • Building on SynerT, we further design two model variants that extend the backbone with structured clinical context: SynerT-MM, a late-fusion multimodal extensio…

  • COPE: Continual Personalization of LLMs under Sparse User Feedback via User Embeddings and Self-Evaluation

    • Publication Time: 2026-09-24 12:00 Beijing Time
    • Abstract: - arXiv:2609.26853v1 Announce Type: new.

  • Abstract: While Large Language Models (LLMs) have achieved remarkable results across various benchmarks, their alignment with normative values often leads to homogenized responses that fail to satisfy diverse user preferences.

  • Existing training-free methods often occupy valuable context windows through prompt engineering, while training-based methods typically remain static post-training, unable to support the continuous optimization required in real-world environments.

  • To address these challenges, we propose COPE (Continual Optimization with Personalized embedding and self-Evaluation), a novel optimization framework tailored for real-world-driven interactive settings with sparse user feedback.

  • EN Highlights:

    • arXiv:2609.26853v1 Announce Type: new
    • Abstract: While Large Language Models (LLMs) have achieved remarkable results across various benchmarks, their alignment with normative values often results in…
    • Existing training-free methods often occupy valuable context windows through prompt engineering, while training-based methods typically remain static post-train…
    • To address these challenges, we propose COPE (Continual Optimization with Personalized embedding and self-Evaluation), a novel optimization framework tailored f…

  • QUARTET: Quad-branch cross-Attention and Random-walk Traces for Enhancing Transformers on Relational Graphs

    • Publication Time: 2026-09-24 12:00 Beijing Time
    • Abstract: - arXiv:2609.26855v1 Announce Type: new.
      • Abstract: Relational Deep Learning (RDL) models multi-table databases as heterogeneous temporal graphs, and graph transformers currently achieve state-of-the-art performance on benchmarks like RelBench.
      • However, the current leading model, RelGT, suffers from two key limitations: its random local sampler yields loosely connected subgraphs that hinder message passing, and its global attention module relies on a single, seed-feature-based memory, ignoring broader macro-dynamics.
      • To overcome these limitations, we introduce QUARTET, an expressive graph transformer architecture that applies full self-attention on local subgraphs while enriching the global context via cross-attention branches.
    • EN Highlights:
      • arXiv:2609.26855v1 Announce Type: new
      • Abstract: Relational Deep Learning (RDL) models multi-table databases as heterogeneous temporal graphs, and graph transformers currently achieve state-of-the-ar…
      • However, the current leading model, RelGT, suffers from two key limitations: its random local sampler yields loosely connected subgraphs that hinder message pas…
      • To overcome these limitations, we introduce QUARTET, an expressive graph transformer architecture that applies full self-attention on local subgraphs while enri…

  • Marginally Correct Tool Caches Can Reverse Group-Normalized Policy Updates

    • Publication Time: 2026-09-24 12:00 Beijing Time
    • Abstract: - arXiv:2609.26866v1 Announce Type: new.

  • Abstract: Tool-result caching reduces repeated execution in agent training, but also couples rollout randomness.

  • We study a two-action model in which independent and shared execution preserve every rollout’s conditional reward distribution.

  • Despite this marginal agreement, sharing one stochastic result per group can reverse the expected group-normalized policy update.

    • EN Key Points:
      • arXiv:2609.26866v1 Announce Type: new
      • Abstract: Tool-result caching reduces repeated execution in agent training, but also couples rollout randomness
      • We study a two-action model in which independent and shared execution preserve every rollout’s conditional reward distribution
      • Despite this marginal agreement, sharing one stochastic result per group can reverse the expected group-normalized policy update

  • PR-Smoother: Simulator-Preserving Non-Gaussian Smoothing for Data Assimilation

    • Published: 2026-09-24 12:00 Beijing Time
    • Abstract: - arXiv:2609.26890v1 Announce Type: new. -Abstract: Many physical data assimilation (DA) workflows require smoothing methods that represent non-Gaussian posteriors over physical state variables, scale to high-dimensional simulators, train only from an observation window, and remain compatible with the calibration of a prescribed simulator.
      • We introduce PR-Smoother, a simulator-preserving amortized smoother designed for this prescribed-simulator DA regime.
      • Its key design principle is to keep the prescribed simulator explicit in both the evidence lower bound and the variational family: rather than learning replacement dynamics or trajectories, PR-Smoother only learns future conditional corrections around a prescribed rollout.
    • EN Key Points:
      • arXiv:2609.26890v1 Announce Type: new
      • Abstract: Many physical data assimilation (DA) workflows require smoothing methods that represent non-Gaussian posteriors over physical state variables, scale t…
      • We introduce PR-Smoother, a simulator-preserving amortized smoother designed for this prescribed-simulator DA regime
      • Its key design principle is to keep the prescribed simulator explicit in both the evidence lower bound and the variational family: rather than learning replacem…