System translated (Gemini)

🤖 AI 速览

OpenAI released Dots, an AI agent capable of persistent work across conversations, marking AI’s evolution from single-turn tools to long-term autonomous execution. Anthropic’s Claude Opus 5.5 surpassed GPT-6 Astra in the Epoch AI Index, shifting the ranking of top model capabilities. …
📋 文章元数据
发布时间
2026-10-03
类型
ai-daily
字数
6647
阅读时长
32 min

2026-10-03 AI Daily | OpenAI Dots Usher in the Era of Persistent Agents, Claude Opus 5.5 Tops Epoch AI Index Link to heading

OpenAI has released Dots, AI agents that persist across conversations, marking an evolution for AI from single-turn tools to long-term autonomous execution. Anthropic’s Claude Opus 5.5 has surpassed GPT-6 Astra on the Epoch AI Index, reshuffling the capability rankings of top models. Meanwhile, the GPT-6 series model guide has been published, offering tiered recommendations for reasoning, tool coordination, and production deployment. Concurrently, multiple papers are beginning to question the systemic biases of existing LLM evaluation metrics, bringing deployability and genuine trustworthiness into the spotlight.

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

Three main themes are worth your time today. First is the “sense of uncertainty” surrounding policy and product roadmaps. One podcast offered an in-depth review of the Trump superintelligence summit and AI safety protocols, placing data center national security, the cybersecurity arms race, and the latest economic data within a single framework. OpenAI also released its GPT-6 series model guide, offering more granular, tiered recommendations for reasoning, tool coordination, and production deployment. Second is a collective rethinking of LLM evaluation. Multiple papers point to the same issue: first-token logic, structures of human bias, and interpretability metrics may systematically overestimate or distort model behavior. Evaluation teams should especially read “The First Token Is Not the Verdict.” Third is deployability and cost. From on-device NER and Text-to-SQL schema budgets to synthetic RLVR data auditing and the carbon footprint of large Arabic-language models, the focus of engineering decisions is shifting from a capabilities race to cost, data provenance, and trustworthiness.

🌐 AI Hot Topics on X Link to heading

Topic 1: Anthropic Launches Mods for Customizing Claude Code Link to heading

  • Category: AI · News
  • Summary: Trending for: 1 day ago, Related posts: 9,100
  • What it is: Anthropic has launched “Mods” for Claude Code, allowing developers to use TypeScript extensions to customize prompts, interfaces, and functionalities.
  • Why it matters: This gives the AI coding agent plug-in-based extensibility, which is expected to strengthen the developer ecosystem and drive the evolution of AI programming tools from general-purpose models to deeply customizable platforms.
  • Discussion summary: Discussions on X are focused on the flexibility offered by TypeScript hooks, comparing it with existing plug-in ecosystems, exploring the limits of customization, and raising concerns about extension security and fragmentation.

Topic 2: Earendil Launches Pi 1.0 Stable AI Coding Agent Link to heading

  • Category: AI · News
  • Summary: Trending for: 1 day ago, Related posts: 2,800
  • What it is: Earendil has released the Pi 1.0 stable version of its AI coding agent.
  • Why it matters: This marks the transition of AI coding agents from an experimental phase to stable, production-ready tools, potentially accelerating the automation of software development workflows.
  • Discussion summary: Discussions on X are mainly focused on its practical coding stability and reliability, its differences from existing AI coding tools, and its potential impact on developer workflows and employment.

Topic 3: OpenAI Launches Persistent AI Agents Called Dots Link to heading

  • Category: AI · News
  • Summary: Trending for: 1 day ago, Related posts: 45,000
  • What it is: OpenAI has launched Dots, AI agents that can persist between chats, and is making them available to select paid users and markets.
  • Why it matters: This marks an evolution for AI from single-turn conversational tools to agents capable of long-term autonomous task execution, raising new requirements for permission management, reliability, and automated workflows.
  • Discussion summary: Discussions on X center on the permission boundaries and reliability risks of persistent agents. Supporters believe they can transform work methods, while critics worry about a lack of transparency and control in autonomous operations.

Topic 4: OpenAI Paid Users Await Delayed Global Usage Reset Link to heading

  • Category: AI · News
  • Summary: Trending for: 15 hours ago, Related posts: 13,000
  • What it is: OpenAI’s paid users experienced a delay in the global reset of their usage limits, with quotas not being restored as scheduled.
  • Why it matters: This incident highlights operational stability issues for commercial AI services under high load and affects the rights of paid users and trust in the platform.
  • Discussion summary: On X, many users complained that the delay impacted their daily work, demanding compensation or a transparent explanation. Others discussed whether this exposes inadequacies in OpenAI’s infrastructure and customer service response.

Topic 5: Anthropic’s Claude Opus 5.5 Tops Epoch AI Index Ahead of OpenAI’s GPT-6 Astra Link to heading

  • Category: AI · News
  • Summary: Trending for: 1 day ago, Related posts: 13,000
  • What it is: Anthropic’s Claude Opus 5.5 has surpassed OpenAI’s GPT-6 Astra to rank first in the Epoch AI Index.
  • Why it matters: This indicates a shift in the capability rankings of frontier large models, with Anthropic taking a temporary lead on an important evaluation benchmark. This could influence industry expectations regarding the competitive landscape of top-tier models.
  • Discussion summary: Discussions on X primarily revolve around the credibility of the index, the extent of Claude Opus 5.5’s lead, and whether the fact that GPT-6 Astra is not yet fully available affects the fairness of the comparison.

Topic 6: PewDiePie Launches Uncensored Local AI Ajax After OpenAI Bans Link to heading

  • Category: AI · News
  • Summary: Trending time: 20 hours ago, Related posts: 3300
  • What it is: PewDiePie has launched a locally-run, uncensored AI tool named Ajax after facing a ban from OpenAI.
  • Why it matters: This event highlights the conflict between creators and AI platforms over content moderation, potentially promoting localized, decentralized AI as an alternative to circumvent censorship.
  • Discussion summary: Discussions on X focus on the implications of uncensored AI for free speech versus its potential risks of misuse. There is also attention on whether PewDiePie’s celebrity influence can accelerate the adoption of local open-source models by the general public.

Topic 7: Ben Affleck Reveals AI Use in Directing Netflix Thriller Animals Link to heading

  • Category: AI · Entertainment
  • Summary: Trending time: 1 day ago, Related posts: 23000
  • What it is: Ben Affleck revealed that he used AI technology while directing the Netflix thriller “Animals.”
  • Why it matters: This news shows that AI is rapidly entering the mainstream film production workflow, sparking widespread industry concern about creative efficiency, its impact on jobs, and artistic originality.
  • Discussion summary: The focus of discussion on X includes the specific applications of AI in filmmaking, whether it threatens traditional film industry jobs, and whether a director’s use of AI in the creative process diminishes artistic value.

AI Public Opinion Summary on X Today Link to heading

Today’s main narrative revolves around AI agents and tools moving from experimentation to production, platformization, and persistence. New releases like Claude Code Mods, Pi 1.0, and OpenAI Dots show that AI coding and task execution are shifting from single-turn conversations to customizable, long-running intelligent agents. Claude Opus 5.5 surpassing GPT-6 Astra on benchmarks has also intensified the discussion about competition at the top. The consensus is that AI is rapidly entering practical scenarios like development, workflows, and film production, leading to increased efficiency and ecosystem expansion. Points of disagreement are centered on whether autonomous agents and local uncensored tools are more trustworthy, and whether AI’s involvement in film creation undermines jobs and artistic originality. Potential risks mainly include operational stability and user trust issues exposed by the delay in restoring OpenAI’s paid quotas, as well as controversies over plugin fragmentation, the misuse of uncensored models, the fairness of benchmark evaluations, and the impact on employment in the film and television industry.

💡 Influencer Insights Link to heading

Influencer insights are not available 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; 22 sources covered; 34 updates in total

All-In Podcast (A_full) Link to heading

  • Trump’s Super Intelligence Summit, AI Safety Accord, GDP Beats, Midterm Predictions
    • Publication time: 2026-10-03 06:30 Beijing Time
    • Summary: - (0:00) Bestie intros.
      • (2:01) Trump’s Super Intelligence Summit: Behind the scenes and how the accord was created and signed.
      • (17:57) Cybersecurity arms race, data center national security, America.gov, coordinated doomer messaging.
      • (45:22) Latest economic data: GDP, jobs, diesel, inflation, interest rates, manufacturing, poverty, etc.
    • EN Highlights:
      • (0:00) Bestie intros
      • (2:01) Trump’s Super Intelligence Summit: Behind the scenes and how the accord was created and signed
      • (17:57) Cybersecurity arms race, data center national security, America.gov, coordinated doomer messaging
  • (45:22) Recent economic data: GDP, jobs, diesel, inflation, rates, manufacturing, poverty, and more

Stratechery by Ben Thompson (A_full) Link to heading

  • 2026.40: Dots and Question Marks
    • Published: 2026-10-03 01:00 Beijing Time
    • Abstract:- (Photo by Heather Diehl/Getty Images).
      • Welcome back to This Week in Stratechery!
      • As a reminder, each week, every Friday, we’re sending out this overview of content in the Stratechery bundle; highlighted links are free for everyone.
      • Additionally, you have complete control over what we send to you.
      • With that said, here are some of our favorites from this week.
    • EN Key Points:
      • (Photo by Heather Diehl/Getty Images)
      • Welcome back to This Week in Stratechery
      • As a reminder, each week, every Friday, we’re sending out this overview of content in the Stratechery bundle; highlighted links are free for everyone
      • Additionally, you have complete control over what we send to you

OpenAI Blog (A_full) Link to heading

  • A model guide for the GPT-6 family

    • Published: 2026-10-03 00:15 Beijing Time
    • Abstract:- Learn how startups can choose GPT-6 models, tune reasoning effort, improve prompts and skills, coordinate tools, and prepare workflows for production.
      • This post from the OpenAI Blog explains how the model guide for the GPT-6 family is shaping the broader AI and infrastructure landscape.
      • It also provides practical implications for founders, operators, and investors following the model guide for the GPT-6 family.
    • EN Key Points:
      • Learn how startups can choose GPT-6 models, tune reasoning effort, improve prompts and skills, coordinate tools, and prepare workflows for production.

  • Chatham scales its capital markets expertise with OpenAI

    • Published: 2026-10-02 08:00 Beijing Time
    • Abstract:- Chatham Financial uses Codex and GPT-5.6 to build technology and redesign workflows, cutting trade validation from 30 minutes to under 4 minutes.
      • This post from the OpenAI Blog explains how Chatham is scaling its capital markets expertise with OpenAI, shaping the broader AI and infrastructure landscape.
      • It also presents practical implications for founders, operators, and investors following Chatham’s scaling of its capital markets expertise with OpenAI.
    • EN Key Points:
      • Chatham Financial uses Codex and GPT-5.6 to build technology and redesign workflows, cutting trade validation from 30 minutes to under 4.

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

  • Heavy-Tailed Memory Traces in Long-Horizon Language Agents

    • Published: 2026-10-02 12:00 Beijing Time
    • Abstract:- arXiv:2610.00010v1 Announcement Type: new.
      • Abstract: Long-horizon language agents increasingly rely on external memory as a frozen world model, yet current memory systems are often judged solely by task success or token cost.

  • We argue that the missing object is the shape of memory use: under finite context and repeated retrieval, agent memory can concentrate on a small core, leaving rare states in the long tail where prediction errors accumulate.

    • We study this effect through a conservative tail audit and find that concentration is reproducible but policy-dependent.
    • EN Highlights:
      • arXiv:2610.00010v1 Announce Type: new
      • Abstract: Long-horizon language agents increasingly rely on external memory as a frozen world model, yet current memory systems are usually judged only by task…
      • We argue that the missing object is the shape of memory use: under finite context and repeated retrieval, agent memory can concentrate on a small core while lea…
      • We study this effect through a conservative tail audit and find that concentration is reproducible but policy-dependent

  • When Do Causal World Models Help Modular LLM Agents

    • Publication Time: 2026-10-02 12:00 Beijing Time
    • Abstract:- arXiv:2610.00012v1 Announce Type: new.
      • Abstract: LLM agents increasingly operate through modular systems (e.g., order, payment, inventory, and shipment services), where actions in one module change the valid transitions in another.
      • Standard world models usually fit observational traces, but this is not the quantity needed for intervention-time planning: a trace may show payment precedes shipment, but cannot determine if payment authorizes shipment, if inventory regulates the effect, or if a hidden trigger explains both.
      • We study this gap through FedCausalCompose, a causal world-model framework for modular LLM agents in which local actions provide intervention-response evidence for cross-module interfaces.
    • EN Highlights:
      • arXiv:2610.00012v1 Announce Type: new
      • Abstract: LLM agents increasingly act through modular systems, such as order, payment, inventory, and shipment services, where actions in one module change whic…
      • Standard world models usually fit observational traces, but this is not the quantity needed for intervention-time planning: a trace may show that payment preced…
      • We study this gap through FedCausalCompose, a causal world-model framework for modular LLM agents in which local actions provide intervention-response evidence…

  • From Proposal to Verified Effect: Praxa, an Evidence-Bound Harness for Governed AI Agent Execution

    • Publication Time: 2026-10-02 12:00 Beijing Time
    • Abstract:- arXiv:2610.00015v1 Announce Type: new.
      • Abstract: Large language model agents can propose and execute actions, but proposal, authorization, scheduling, verification of external effects, and service promotion are different matters.
      • We introduce Praxa, an agent tool that explicitly represents these states through deterministic admission, agent execution, external read-back, coordination, and review elevation.
      • We report four evidence channels.
    • EN Highlights:
      • arXiv:2610.00015v1 Announce Type: new

  • Abstract: Large-language-model agents can propose and execute actions, but proposal, authority, dispatch, verified external effect, and serving promotion are di…

  • We present Praxa, an agent harness that represents these states explicitly through deterministic admission, brokered execution, external read-back, reconciliati…

  • We report four evidence lanes

  • What Do Rationales Communicate? A Message-Intervention Study in Role-Specialized QA

    • Release Time: 2026-10-02 12:00 Beijing Time
    • Abstract: - arXiv:2610.00018v1 Announce Type: new.
      • Abstract: Role-specialized QA pipelines increasingly pass rationales from a reasoner to a verifier, but it is unclear what this message actually buys: better answers, stronger support evaluations, or new failure surfaces.
      • We introduce a message-intervention diagnostic that fixes the evidence and candidate answer while varying only the rationale passed across the reasoner-to-verifier boundary.
      • On 400 MuSiQue, HotpotQA, and 2WikiMultiHopQA examples using DeepSeek as both generator and verifier, faithful rationales add almost no answer accuracy over having no rationale, while corrupted rationales strongly alter support judgments.
    • EN Highlights:
      • arXiv:2610.00018v1 Announce Type: new
      • Abstract: Role-specialized QA pipelines increasingly pass rationales from a reasoner to a verifier, but it is unclear what this message actually buys: better an…
      • We introduce a message-intervention diagnostic that fixes the evidence and candidate answer while varying only the rationale passed across the reasoner-to-verif…
      • On 400 MuSiQue, HotpotQA, and 2WikiMultiHopQA examples with DeepSeek as generator and verifier, faithful rationales add almost no answer accuracy over no ration…

  • Measuring the Microtask Eligibility Gap: When Is an Off-the-Shelf SLM Enough for an Agent Harness?

    • Release Time: 2026-10-02 12:00 Beijing Time
    • Abstract: - arXiv:2610.00025v1 Announce Type: new.
      • Abstract: Agents increasingly seek to run Small Language Models (SLMs) on microtasks surrounding a frontier Large Language Model (LLM) planner: auto-approving shell commands, writing to memory, selecting tools, ranking past turns.
      • We ask whether off-the-shelf SLMs meet practitioner-defined thresholds, why they fail when they do, and whether quantization changes the answer.
      • We build a benchmark of 4 such microtasks with fixed prompts and automated metrics, each with a pre-specified threshold $\tau$ anchored to cheap non-LLM baselines and a CI-aware eligibility rule (a configuration passes only if its confidence bounds clear $\tau$).
    • EN Highlights:
      • arXiv:2610.00025v1 Announce Type: new

  • Abstract: Agent harnesses increasingly want to run small language models (SLMs) on the microtasks around a frontier large language model (LLM) planner: auto-app…

  • We ask whether off-the-shelf SLMs meet practitioner-defined thresholds and, when they fail, why, and whether quantization changes the answer

  • We build a benchmark of 4 such microtasks with fixed prompts and automatic metrics, each with a pre-specified threshold $\tau$ anchored to a cheap non-LLM basel…

  • Characterizing a Configuration Where Inference-Time PRM-Pruned Fragment Grafting Is Inert: Evidence from Three Reasoning LMs

    • Publication Time: 2026-10-02 12:00 Beijing Time
    • Abstract: - arXiv:2610.00047v1 Announcement Type: New.
      • Abstract: Diversity collapse in parallel chains of thought has motivated inference-time interventions based on a natural design: when a process reward model (PRM) prunes a chain, its high-PRM prefix is extracted and transplanted verbatim as a contextual demonstration into its still-decoding siblings.
      • We isolate this mechanism (PRM-Pruned Fragment Grafting (PPFG)) as the most cost-minimal operationalization of cross-trajectory step-level transfer and test it at an operating point where prior fragment grafting work reports gains only with additional compensating components.
      • On Qwen2.5-7B-Instruct with Math-Shepherd on the full MATH500 (n=500, three seeds), PPFG in both stagnation- and random-targeting variants is statistically indistinguishable from an independent parallel CoT baseline on every measured axis.
    • EN Highlights:
      • arXiv:2610.00047v1 Announce Type: new
      • Abstract: Diversity collapse in parallel chain-of-thought has motivated inference-time interventions built on a natural design: when a process reward model (PRM…
      • We isolate this mechanism, PRM-Pruned Fragment Grafting (PPFG), as the most cost-minimal operationalization of cross-trajectory step-level transfer, and test it…
      • On Qwen2.5-7B-Instruct with Math-Shepherd on full MATH500 (n=500, three seeds), PPFG in both stagnation- and random-targeting variants is statistically indistin…

  • Gradient-Aligned Pair Selection for Personalized Preference Optimization

    • Publication Time: 2026-10-02 12:00 Beijing Time
    • Abstract: - arXiv:2610.00061v1 Announcement Type: New.
      • Abstract: Personalizing large language models (LLMs) requires aligning generative behavior with user-specific preferences rather than overall quality.
      • While Direct Preference Optimization (DPO) provides a stable framework for preference learning, its effectiveness in personalized settings critically depends on how preference pairs are selected.
      • Existing methods often rely on heuristic criteria, such as likelihood-based extreme values, which decouples the optimization from explicit user utility and can lead to suboptimal personalization.
    • EN Highlights:
      • arXiv:2610.00061v1 Announce Type: new

  • Abstract: Personalizing large language models (LLMs) requires aligning generation behavior with user-specific preferences rather than aggregate quality

  • While Direct Preference Optimization (DPO) provides a stable framework for preference learning, its effectiveness in personalized settings critically depends on…

  • Existing approaches typically rely on heuristic criteria, such as likelihood-based extremes, which decouple optimization from explicit user utility and can lead…

  • K-Dense BYOK: An Open-Source AI Research Assistant That Runs Locally and Keeps a Hash-Chained Lab Notebook

    • Publication Time: 2026-10-02 12:00 Beijing Time
    • Abstract: - arXiv:2610.00074v1 Announcement Type: new.
      • Abstract: K-Dense BYOK (bring your own keys) is a free, open-source AI research assistant for scientists in any field that runs on the researcher’s own computer.
      • The researcher provides access to a model of their choice (hosted or running locally), and the application provides everything else: a place for the work to run, a layer of scientific scaffolding, and a complete record.
      • Each project is an ordinary folder, so the data, code, results, and record remain on a machine administered by the researcher and can be read years later without the application.
    • EN Highlights:
      • arXiv:2610.00074v1 Announce Type: new
      • Abstract: K-Dense BYOK (bring your own keys) is a free, open-source AI research assistant for scientists in any field that runs on the researcher’s own computer
      • The researcher supplies access to a model of their choice, hosted or running locally, and the application supplies everything else: a place for the work to run,…
      • Each project is an ordinary folder, so the data, the code, the results, and the record stay on a machine the researcher administers and can be read years later…

  • Scientific Agents: Evaluating Profession-Specific System Prompts on Scientific Tasks

    • Publication Time: 2026-10-02 12:00 Beijing Time
    • Abstract: - arXiv:2610.00084v1 Announcement Type: new.
      • Abstract: Detailed, profession-specific system prompts increased token usage and estimated cost per response, but without consistent gains in accuracy.
      • We evaluate Scientific Agents using Gemini 3.8 Flash via OpenRouter in the Pi agent tool, an open-source corpus of 503 professional AGENTS.md profiles.
      • We compare matched profiles against four controls: a minimal baseline (“You are a helpful assistant”), the profile’s opening role sentence, general scientific rigor guidelines, and a profile from an unrelated domain.
    • EN Highlights:
      • arXiv:2610.00084v1 Announce Type: new

  • Abstract: Detailed profession-specific system prompts raise token use and estimated cost per response without a consistent accuracy gain

  • We evaluate Scientific Agents, an open-source corpus of 503 profession-specific AGENTS.md profiles, with Gemini 3.8 Flash via OpenRouter in the Pi agent harness

  • We compare matched profiles with four controls: a minimal baseline (“You are a helpful assistant”), the profile’s opening role sentence, a generic scientific ri…

  • Comedic Fool’s Gold: Reward Exploits and Countermeasures in Conversational Humor

    • Publication Time: 2026-10-02 12:00 Beijing Time
    • Abstract: - arXiv:2610.00197v1 Announce Type: new.
      • Abstract: We investigate automated rewards for training language models in conversational humor, focusing on reward exploits and countermeasures.
      • Two approaches aim to capture understandable surprise and predicted audience amusement.
      • Controlled tests show that an embedding-based surprise reward accepts word-shuffled replies as readily as witty ones.
    • EN Key Points:
      • arXiv:2610.00197v1 Announce Type: new
      • Abstract: We investigate automated rewards for training language models in conversational humor, focusing on reward exploits and countermeasures
      • Two approaches aim to capture understandable surprise and predicted audience amusement
      • Controlled tests show that an embedding-based surprise reward accepts word-shuffled replies as readily as witty ones

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

  • On-Device Named-Entity Recognition: A Deployability Study of Accuracy, Cost, Reliability, and Confidence

    • Publication Time: 2026-10-02 12:00 Beijing Time
    • Abstract: - arXiv:2610.00007v1 Announce Type: new.
      • Abstract: Named-entity recognition (NER) is increasingly wanted on-device (no API, low latency, data kept local).
      • The question for practitioners is not about leaderboards, but which model is deployable, how to evaluate it without human annotation, and whether its confidence is trustworthy.
      • We place 9 systems across three paradigms and 13M to 8B parameters on three datasets with different characteristics: a classic tagger (spaCy), bidirectional encoder experts (GLiNER, 166 to 460M), and locally run generative LLMs (Qwen3-0.6B/1.7B/4B-Instruct, DeepSeek-R1-1.5B/8B), and report precision plus two axes the literature overlooks: latency and output validity.
    • EN Key Points:
      • arXiv:2610.00007v1 Announce Type: new
      • Abstract: Named-entity recognition (NER) is increasingly wanted on-device (no API, low latency, data kept local)

  • The practitioner’s question is not the leaderboard but which model is deployable, how to evaluate it without human annotation, and whether its confidence can be…

  • We answer these jointly

  • SCM-based Fairness and Faithful Explainability for Legal Document Classification

    • Publication Time: 2026-10-02 12:00 Beijing Time
    • Abstract: - arXiv:2610.00045v1 Announcement Type: New.
      • Abstract: Transformer models such as LegalBERT are increasingly used in legal decision support, raising concerns about the fairness and transparency of model explanations.
      • These properties are typically evaluated separately, leaving it unclear whether debiasing interventions that alter fairness also change how faithfully explanations reflect the model’s reasoning.
      • This study investigates this relationship in the ECtHR alleged violations corpus from LexGLUE.
    • EN Highlights:
      • arXiv:2610.00045v1 Announce Type: new
      • Abstract: Transformer models such as LegalBERT are increasingly used in legal decision support, raising concerns about both fairness and the transparency of mod…
      • These properties are usually evaluated separately, leaving open whether a debiasing intervention that changes fairness also changes how faithfully explanations…
      • This study investigates that relationship on the ECtHR alleged-violations corpus from LexGLUE

  • The First Token Is Not the Verdict: Hidden Costs of Reading LLM Judges Without Generating

    • Publication Time: 2026-10-02 12:00 Beijing Time
    • Abstract: - arXiv:2610.00054v1 Announcement Type: New.
      • Abstract: Reading an LLM judge’s verdict from the logits of the first generated token is cheap, requires no generation, and is precisely what constrained decoding and likelihood scoring evaluation tools produce.
      • We show that this reading distorts positional bias in one direction: it exaggerates it in every condition we test, so the figures obtained this way act as an upper bound.
      • The mechanism is that the judge is not always led by the verdict token; for the three Qwen3 judges, this occurs on 12% to 49% of pairs, while for Llama-3.1-8B and Phi-3.5-mini, this figure is below 3%. Forcing a reading on these pairs returns the response that appears first, rather than the judgment.
    • EN Highlights:
      • arXiv:2610.00054v1 Announce Type: new
      • Abstract: Reading an LLM judge’s verdict from the logits of its first generated token is cheap, requires no generation, and is exactly what constrained decoding…
      • We show that this readout distorts position bias in one direction: it overstates it in every condition we test, so figures obtained this way behave as upper bou…

  • The mechanism is that judges do not always lead with a verdict token, on 12% to 49% of pairs for three Qwen3 judges and under 3% for Llama-3.1-8B and Phi-3.5-mi…

  • Measuring Human-Like Bias in LLMs? A Critique of Human-Derived Bias Constructs in LLM Evaluation

    • Publication Time: 2026-10-02 12:00 Beijing Time
    • Abstract: - arXiv:2610.00070v1 Announce Type: New.
      • Abstract: Researchers are increasingly using human-derived bias constructs to study Large Language Models (LLMs), including social-cognitive constructs like implicit bias and stereotype activation, as well as cognitive biases such as anchoring, framing effects, and confirmation bias.
      • Such approaches provide an alternative to overt bias probes, especially when direct questioning might obscure bias or when the model’s behavior is normatively acceptable.
      • However, adapting human bias constructs to LLMs introduces an inferential gap.
    • EN Highlights:
      • arXiv:2610.00070v1 Announce Type: new
      • Abstract: Researchers increasingly use human-derived bias constructs to study Large Language Models (LLMs), including social-cognitive constructs such as implic…
      • Such approaches offer alternatives to overt bias probes, particularly when direct questioning may obscure bias or when model behaviour appears normatively accep…
      • However, adapting human bias constructs to LLMs introduces an inferential gap

  • Legal text classification in Korean sexual offense cases: from traditional machine learning to large language models with XAI insights

    • Publication Time: 2026-10-02 12:00 Beijing Time
    • Abstract: - arXiv:2610.00087v1 Announce Type: New.
      • Abstract: The advancement of Natural Language Processing (NLP) has expanded AI-based text classification in the legal domain.
      • However, accurately classifying legal documents remains challenging due to the complexity of legal texts and subtle differences between legal categories.
      • This study evaluates legal text classification models, ranging from traditional machine learning techniques to Large Language Models (LLMs), using ten categories of South Korean sexual offense precedents.
    • EN Highlights:
      • arXiv:2610.00087v1 Announce Type: new
      • Abstract: The advancement of natural language processing (NLP) has expanded AI-based text classification in the legal domain
      • However, accurately classifying legal documents remains challenging due to the complexity of legal texts and subtle differences between legal categories
      • This study evaluates legal text classification models ranging from traditional machine learning techniques to large language models (LLMs) using ten categories…

  • BudgetSchemaBench: A Budget-Swept Diagnostic for Schema Context in Text-to-SQL

  • Publication Time: 2026-10-02 12:00 Beijing Time

    • Abstract: - arXiv:2610.00092v1 Announcement Type: new.
      • Abstract: Data agents over structured sources must fit the database schema into the model’s context window.
      • Large catalogs can span many databases and thousands of columns, so cost constraints may require choosing between table coverage and serialization detail before the context window is filled.
      • We introduce BudgetSchemaBench, an execution-based diagnostic for this setting.
    • EN Highlights:
      • arXiv:2610.00092v1 Announce Type: new
      • Abstract: Data agents over structured sources must fit database schema into the model’s context window
      • Large catalogs can span many databases and thousands of columns, so cost constraints may require choosing between table coverage and serialization detail well b…
      • We introduce BudgetSchemaBench, an execution-grounded diagnostic for this setting

  • FACET at WMT 2026 Automated Translation Quality Evaluation Task

    • Publication Time: 2026-10-02 12:00 Beijing Time
    • Abstract: - arXiv:2610.00096v1 Announcement Type: new.
      • Abstract: Different error types in machine translation require different evidence.
      • Whether the meaning is preserved can only be judged based on the source, while whether the target is well-formed, or whether it consistently names an entity, can be judged based on the target alone.
      • We present FACET, our reference-free submission to the WMT26 Automated Translation Quality Evaluation Task, which breaks down the evaluation into fluency, accuracy, and consistency stages, and provides only the necessary context for the error types in each stage.
    • EN Highlights:
      • arXiv:2610.00096v1 Announce Type: new
      • Abstract: Different error types in machine translation require different evidence
      • Whether meaning is preserved can be judged only against the source, while whether the target is well-formed, or whether it names one entity consistently, can be…
      • We present FACET, our reference-free submission to the WMT26 Automated Translation Quality Evaluation Task, which decomposes evaluation into Fluency, Accuracy,…

  • Metric-Construction Coupling Inflates Measured Synthetic Dialect Recovery

    • Publication Time: 2026-10-02 12:00 Beijing Time
    • Abstract: - arXiv:2610.00164v1 Announcement Type: new.
      • Abstract: Korean dialect corpora are available, but not redistributable: weights may be released, but training and evaluation cannot be reproduced from the underlying data.
      • We ask how much is recovered by supervised synthetic data and whether recovery can be measured independently of the synthesis pipeline.
      • We contribute KoDialectBench, 1,000 items from five regions on three axes, released as identifier hashes and scoring code so users can reconstruct items from their own licensed copies.
    • EN Highlights:
      • arXiv:2610.00164v1 Announce Type: new

  • Abstract: Korean dialect corpora are available but not redistributable: weights may be released, while reproducing training and evaluation from the underlying d…

    • We ask how much of that supervision synthetic data recovers, and whether that recovery can be measured independently of the synthesis pipeline
    • We contribute KoDialectBench, 1,000 items across five regions on three axes, released as identifier hashes and scoring code so users reconstruct the items from…

  • When a Data Artifact Isn’t a Shortcut: Causal Auditing of Synthetic RLVR Corpora

    • Publication Time: 2026-10-02 12:00 Beijing Time
    • Abstract: - arXiv:2610.00202v1 Announcement Type: New.
      • Abstract: Several recent pipelines build RLVR training data by masking a span of real corpus text and asking a language model to invent plausible wrong answers around it.
      • Therefore, the correct option is genuine human prose; every distractor is synthetic.
      • Correctness and provenance become entangled, and a policy could in principle learn the second instead of the first.
    • EN Key Points:
      • arXiv:2610.00202v1 Announce Type: new
      • Abstract: Several recent pipelines build RLVR training data by masking a span of real corpus text and asking a language model to invent plausible wrong answers…
      • The correct option is therefore genuine human prose; every distractor is synthetic
      • Correctness and provenance become entangled, and a policy could in principle learn the second instead of the first

  • A Holistic Assessment of the Carbon Footprint of Noor, a Very Large Arabic Language Model

    • Publication Time: 2026-10-02 12:00 Beijing Time
    • Abstract: - arXiv:2610.00223v1 Announcement Type: New.
      • Abstract: As ever larger language models grow more ubiquitous, it is crucial to consider their environmental impact.
      • Characterized by extreme size and resource use, recent generations of models have been criticized for their voracious appetite for compute, and thus their significant carbon footprint.
      • Although carbon impact reporting is becoming more common in machine learning papers, such reporting is often limited to the computational resources used strictly for training.
    • EN Key Points:
      • arXiv:2610.00223v1 Announce Type: new
      • Abstract: As ever larger language models grow more ubiquitous, it is crucial to consider their environmental impact
      • Characterised by extreme size and resource use, recent generations of models have been criticised for their voracious appetite for compute, and thus significant…

  • Although reporting of carbon impact has grown more common in machine learning papers, this reporting is usually limited to compute resources used strictly for t…

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

  • Reverse Item Response Theory for Sparsity-Robust Ranking in Fragmented Cancer Drug-Response Matrices

    • Publication Time: 2026-10-02 12:00 Beijing Time
    • Abstract: - arXiv:2610.00002v1 Announcement Type: New.
      • Abstract: We introduce reverse Item Response Theory (IRT) into pharmacogenomic drug-response analysis by treating cancer types as latent “subjects” with resistance, and drugs as “items” with evasion difficulty.
      • The model was applied to 242,036 drug sensitivity measurements in the Genomics of Drug Sensitivity in Cancer (GDSC2) database, estimating cancer-type-level in vitro drug resistance and drug-level broad activity within a shared latent space.
      • Validation across four missingness regimes shows that reverse IRT better recovers the full-data latent ranking than simple averaging, with advantages of Delta-rho = +0.089 to +0.095 in 60% missing cases under MCAR, cancer-biased, and drug-biased sparsity.
    • EN Key Points:
      • arXiv:2610.00002v1 Announce Type: new
      • Abstract: We introduce reverse Item Response Theory (IRT) to pharmacogenomic drug-response analysis by treating cancer types as latent “subjects” with resistanc…
      • Applied to 242,036 drug sensitivity measurements from the Genomics of Drug Sensitivity in Cancer (GDSC2) database, the model estimates cancer-type-level in-vitr…
      • Validation across four missingness regimes demonstrates that reverse IRT better recovers the full-data latent ranking than simple averaging, with advantages of…

  • How Far is Adam from Natural Gradient Descent?

    • Publication Time: 2026-10-02 12:00 Beijing Time
    • Abstract: - arXiv:2610.00004v1 Announcement Type: New.
      • Abstract: Adam is the standard optimizer in deep learning, but its geometric relationship to natural gradient descent (NGD) contains unresolved questions.
      • We investigate Adam’s full update rule, including momentum, as a diagonal empirical Fisher approximation affected by diagonal truncation, empirical label substitution, and temporal lag.
      • Using the scale-invariant $\gamma(\Delta\theta)$ metric, we measure Adam’s geometric deviation from true NGD in four loss scenarios: well-conditioned linear regression, ill-conditioned linear regression, logistic regression, and non-convex small neural networks.
    • EN Key Points:
      • arXiv:2610.00004v1 Announce Type: new
      • Abstract: Adam is the standard optimizer in deep learning, yet its geometric relationship to natural gradient descent (NGD) contains unresolved questions
      • We study Adam’s full update rule, including momentum, as a diagonal empirical Fisher approximation subject to diagonal truncation, empirical label substitution,…

  • Using the scale-invariant $\gamma(\Delta\theta)$ metric, we measure Adam’s geometric deviation from true NGD across four loss landscapes: well-conditioned linea…

  • FourierQK: Filter Shape, Admissibility and the Leakage-Coverage Law

    • Posted: 2026-10-02 12:00 Beijing Time
    • Abstract: - arXiv:2610.00009v1 Announce Type: new.
      • Abstract: Frequency-collapse attention [Zeris, 2026e] achieves large gains over standard dot-product attention by replacing the Q/K dot product with a bandpass-filtered inner product at learned frequencies.
      • A natural follow-up question is: which filter shape works best, and why?
      • We test five hypotheses about filter properties – DC suppression, Nyquist suppression, bandwidth, center frequency, and multi-scale coverage – using controlled ablations on a character-level language model (TinyShakespeare, 6-layer GPT).
    • EN Highlights:
      • arXiv:2610.00009v1 Announce Type: new
      • Abstract: Frequency-collapse attention [Zeris, 2026e] achieves large gains over standard dot-product attention by replacing the Q/K dot product with a bandpass-…
      • A natural follow-up question is: which filter shape works best, and why
      • We test five hypotheses about filter properties – DC suppression, Nyquist suppression, bandwidth, centre frequency, and multi-scale coverage – using a control…

  • Integrating Fairness and Explainability in a Multiple Instance Reinforcement Learning System

    • Posted: 2026-10-02 12:00 Beijing Time
    • Abstract: - arXiv:2610.00035v1 Announce Type: new. -Abstract: Predicting student performance from educational interaction data requires models that are both accurate and sufficiently transparent to support meaningful interventions, while demographic information introduces additional risks of unfair predictions.
      • This study investigates a multi-objective framework that combines reinforcement learning-based multiple instance learning (RL-MIL), adversarial debiasing, and preference-conditioned hypernetworks for student risk prediction.
      • MIL represents each student as a bag of weakly labeled interactions, while an RL agent selects informative instances for downstream classification.
    • EN Highlights:
      • arXiv:2610.00035v1 Announce Type: new
      • Abstract: Predicting student performance from educational interaction data requires models that are both accurate and sufficiently transparent to support meanin…
      • This study investigates a multi-objective framework that combines reinforcement learning-based multiple instance learning (RL-MIL), adversarial debiasing, and p…
      • MIL represents each student as a bag of weakly labeled interactions, while an RL agent selects informative instances for downstream classification

  • Fast Polynomial Transcendentals for LLMs

  • Publication Time: 2026-10-02 12:00 Beijing Time

    • Abstract: - arXiv:2610.00049v1 Announce Type: new.
      • Abstract: Graphics processing unit (GPU) generations scale matrix, special-function, and memory pipelines at different rates, so kernel bottlenecks change as hardware evolves.
      • FlashAttention-4 exposed an internal attention imbalance on NVIDIA Blackwell.
      • We test whether short polynomial programs can accelerate other special-function-unit (SFU) operations in large language models (LLMs).
    • EN Key Points:
      • arXiv:2610.00049v1 Announce Type: new
      • Abstract: Graphics processing unit (GPU) generations scale matrix, special-function, and memory pipelines at different rates, so kernel bottlenecks move as hard…
      • FlashAttention-4 exposed this imbalance inside attention on NVIDIA Blackwell
      • We test whether short polynomial programs can accelerate other special-function-unit (SFU) operations in large language models (LLMs)

  • SW-KAN: Kolmogorov-Arnold Networks with Stieltjes-Wigert q-Orthogonal Polynomials

    • Publication Time: 2026-10-02 12:00 Beijing Time
    • Abstract: - arXiv:2610.00050v1 Announce Type: new.
      • Abstract: Kolmogorov-Arnold Networks (KANs) represent a paradigm shift in deep learning, replacing fixed node activations with learnable univariate functions on the edges, thereby providing enhanced interpretability and parameter efficiency.
      • While recent polynomial-based KAN variants have addressed the computational overhead of the original B-spline implementations, they introduce a fundamental yet underexplored challenge: the domain mismatch between unbounded real-valued inputs and the bounded or semi-infinite support of orthogonal polynomial bases.
      • To address this limitation, we propose the Stieltjes-Wigert Kolmogorov-Arnold Network (SW-KAN), a novel architecture that employs Stieltjes-Wigert q-orthogonal polynomials defined on the semi-infinite domain (0, infinity).
    • EN Key Points:
      • arXiv:2610.00050v1 Announce Type: new
      • Abstract: Kolmogorov-Arnold Networks (KANs) represent a paradigmatic shift in deep learning by replacing fixed node activations with learnable univariate functi…
      • While recent polynomial-based KAN variants have addressed the computational overhead of original B-spline implementations, they introduce a fundamental yet unde…
      • To address this limitation, we propose the Stieltjes-Wigert Kolmogorov-Arnold Network (SW-KAN), a novel architecture that employs Stieltjes-Wigert q-orthogonal…

  • Format-Aware Fusion for Fast FP4 Pretraining

    • Publication Time: 2026-10-02 12:00 Beijing Time
    • Abstract: - arXiv:2610.00053v1 Announce Type: new.
      • Abstract: Four-bit floating-point (FP4) tensor cores accelerate matrix multiplication, but scale computation, operand packing, layout construction, and saving backward states can erase the gains.

  • We propose \emph{format-aware fusion}, which co-designs each quantization producer with its scale domain and consumer layout for native \mxfp{}, global \nvfp{}, and collaborative thread array local \nvfp{}.

  • We evaluate Llama-3-family 8B pretraining through 160 billion tokens using bfloat16 output projections and compiled cross entropy.

  • EN Highlights:

    • arXiv:2610.00053v1 Announce Type: new
    • Abstract: Four-bit floating-point (FP4) Tensor Cores accelerate matrix multiplication, but scale computation, operand packing, layout construction, and saved ba…
    • We present \emph{format-aware fusion}, which co-designs each quantization producer with its scale domain and consumer layout for native \mxfp{}, global \nvfp{},…
    • We evaluate Llama-3-family 8B pretraining through 160 billion tokens using bfloat16 output projections and compiled cross entropy

  • “very likely” Means “uncertain”? How LLMs Diverge from Humans in Linguistic Uncertainty Quantification

    • Release Time: 2026-10-02 12:00 Beijing Time
    • Abstract: - arXiv:2610.00083v1 Announce Type: new.
      • Abstract: Humans express uncertainty verbally via markers (e.g., “possible,” “likely”), yet most LLM uncertainty quantification (UQ) relies on signals based on cost likelihood or consistency.
      • From a cognitive perspective, accurate verbal uncertainty reflects metacognitive monitoring, representing knowledge boundaries (“knowing that you don’t know”) to support regulation and information seeking.
      • In this paper, we investigate how LLMs diverge from humans in verbal uncertainty quantification and whether verbal markers can reliably quantify LLM uncertainty.
    • EN Highlights:
      • arXiv:2610.00083v1 Announce Type: new
      • Abstract: Humans express uncertainty verbally via markers (e.g., “possible,” “likely”), yet most LLM uncertainty quantification (UQ) relies on costing likelihoo…
      • From a cognitive perspective, accurate verbal uncertainty reflects metacognitive monitoring, representing knowledge boundaries (“knowing that you don’t know”) t…
      • In this paper, we investigate how LLMs diverge from humans in verbal uncertainty quantification and whether verbal markers can reliably quantify LLM uncertainty

  • Nous: Learning and Certifying Memory Decisions Before Source Calibration

    • Release Time: 2026-10-02 12:00 Beijing Time
    • Abstract: - arXiv:2610.00094v1 Announce Type: new.
      • Abstract: Belief-based agent memory requires making reliable decisions about the current state, but its evidence can be noisy, duplicated, or stale.
      • Must memory first calibrate its sources to improve its decisions?
      • We separate learning, calibration, and revision certification.
    • EN Highlights:
      • arXiv:2610.00094v1 Announce Type: new

  • Abstract: Belief-based agent memory needs reliable decisions about current state, yet its evidence may be noisy, copied, or stale

  • Must a memory calibrate its sources before it can improve its decisions

  • We separate learning, calibration, and revision certification

  • One Mastery Threshold Does Not Fit All Knowledge Tracing Models

    • Publication Time: 2026-10-02 12:00 Beijing Time
    • Abstract: - arXiv:2610.00095v1 Announcement Type: new.
      • Abstract: Tutoring systems use mastery thresholds to decide when students can stop practicing and move on, but when the underlying knowledge tracing (KT) model changes, the same numerical threshold can lead to very different decisions.
      • We examined six KT models in four public educational datasets and evaluated 12 thresholds from 0.50 to 0.99 using post-advancement performance, advancement coverage, practice burden, and differences between prior performance groups.
      • We also identified thresholds that balance performance, extra practice, and advancement under 30 predefined instructional settings.
    • EN Key Points:
      • arXiv:2610.00095v1 Announce Type: new
      • Abstract: Tutoring systems use mastery thresholds to decide when students can stop practicing and advance, but the same numerical threshold can lead to very dif…
      • We examine six KT models across four public educational datasets and evaluate 12 thresholds from 0.50 to 0.99 using post-advancement performance, advancement co…
      • We also identify thresholds that balance performance, extra practice, and advancement under 30 predefined instructional settings