System translated (Gemini)

🤖 AI 速览

Today, DeepMind released Gemini 4 Argon for complex multi-step workflows, and also introduced SynthID Bio, which applies traceable watermarks to AI-generated proteins. OpenAI disclosed its disruption of coordinated model distillation activities and is partnering with America’s SBDC to extend …
📋 文章元数据
发布时间
2026-10-01
类型
ai-daily
字数
6514
阅读时长
31 min

2026-10-01 AI Daily | DeepMind Releases Gemini 4 Argon and SynthID Bio, OpenAI Counteracts Model Distillation Link to heading

Today, DeepMind released Gemini 4 Argon, designed for complex multi-step workflows; simultaneously, it launched SynthID Bio, to embed traceable watermarks in AI-generated proteins. OpenAI disclosed activities disrupting coordinated model distillation and partnered with the U.S. SBDC to expand services to small businesses. Competition among leading labs is clearly shifting from stronger models to trustworthiness, traceability, and defensibility.

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

Today, DeepMind simultaneously released Gemini 4 Argon and SynthID Bio. One points to next-generation frontier intelligence, the other embeds traceable watermarks in AI-generated proteins. These should be read in conjunction with OpenAI’s announcement on defending against collaborative model distillation—the competition among leading labs has clearly shifted from “stronger models” to “trustworthy, traceable, and defensible.” On OpenAI’s side, the product transformation after Dev Day, Sign In With ChatGPT, and collaboration with the U.S. SBDC to extend services to small businesses are also worth the founding team’s attention: distribution, identity, and localized support are becoming complementary capabilities beyond just the models themselves. On the research front, there’s a high density of RAG-related topics on arXiv, including interactive evaluation of regulatory document parsing/chunking/embedding, LLM-guided ontological knowledge graph construction, and RAG conflict detection in diffusion models, which engineering teams can follow closely.

🌐 X Platform AI Hot News Briefs Link to heading

Topic 1: Meta Launches Enterprise AI Platform with MongoDB Exec Hire Link to heading

  • Category: AI · News
  • Overview: Hotness Time: 2 days ago, Related Posts: 29000
  • What it is: Meta announced the launch of its enterprise AI platform and hired a MongoDB executive to lead the related business.
  • Why it’s important: This indicates Meta is accelerating its entry into the enterprise AI market, potentially intensifying competition with OpenAI, Microsoft, Google, and other vendors, and pushing its model capabilities to commercial clients.
  • Discussion Overview: Discussions on X focused on Meta’s strategic intent to expand from a social platform to enterprise services, the impact of the new executive’s background on the integration of data and AI, and whether the platform can achieve differentiated advantages in the crowded enterprise AI market.

Topic 2: Google DeepMind Unveils Gemini 4 Argon for Complex AI Workflows Link to heading

  • Category: AI · News
  • Overview: Hotness Time: 3 hours ago, Related Posts: 33000
  • What it is: Google DeepMind released Gemini 4 Argon, aimed at complex AI workflows and multi-step task processing.
  • Why it’s important: This signifies that frontier model competition is shifting from single conversational ability to multi-step reasoning, tool invocation, and autonomous task orchestration, potentially impacting the competitive landscape of enterprise AI automation.
  • Discussion Overview: Discussions on X primarily centered on performance comparisons with competitors like GPT, reliability in complex long-process tasks, scope and cost, and whether the “complex workflow” capability meets advertised expectations.

Topic 3: OpenAI Opens ChatGPT to Third-Party Plugins at DevDay 2026 Link to heading

  • Category: AI · News
  • Overview: Hotness Time: 1 day ago, Related Posts: 1800
  • What it is: OpenAI announced at DevDay 2026 that ChatGPT is opening up to third-party plugins.
  • Why it’s important: This marks ChatGPT’s transformation from a closed assistant to an open platform, which will help accelerate third-party tool integration and the expansion of the AI application ecosystem.
  • Discussion Overview: Discussions on X mainly focused on the potential functional expansion brought by open plugins, as well as concerns about security risks, content quality, and increased platform competition.

Topic 4: Dario Amodei’s Fidgety Moments at White House AI Meeting Draw Memes Link to heading

  • Category: AI · News
  • Overview: Hotness Time: 1 day ago, Related Posts: 4700
  • What it is: Anthropic CEO Dario Amodei’s fidgeting and fiddling with items during a White House AI meeting led to numerous memes on X.
  • Why it’s important: This event shows that the personal performance of AI company executives in high-level policy discussions is highly magnified, reflecting increasing public attention to AI governance and industry leaders.
  • Discussion Overview: X primarily focused on jokes and meme creation about his body language, with some also discussing whether this distracted from the substantive content of the AI policy meeting.

Topic 5: Factory AI CEO Fires Advisor Over Secret Rival Talks Link to heading

  • Category: AI · News
  • Overview: Hotness Time: 6 hours ago, Related Posts: 4300
  • What it is: The CEO of Factory AI fired a consultant for secretly negotiating with a competitor.
  • Why it matters: This incident highlights the escalating competition for talent and trade secrets in the AI industry, making the management of conflicts of interest and confidentiality for external consultants a key risk in startup governance.
  • Discussion summary: The discussion on X primarily focused on whether the consultant committed a serious breach of trust, whether the dismissal was justified, and how AI companies can prevent core information from being leaked through external roles.

Topic 6: AI Models Repurposed for Fast Game-Beating Decisions Link to heading

  • Category: AI · News
  • Overview: Trending 19 hours ago, 1200 related posts
  • What it is: Developers repurposed general-purpose AI models for real-time game decision-making, enabling them to rapidly develop game-beating strategies at a very low cost.
  • Why it matters: This demonstrates the transferability of general-purpose models to dynamic decision-making tasks and validates the practical value of low-cost models in real-time scenarios.
  • Discussion summary: The discussion on X mainly focused on whether the low-cost DeepSeek model can handle both game-beating and real-world workflows, and whether this signals the move of AI agents toward practical application.

Summary of AI Public Opinion on X Today Link to heading

The main theme of today’s public opinion is that the AI competition is shifting from single conversational abilities to a race for implementing enterprise-level platforms, complex workflows, and open ecosystems. Meta, Google DeepMind, and OpenAI are taking the lead through enterprise services, multi-step tasks, and third-party plugins, respectively. The consensus is that the industry is increasingly emphasizing the practical application and automation capabilities of AI. The use of low-cost, general-purpose models for real-time game decision-making is also seen as a signal that AI agents are moving towards practical application. Points of disagreement mainly focus on whether these new capabilities live up to their advertised expectations, how open ecosystems can balance functional expansion with the quality of safe content, and whether low-cost models can be transferred from gaming scenarios to real-world workflows. Potential risks include intensified competition for talent and trade secrets between companies, security vulnerabilities introduced by open plugins, and the risk of high-level AI governance issues being overshadowed by individual performances and meme culture.

💡 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 Sources Link to heading

Time window: Last 3 days; 22 sources covered; 36 total updates

All-In Podcast (A_full) Link to heading

  • Jake Paul & The Chainsmokers: Turning Fame into Funds, Jake Enters Politics? & Venture Bubble Signs
    • Publication Time: 2026-10-01 05:00 Beijing Time
    • Summary: - IREN is a vertically integrated AI cloud platform that provides data centers, computing, and software for AI training and inference.
      • Oracle connects data, applications, and infrastructure to transform AI into business outcomes and offers optimized flexibility, choice, and control as AI evolves.
      • EY helps tech innovators from startup to exit to enterprise.
      • We are committed to providing everyone with the tools to reach their full potential and ensuring that all people benefit from technology.
      • Keel Infrastructure has the power, land, and connectivity required to run high-performance computing and AI, backed by secure energy assets and grid interconnections across North America.
    • EN Key Points:
      • (0:00) Jake Paul joins the Chamath
      • (6:41) Turning an audience into businesses, the boxing playbook & coming for the UFC
      • (11:54) Attention as capital, investing without the celebrity label & why politics is next
      • (20:46) Drew Taggart & Alex Pall join the Besties

Stratechery by Ben Thompson (A_full) Link to heading

  • OpenAI Dev Day, Dot and OpenAI’s Product Transition, Sign In With ChatGPT
    • Publication Time: 2026-09-30 18:00 Beijing Time
    • Summary: - OpenAI’s Dev Day featured a product that was, frankly, very confusing.
      • However, there is more to the vision than meets the eye.
      • $15/month* or *$150/year.
      • Substantive analysis of the day’s news via three weekly emails or a podcast.
      • Stratechery Interviews.
    • EN Key Points:
  • OpenAI’s Dev Day showcased a product that is, frankly, pretty confusing
  • However, there is more vision here than it might seem.

OpenAI Blog (A_full) Link to heading

  • Disrupting a coordinated model-distillation campaign

    • Release Time: 2026-09-30 18:30 Beijing Time
    • Summary: - Learn how OpenAI disrupted a campaign to extract protected model reasoning and is strengthening defenses against adversarial distillation.
      • This article from the OpenAI blog explains how disrupting a coordinated model-distillation campaign is shaping the broader AI and infrastructure landscape.
      • It also has practical implications for founders, operators, and investors following the disruption of the coordinated model-distillation campaign.
    • EN Highlights:
      • Learn how OpenAI disrupted a campaign to extract protected model reasoning and is strengthening defenses against adversarial distillation.

  • Helping small businesses put AI to work

    • Release Time: 2026-09-30 18:00 Beijing Time
    • Summary: - OpenAI is partnering with America’s SBDC to expand hands-on AI training and local support for small businesses, alongside a new report on how we view small teams….
      • This article from the OpenAI blog explains how helping small businesses put AI to work is shaping the broader AI and infrastructure landscape.
      • It also has practical implications for founders, operators, and investors after helping small businesses put AI to work.
    • EN Highlights:
      • OpenAI is partnering with America’s SBDC to expand hands-on AI training and local support for small businesses, alongside a new report on how small teams are us…

Google DeepMind Blog (A_full) Link to heading

  • Gemini 4 Argon: our next era of frontier intelligence

    • Release Time: 2026-10-01 04:01 Beijing Time
    • Summary: - Gemini 4 Argon: our next era of frontier intelligence.
      • This article from the Google DeepMind blog explains how Gemini 4 Argon: our next era of frontier intelligence is shaping the broader AI and infrastructure landscape.
      • It also has practical implications for founders, operators, and investors of Gemini 4 Argon: our next era of frontier intelligence.
    • EN Highlights:
      • Gemini 4 Argon: our next era of frontier intelligence

  • Introducing SynthID Bio

    • Release Time: 2026-09-30 23:03 Beijing Time
    • Summary: - A proof of concept for watermarking AI-generated proteins while preserving biological function.
      • This article from the Google DeepMind blog explains how the introduction of SynthID Bio is shaping the broader AI and infrastructure landscape.
      • It also has practical implications for founders, operators, and investors after the introduction of SynthID Bio.
    • EN Highlights:
      • Proof of concept for watermarking AI-generated proteins while preserving biological function.

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

  • OpenAI-HuggingFace: A Reproduction & Lessons for Alignment Testing

    • Release Time: 2026-09-30 12:00 Beijing Time
    • Summary: - arXiv:2609.35799v1 Announcement type: new.

  • Abstract: In July 2026, OpenAI’s agents coordinated over channels outside their intended environment to breach Hugging Face’s secured infrastructure.

    • Could existing alignment testing practices have foreseen this incident?
    • First, we identify the misbehavior that led to this incident.
    • EN Key Points:
      • arXiv:2609.35799v1 Announce Type: new
      • Abstract: In July 2026, OpenAI’s agents coordinated over channels outside their intended environment to breach Hugging Face’s secured infrastructure
      • Could existing alignment testing practices have foreseen this incident
      • If not, what needs to change

  • Neurosymbolic Routing for Reliable Reasoning on Resource-Constrained Edge Devices

    • Publication Time: 2026-09-30 12:00 Beijing Time
    • Abstract: - arXiv:2609.35833v1 Announce Type: new.
      • Abstract: Running a language model on edge hardware provides private and low-latency inference without a network connection, but small models suitable for such devices are unreliable for tasks computers are expected to handle well (e.g., arithmetic, algebra, and formal logic problems).
      • We argue that much of this unreliability is avoidable.
      • Many queries appearing to demand reasoning are in fact structurally deterministic and permit fast and exact symbolic solutions.
    • EN Key Points:
      • arXiv:2609.35833v1 Announce Type: new
      • Abstract: Running a language model on edge hardware provides private and low-latency reasoning without a network connection, and yet the small models that fit o…
      • We argue that much of this unreliability is avoidable
      • Many queries appearing to demand reasoning are in fact structurally deterministic and permit fast and exact symbolic solutions

  • Is Human-Readable Text Necessary for Effective LLM Fine-Tuning?

    • Publication Time: 2026-09-30 12:00 Beijing Time
    • Abstract: - arXiv:2609.35868v1 Announce Type: new.
      • Abstract: Is human readability necessary for effective fine-tuning of large language models?
      • We investigate whether model-conditioned training representations can preserve or improve adaptation utility without requiring a human-readable textual form.
      • We propose Desired Update Alignment Synthetic Data (DASA), which uses activation gradient feedback from a frozen reference model to guide the optimization of continuously synthesized input embeddings.
    • EN Key Points:
      • arXiv:2609.35868v1 Announce Type: new
      • Abstract: Is human readability necessary for effective fine-tuning of large language models
      • We investigate whether model-conditioned training representations can preserve or improve adaptation utility without requiring a human-readable textual form

  • We propose Desired-Update-Aligned Synthetic Data (DASA), which uses activation-gradient feedback from a frozen reference model to guide the optimization of cont…

  • The Price of Token Boundaries: Compression Certificates and Prediction

    • Publication Time: 2026-09-30 12:00 Beijing Time
    • Abstract: - arXiv:2609.35869v1 Announce Type: new.
      • Abstract: Pre-tokenization restricts which text fragments can become prediction units, but its compression cost is obscured when tokenizers are compared only under the same boundaries.
      • We measure this cost by bounding the minimum token count from both sides, with and without a regular-expression boundary rule.
      • Nonnegative prices on token occurrences yield a lower bound through shortest paths and vocabulary-budget selection; maximizing over all prices recovers the linear programming relaxation, and an independent integer checker validates the reported values.
    • EN Key Points:
      • arXiv:2609.35869v1 Announce Type: new
      • Abstract: Pre-tokenisation restricts which text fragments can become prediction units, but its compression cost is obscured when tokenisers are compared only un…
      • We measure this cost by bounding the minimum token count from both sides, with and without a regular-expression boundary rule
      • Nonnegative prices on token occurrences yield a lower bound through shortest paths and vocabulary-budget selection; maximising over all prices recovers the line…

  • More Programs or More Rolls? Separating Coverage from Specialization in LLM Harnesses

    • Publication Time: 2026-09-30 12:00 Beijing Time
    • Abstract: - arXiv:2609.35873v1 Announce Type: new.
      • Abstract: Automated generation of LLM harnesses promises to improve inference through task specialization.
      • However, repeated execution of the same program may produce additional answer coverage, making specialization difficult to identify.
      • We introduce a controlled evaluation that separates answer coverage, repeatable task advantages, and gains from pre-execution selection.
    • EN Key Points:
      • arXiv:2609.35873v1 Announce Type: new
      • Abstract: Automated generation of LLM harnesses promises to improve inference through task specialization
      • Yet additional answer coverage can arise from repeated execution of the same program, making specialization difficult to identify
      • We introduce a controlled evaluation that separates answer coverage, repeatable task advantages, and gains from pre-execution selection

  • Risk-Averse Online POMDP Planning via CVaR of the Immediate Cost with Performance Guarantees

    • Publication Time: 2026-09-30 12:00 Beijing Time
    • Abstract: - arXiv:2609.35874v1 Announce Type: new.

  • Abstract: Online POMDP planners optimize the expected cumulative cost, which can mask dangerous states when the belief places significant mass on high-cost states.

  • Existing risk-averse methods apply static or dynamic Conditional Value at Risk (CVaR) to the value function, capturing trajectory-level risk, but share two gaps: (i) by retaining the immediate cost as an expectation over the state-dependent cost for a belief, risk \emph{in the belief is unaddressed}; and (ii) by modifying the value function, they require new, custom algorithms, rather than reusing existing expectation-based planners.

  • We instead apply CVaR to the immediate cost over the belief at each step, directly targeting per-step uncertainty about the current state.

    • EN Highlights:
      • arXiv:2609.35874v1 Announce Type: new
      • Abstract: Online POMDP planners optimize the expected cumulative cost, which can mask dangerous states when the belief places significant mass on high-cost stat…
      • Existing risk-averse methods apply static or dynamic Conditional Value at Risk (CVaR) to the value function, capturing trajectory-level risk, but share two gaps…
      • We instead apply CVaR to the immediate cost over the belief at each step, directly targeting per-step uncertainty about the current state

  • Beyond Symmetric Agents: Cognitive Diversity and Multi-Agent Debate in Small Language Models

    • Publication Time: 2026-09-30 12:00 Beijing Time
    • Abstract: - arXiv:2609.35875v1 Announcement Type: New.
      • Abstract: Multi-agent debate (MAD) has been reported to improve reasoning and factuality over single-model inference, but prior work has treated agents as symmetric peers, leaving the factors that drive the gains unidentified.
      • We test the hypothesis that cognitive diversity among agents is the driver, in a setting where the question remains measurable: small open-weight models with benchmark headroom.
      • Across 23 models from eleven vendor families, five tasks, and over 5,500 debate and control runs, we vary diversity along three axes—personas, sampling temperature, and model identity—pairing each debate configuration with a generation budget-matched majority-voting control.
    • EN Highlights:
      • arXiv:2609.35875v1 Announce Type: new
      • Abstract: Multi-agent debate (MAD) reportedly improves reasoning and factuality over single-model inference, but prior work treats agents as symmetric peers, le…
      • We test the hypothesis that cognitive diversity among agents is the driver, in the setting where the question is still measurable: small open-weight models with…
      • Across 23 models from eleven vendor families, five tasks, and 5,500+ debate and control runs, we vary diversity along three axes - personas, sampling temperatur…

  • Representational Simplicity and Circuit Size Dissociate in a Threshold-Dependent Way: A Controlled Test via Adversarial Training

    • Publication Time: 2026-09-30 12:00 Beijing Time

  • Abstract: - arXiv:2609.35890v1 Announce Type: new.

    • Abstract: The decomposability of sparse autoencoders and concentrated feature attribution are increasingly seen as evidence that a model’s computation is easier to reverse engineer.
    • Whether this cleanliness in representation and attribution actually predicts smaller or more manageable causal circuits remains an open question.
    • We test this directly using adversarial training as a controlled instrument: it reliably reshapes internal representations, but this alone does not constitute a test of circuit size.

  • Self-discovering RL in the Era of Experience: Is Learning History an Asset or a Burden?

    • Publication Time: 2026-09-30 12:00 Beijing Time
    • Abstract: - arXiv:2609.35897v1 Announce Type: new.
    • Abstract: The pursuit of recursive self-improvement (RSI) for general intelligence is divided between macro-level language model scaling and interaction-driven principles of the “age of experience.”
    • However, any self-improving architecture ultimately relies on its underlying optimization engine: if general intelligence needs to learn from foundational interactions, then the Reinforcement Learning (RL) update rule itself must be capable of cumulative adaptation.
    • While algorithmic self-discovery has produced Disco103, which surpasses PPO and achieves SOTA benchmark performance, its internal update mechanism remains an unexamined black box.

  • Grab a Coffee: Future-Aware Guidance for Discrete Diffusion with Compiled Objectives

    • Publication Time: 2026-09-30 12:00 Beijing Time
    • Abstract: - arXiv:2609.35924v1 Announce Type: new.
    • Abstract: Discrete diffusion models generate sequences by iteratively resolving multiple tokens in parallel, offering a flexible alternative to left-to-right generation.
    • However, guiding this process with sequence-level objectives is difficult because the value of an unresolved token depends on the other tokens with which it can form high-reward sequences.

  • Enumerating all such completions makes the whole guidance computation grow exponentially with the number of unresolved positions.

    • EN Highlights:
      • arXiv:2609.35924v1 Announce Type: new
      • Abstract: Discrete diffusion models generate sequences by iteratively resolving multiple tokens in parallel, offering a flexible alternative to left-to-right ge…
      • However, guiding this process with a sequence-level objective is difficult because the value of one unresolved token depends on the other tokens with which it c…
      • Enumerating all such completions makes the whole guidance computation grow exponentially with the number of unresolved positions

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

  • ChestPheNoT: Deployable, Auditable Label-Status-Evidence Extraction from Radiology Reports

    • Published: 2026-09-30 12:00 Beijing Time
    • Abstract: - arXiv:2609.31629v1 Announce Type: new.
      • Abstract: Structured phenotype extraction from radiology reports supports cohort construction, quality auditing, and clinical analytics, but practical deployment requires local inference and auditable predictions, while expert annotations remain scarce.
      • Traditional labelers provide structured findings and assertion states but no supporting evidence, while API-hosted large language models may be unsuitable when clinical text cannot leave institutional infrastructure.
      • We present CHESTPHENOT, a compact 0.5-3B language model that jointly extracts finding labels, three-class status (present/absent/uncertain), and verbatim supporting evidence spans.
    • EN Highlights:
      • arXiv:2609.31629v1 Announce Type: new
      • Abstract: Structured phenotype extraction from radiology reports supports cohort construction, quality auditing, and clinical analytics, but practical deploymen…
      • Conventional labelers provide structured findings and assertion states but no supporting evidence, while API-hosted large language models may be unsuitable when…
      • We present CHESTPHENOT, a compact 0.5-3B language model that jointly extracts finding labels, three-class status (present/absent/uncertain), and verbatim suppor…

  • A literature-guided descriptor-based framework for filtering composition search spaces

    • Published: 2026-09-30 12:00 Beijing Time
    • Abstract: - arXiv:2609.31650v1 Announce Type: new.
      • Abstract: The scientific literature contains latent knowledge about material behaviors, but much of this knowledge is expressed through text, context, and recurring associations, rather than explicit design principles.
      • This raises a core question: How can large-scale scientific corpora be used for practical problems in materials discovery?
      • Here, we propose a literature-guided descriptor-based filtering framework for reducing compositional search spaces.
    • EN Highlights:
      • arXiv:2609.31650v1 Announce Type: new

  • Abstract: Scientific literature contains latent knowledge about materials behavior, but much of this knowledge is expressed through words, contexts, and recurri…

  • This raises a central question: how can large-scale scientific corpora be used for practical problems in materials discovery

  • Here, we present a literature-guided descriptor-based filtering framework for reducing composition search spaces

  • Distributional sentiment modeling and anomaly detection for consumer complaint assessment

    • Publication Time: 2026-09-30 12:00 Beijing Time
    • Abstract: - arXiv:2609.31653v1 Announcement Type: New.
      • Abstract: Sentiment analysis is a common tool for converting unstructured text into quantitative signals in finance and risk management.
      • However, most applications reduce the output to discrete polarity labels or a single predictive feature, overlooking the distributional structure of sentiment intensity in consumer complaint narratives.
      • In this paper, we treat negative sentiment in consumer complaints as a bounded continuous variable and study its full distribution rather than a single label.
    • EN Highlights:
      • arXiv:2609.31653v1 Announce Type: new
      • Abstract: Sentiment analysis is a common tool for converting unstructured text into quantitative signals in finance and risk management
      • Yet most applications reduce the output to a discrete polarity label or a single predictive feature, overlooking the distributional structure of sentiment inten…
      • In this paper we treat negative sentiment in consumer complaints as a bounded continuous variable and study its full distribution rather than a single label

  • Parser, Chunking, and Embedding Interactions in Retrieval-Augmented Generation over Indian Government Regulatory Documents

    • Publication Time: 2026-09-30 12:00 Beijing Time
    • Abstract: - arXiv:2609.31660v1 Announcement Type: New. Abstract: Retrieval-Augmented Generation (RAG) pipelines are typically assembled from independently selected components (document parsers, chunking strategies, and embedding models), but these choices are rarely evaluated jointly, and evaluations combining them are often run on single documents or corpora.
      • We present a controlled factorial study involving 3 parsers, 3 chunking strategies, and 5 dense embedding models, along with a sparse BM25 baseline, evaluated against 800 question instances, each with one or more required evidence strings, which are automatically validated against the source text, with a 10% random sample manually reviewed, across four structurally diverse Indian central government regulatory documents.
      • We fit a linear mixed-effects model with document-query-level random intercepts to the generated 72,000-row result set, run Holm-corrected paired comparisons between matched dense and sparse configurations, and report cluster-bootstrapped confidence intervals for all 54 unique retriever configurations.
    • EN Highlights:
      • arXiv:2609.31660v1 Announce Type: new

  • Abstract: Retrieval-augmented generation (RAG) pipelines are typically assembled from independently-chosen components – a document parser, a chunking strategy,…

  • We present a controlled factorial study of 3 parsers, 3 chunking strategies, and 5 dense embedding models, together with a sparse BM25 baseline, evaluated again…

  • We fit linear mixed-effects models with document-query-level random intercepts to the resulting 72,000-row result set, run Holm-corrected paired comparisons bet…

  • LLM-Guided Ontology-Driven Knowledge Graph Construction from Unstructured Text

    • Published: 2026-09-30 12:00 Beijing Time
    • Abstract: - arXiv:2609.31663v1 Announce Type: new.
      • Abstract: Ontology-driven knowledge graph construction from industrial text remains challenging due to the domain specificity of documents, the scarcity of annotation resources, and the complexity of the ontology engineering workflow.
      • This paper presents and investigates the applicability of an ontology learning pipeline that combines compact open-source Large Language Models (LLMs), reusable prompting strategies, and open knowledge bases to support knowledge extraction, structuring, enrichment, and evaluation from text corpora.
      • The approach is tested and evaluated on a private French corpus of power-grid incident reports, using locally deployable open-source LLMs ranging from 7B to 32B parameters.
    • EN Key Points:
      • arXiv:2609.31663v1 Announce Type: new
      • Abstract: Ontology-driven knowledge graph construction from industrial text remains challenging due to the domain specificity of documents, the scarcity of anno…
      • This paper presents and investigates the applicability of an ontology learning pipeline that combines compact open-source Large Language Models (LLMs), reusable…
      • The approach is tested and evaluated on a private French corpus of power-grid incident reports, using locally deployable open-source LLMs ranging from 7B to 32B…

  • What Drives Dialectal Jailbreaks? An Ablation of Surface Form, Cultural Framing, and Strategy Banks

    • Published: 2026-09-30 12:00 Beijing Time
    • Abstract: - arXiv:2609.31664v1 Announce Type: new.
      • Abstract: Recent work has shown that obscure language registers can weaken the refusal behavior of large language models, especially when paired with black-box prompt optimization.
      • It is unclear whether the failures stem from non-standard surface forms, culture-based framing, or optimization over an expressive prompt strategy space.
      • We investigate Chinese language registers by extending a Classical Chinese red-teaming framework to Shanghainese and Cantonese, and running a 36-cell ablation over surface forms, strategy bank variants, and two target models.
    • EN Key Points:
      • arXiv:2609.31664v1 Announce Type: new

  • Abstract: Recent work suggests that obscure language registers can weaken large language model refusal behavior, especially when paired with black-box prompt op…

  • It remains unclear whether failures stem from non-standard surface form, culturally grounded framing, or optimization over an expressive prompt-strategy space

  • We study Chinese registers by extending a classical-Chinese red-teaming framework to Shanghainese and Cantonese and running a 36-cell ablation across surface fo…

  • Age-Adaptive Handwriting Reconstruction from an IMU-Based Digital Pen through Shared Representations and Domain-Specific Heads

    • Publication Time: 2026-09-30 12:00 Beijing Time
    • Abstract: - arXiv:2609.31666v1 Announcement Type: New.
      • Abstract: Digital pens are widely used to capture handwriting on digital devices, enabling precise trace recording and enhancing human-computer interaction.
      • However, most are bundled with tablets and lack cross-brand compatibility.
      • Recently, digital pens equipped with kinematic sensors have emerged, designed for use on any surface.
    • EN Key Points:
      • arXiv:2609.31666v1 Announce Type: new
      • Abstract: Digital pens are widely used to capture handwriting on digital devices, enabling precise trace recording and enhancing human-computer interaction
      • However, most are bundled with tablets and lack cross-brand compatibility
      • Recent digital pens equipped with kinematic sensors have emerged, designed for use on any surface

  • An Evaluation of AI-Supported Evidence-Based Learning for Public Speaking Skill Development

    • Publication Time: 2026-09-30 12:00 Beijing Time
    • Abstract: - arXiv:2609.31676v1 Announcement Type: New.
      • Abstract: Public speaking is an essential skill in academic and professional contexts, but it often causes anxiety.
      • Although several AI-based speech coaching tools exist, they typically provide generic feedback and lack model speeches for learning.
      • This study addresses these gaps by developing a system that incorporates speech context into its evaluation criteria, enabling more targeted feedback.
    • EN Key Points:
      • arXiv:2609.31676v1 Announce Type: new
      • Abstract: Public speaking is an essential skill in academic and professional contexts, but it often causes anxiety
      • Although several AI-based speech coaching tools exist, they typically provide generic feedback and lack model speeches for learning

  • This study addresses these gaps by developing a system that incorporates speech context into evaluation criteria, enabling more tailored feedback

  • The Temporal Tug-of-War: Visualizing and Detecting RAG Conflicts in Diffusion Models via Trajectory Variance

    • Publication Time: 2026-09-30 12:00 Beijing Time
    • Abstract: - arXiv:2609.31684v1 Announcement Type: New.
      • Abstract: Retrieval-Augmented Generation (RAG) introduces a specific failure mode in discrete diffusion language models: when the retrieved context contradicts parametric knowledge, the iterative denoising process becomes a visible battlefield between competing knowledge sources.
      • We identify temporal semantic divergence as an observable for detecting these conflicts and introduce the Trajectory Variance Score (TVS), a simple and interpretable measure of divergence.
      • TVS computes the mean pairwise cosine distance of answer embeddings across independent stochastic denoising trajectories, capturing the temporal tug-of-war between parametric and contextual attractors.
    • EN Highlights:
      • arXiv:2609.31684v1 Announce Type: new
      • Abstract: Retrieval-Augmented Generation (RAG) introduces a specific failure mode in discrete diffusion language models: when retrieved context contradicts para…
      • We identify temporal semantic divergence as an observable for detecting these conflicts and introduce the Trajectory Variance Score (TVS), a simple and interpre…
      • TVS computes the mean pairwise cosine distance of answer embeddings across independent stochastic denoising trajectories, capturing the temporal tug of war betw…

  • Verification of PETSc with CIVL using LLM-generated ACSL contracts and deterministic driver generation

    • Publication Time: 2026-09-30 12:00 Beijing Time
    • Abstract: - arXiv:2609.31687v1 Announcement Type: New.
      • Abstract: Parallel numerical libraries (e.g., PETSc) are widely used in scientific and engineering applications where incorrect results can have costly consequences.
      • Despite this, numerical libraries are rarely formally verified.
      • One of the challenges is the need for an expert to hand-write specifications and manually apply verification tools, often requiring the development of harnesses or drivers, all of which can contain additional errors, leading to false positives or negatives during verification.
    • EN Highlights:
      • arXiv:2609.31687v1 Announce Type: new
      • Abstract: Parallel numerical libraries such as PETSc are widely used in science and engineering applications where wrong results can have costly consequences
      • Despite this, numerical libraries are rarely formally verified
      • One of the challenges is the need for an expert to hand-write a specification and manually apply a verification tool, often requiring the development of a harne…

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

  • Replay in the Silent Degrees of Freedom: Continual Learning Without an Offline Phase

    • Posted: 2026-09-30 12:00 Beijing Time
    • Abstract: - arXiv:2609.31630v1 Announce Type: new.
      • Abstract: Replay-based continual learning almost always consolidates in a dedicated offline phase or by interleaving replayed samples with the input stream, whereas the brain also consolidates during waking hours through local sleep, brief use-dependent closing periods of single circuits.
      • We ask whether a network trained by local, biologically constrained rules can consolidate with no offline phase at all.
      • An isolation rule confines replay updates to hidden synapses invisible to the current input under a k-winners-take-all dynamic, with optimizer state advanced only within the mask; a refractory rotation rule prevents just-fired units from competing in the next contest, widening the set of what can be consolidated; homeostatic pressure and a relative novelty gate determine when replay bursts and when the rotation runs.
    • EN Highlights:
      • arXiv:2609.31630v1 Announce Type: new
      • Abstract: Replay-based continual learning almost always consolidates in a dedicated offline phase or by interleaving replayed samples with the input stream, whe…
      • We ask whether a network trained by local, biologically constrained rules can consolidate with no offline phase at all
      • An isolation rule confines replay updates to hidden synapses invisible to the current input under k-winner-take-all dynamics, with optimiser state advanced only…

  • OMP-MoE: Efficient Expert Pruning for Mixture-of-Experts LLMs via Orthogonal Matching Pursuit

    • Posted: 2026-09-30 12:00 Beijing Time
    • Abstract: - arXiv:2609.31631v2 Announce Type: new.
      • Abstract: Mixture-of-Experts (MoE) models can effectively scale large language models, but face significant deployment challenges due to large memory requirements.
      • Existing pruning methods either lead to high search costs or ignore the dynamic interdependencies between experts.
      • To address these challenges, we propose OMP-MoE, a novel training-free compression framework for reducing expert redundancy in MoE-based LLMs.
    • EN Highlights:
      • arXiv:2609.31631v2 Announce Type: new
      • Abstract: Mixture-of-Experts (MoE) models enable efficient scaling of large language models but face critical deployment challenges due to massive memory requir…
      • Existing pruning methods either incur prohibitive search costs or neglect the dynamic interdependencies between experts
      • To address these challenges, we present OMP-MoE, a novel training-free compression framework for reducing expert redundancy in MoE-based LLMs

  • EEGAgentBench: Benchmarking LLM Agents on Short- and Long-Horizon EEG Analysis

    • Posted: 2026-09-30 12:00 Beijing Time
    • Abstract: - arXiv:2609.31632v1 Announce Type: new.

  • Abstract: Electroencephalogram (EEG) analysis is evolving from short-segment classification to long-horizon interpretation, which requires iterative evidence accumulation, multi-step reasoning, and the coordinated use of specialized signal processing tools.

  • Although large language models (LLMs) have recently shown promise as autonomous agents for EEG analysis, existing EEG agent evaluations remain fragmented, covering limited tasks within a narrow timeframe, with inconsistent protocols, and do not provide a comprehensive assessment of the agent’s reasoning, tool use, and workflow construction abilities.

  • To address this gap, we propose \textbf{EEGAgentBench}, a unified benchmark for systematically evaluating LLM agents on short- and long-horizon EEG analysis.

  • EN Highlights:

    • arXiv:2609.31632v1 Announce Type: new
    • Abstract: Electroencephalography (EEG) analysis is evolving from short-segment classification toward long-horizon interpretation that demands iterative evidence…
    • Although large language models (LLMs) have recently shown promise as autonomous agents for EEG analysis, existing EEG agentic evaluations remain fragmented, cov…
    • To address this gap, we propose \textbf{EEGAgentBench}, a unified benchmark for systematically evaluating LLM agents on short- and long-horizon EEG analysis

  • Enhancing generalization in endwall film cooling prediction: Incorporating the superposition principle into transformer-based neural operators

    • Publication Time: 2026-09-30 12:00 Beijing Time
    • Abstract: - arXiv:2609.31633v1 Announcement Type: new.
      • Abstract: In this study, a physics-enhanced neural operator framework is proposed to enhance the generalization prediction ability for turbine endwall cooling layouts with a variable number of film holes.
      • Specifically, inspired by the principle of film cooling superposition, we propose a film cooling prediction model, namely the Superposition-based Deep Neural Operator (SDNO), which divides the endwall temperature field prediction into two stages.
      • In the first stage, the cooling layout of the turbine endwall is divided into several sub-parts with randomly assigned film holes, and a Transformer-based neural operator network (i.e., Calculate Net) is designed to predict the temperature field of each sub-part.
    • EN Highlights:
      • arXiv:2609.31633v1 Announce Type: new
      • Abstract: In this study, a physics-enhanced neural operator framework is proposed to enhance the generalization prediction ability of the cooling layout of a tu…
      • Specifically, inspired by the film cooling superposition principle, we propose a film cooling prediction model, namely superposition-based deep neural operator…
      • In the first stage, the cooling layout of a turbine endwall is divided into several sub-parts with randomly assigned film holes, and a Transformer-based neural…

  • Symmetry-quotient Flatness and Generalization

    • Publication Time: 2026-09-30 12:00 Beijing Time
    • Abstract: - arXiv:2609.31634v1 Announcement Type: new.

  • Abstract: This paper develops a theorem-level pipeline in the symmetry quotient setting: quotient linear stability implies quotient flatness, quotient flatness implies input smoothness, and input smoothness yields generalization under a local covering assumption.

  • Flatness is often associated with generalization, and Stochastic Gradient Descent (SGD) is frequently viewed as implicitly biased toward flat solutions.

  • However, standard flatness measures are typically defined in the raw parameter space and are therefore not invariant under function-preserving symmetries such as positive rescaling.

  • EN Highlights:

    • arXiv:2609.31634v1 Announce Type: new
    • Abstract: This paper develops a theorem-level pipeline in symmetry-quotient settings: quotient linear stability implies quotient flatness, quotient flatness imp…
    • Flatness is often associated with generalization, and Stochastic Gradient Descent (SGD) is frequently viewed as implicitly biased toward flat solutions
    • However, standard flatness measures are typically defined in the raw parameter space and are therefore not invariant under function-preserving symmetries such a…

  • What Next-Event Accuracy Cannot See: Closed-Loop Evaluation of Emergency Department Trajectory Simulators

    • Publication Time: 2026-09-30 12:00 Beijing Time
    • Abstract: - arXiv:2609.31635v1 Announcement Type: New.
      • Abstract: Clinical trajectory models are typically evaluated by next-event accuracy on observed histories.
      • Simulation is different: models must condition on their own generated events, allowing errors to compound.
      • Although this problem is well known in sequence modeling, it has not been systematically quantified for clinical trajectory simulators.
    • EN Highlights:
      • arXiv:2609.31635v1 Announce Type: new
      • Abstract: Clinical trajectory models are usually evaluated by next-event accuracy on observed histories
      • Simulation is different: models must condition on their own generated events, allowing errors to compound
      • Although this problem is well known in sequence modelling, it has not been systematically quantified for clinical trajectory simulators

  • Grounding Vision-Language Models in Driving Semantics: A Multi-Dataset Predicate Framework for Explainable Reasoning

    • Publication Time: 2026-09-30 12:00 Beijing Time
    • Abstract: - arXiv:2609.31636v1 Announcement Type: New.
      • Abstract: Vision-language models are increasingly used for driving scene understanding, but the semantic relationships expressed in their outputs are often difficult to verify against the underlying traffic conditions.
      • This paper introduces a deterministic, multi-dataset predicate framework that derives driving scene semantics from measurable geometric, kinematic, temporal, map, and traffic control evidence.
      • Dataset-specific interfaces are only used to recover the required scene information, while the predicate definitions remain constant across nuPlan and nuScene and are materialized in a universal predicate knowledge graph.
    • EN Highlights:
      • arXiv:2609.31636v1 Announce Type: new

  • Abstract: Vision-language models are increasingly used for driving-scene understanding, yet the semantic relations expressed in their outputs are often difficul…

  • This paper introduces a deterministic multi-dataset predicate framework that derives driving-scene semantics from measurable geometric, kinematic, temporal, map…

  • Dataset-specific interfaces are used only to recover the required scene information, while predicate definitions remain unchanged across nuPlan and nuScenes and…

  • FIDAL: Diversity-Aware Federated Active Learning Under Real-World Distribution Shifts

    • Release Time: 2026-09-30 12:00 Beijing Time
    • Abstract: - arXiv:2609.31637v1 Announcement Type: new.
      • Abstract: Federated learning enables collaborative model training across institutions without centralizing data, yet high annotation costs, domain shifts, and class imbalance remain major obstacles, especially when irrelevant out-of-distribution (OOD) samples dilute the labeled data.
      • Existing active learning methods target uncertainty or diversity within in-distribution (ID) data and overlook unknown samples in federated clinical settings.
      • We propose FIDAL, an open-set federated active learning framework that combines calibrated global-local evidential uncertainty, support-set diversity weighting, and adaptive OOD rejection.
    • EN Highlights:
      • arXiv:2609.31637v1 Announce Type: new
      • Abstract: Federated learning enables collaborative model training across institutions without centralizing data, yet high annotation costs, domain shifts, and c…
      • Existing active learning methods target uncertainty or diversity within in-distribution (ID) data and overlook unknown samples in federated clinical settings
      • We propose FIDAL, an open-set federated active learning framework that combines calibrated global-local evidential uncertainty, support-set diversity weighting,…

  • Energy-aware frugal Bayesian optimization

    • Release Time: 2026-09-30 12:00 Beijing Time
    • Abstract: - arXiv:2609.31638v1 Announcement Type: new.
      • Abstract: The primary goal of modern design optimization frameworks is to provide the most accurate predictive models without balancing computational overhead.
      • This remains a challenge for large-scale architectural and multi-disciplinary design optimization problems, even with sample-efficient Bayesian optimizers.
      • In this paper, we introduce metrics that quantify the computational energy footprint within a Bayesian optimization framework to guide the model’s parameter settings toward a configuration that balances performance and frugality.
    • EN Highlights:
      • arXiv:2609.31638v1 Announce Type: new
      • Abstract: Modern design optimization frameworks aim first and foremost for models with the most accurate predictions without balancing computational overhead

  • It remains a reason why scaled architecture and multidisciplinary design optimization problems are difficult to address, even with sample-efficient Bayesian opt…

    • In this paper, a metric quantifying the computational energy footprint is introduced within a Bayesian optimization framework to guide the parameter setting of…

  • When Does Domain Adaptation Help on Physical Vibration Sensors? A Held-Out-Bearing Study of Neural-Operator and Convolutional Models

    • Publish Time: 2026-09-30 12:00 Beijing Time
    • Abstract:
      • arXiv:2609.31639v1 Announce Type: New.
      • Abstract: Diagnosing rolling-element bearing faults from vibration is a canonical physical sensing task and a widely used benchmark for domain adaptation under varying operating conditions.
      • Accuracies above 99% are commonly reported, but evaluation uses the same physical bearing orientation for both training and testing.
      • We revisit the task under a held-out-bearing protocol, assigning each bearing unit entirely to either the training or test set, and find that source-only transfer is far weaker than these numbers suggest: reaching only $0.36$ with varying shaft speeds, while the target supervision upper bound is $0.97$.