System translated (Gemini)

🤖 AI 速览

Today, there are three main themes: The topic of superintelligence has entered high-level US-China dialogue, intensifying competition in AI governance and geopolitics; Anthropic has opened its Claude plugin submission portal, shifting the model competition toward ecosystems and platformization; and …
📋 文章元数据
发布时间
2026-09-26
类型
ai-daily
字数
5687
阅读时长
27 min

2026-09-26 AI Daily | Superintelligence Enters High-Level U.S.-China Dialogue, Agents Shift from Efficiency Gains to Reliability Scrutiny Link to heading

Today’s three main themes are: The topic of superintelligence is entering high-level U.S.-China talks, heating up AI governance and geopolitical competition; Anthropic is opening its Claude plugin submission portal, shifting the model competition towards ecosystems and platformization; The Watch List points to agent reliability, with Proaction’s sales efficiency gains and predictive agent stress tests showing that the deeper the automation, the more indispensable supervision and calibration become.

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

Two main themes are worth a deep dive today. First, the narrative around agents is shifting from efficiency gains to a focus on reliability: An OpenAI blog post revealed Proaction used Codex to increase sales by 60% and save over 75 hours; Stratechery this week pointed out that agents are well-suited for handling routine tasks. However, papers on arXiv regarding reward hacking in autonomous research agents and stress-testing predictive agent behavior routing serve as a reminder that the deeper the automation, the more critical supervision and calibration become. Second, model behavior auditing and fairness: misaligned incentives can lead witnesses to defeat contextual grounding, post-training compression of Whisper can redistribute the burden of errors, and COILD reinforces machine translation for Indic languages. These are topics for alignment, speech, and multilingual evaluation teams to follow up on.

🌐 AI Hot Topics on X Link to heading

Topic 1: Trump Pitches ‘Super Intelligence’ to Xi at White House Summit Link to heading

  • Category: AI · Other
  • Overview: Trending for: 1 day ago, Related posts: 30,000
  • What happened: Trump raised issues or proposed cooperation related to “superintelligence” with Xi Jinping during the White House summit.
  • Why it matters: This indicates that superintelligence has become a topic of high-level diplomacy between major powers, potentially impacting the landscape of global AI governance, technological competition, and U.S.-China cooperation.
  • Discussion summary: The main focus on X is whether the U.S. and China will cooperate or compete in the field of superintelligence. Topics also include AI chip export controls and the deployment of AI chips in space. The disagreement centers on technological nationalism versus global governance.

Topic 2: Anthropic Launches Plugin Submission Portal for Claude Developers Link to heading

  • Category: AI · News
  • Overview: Trending for: 2 hours ago, Related posts: 308
  • What happened: Anthropic has launched a plugin submission portal for Claude developers, opening up access for third-party plugins.
  • Why it matters: This marks a shift for Claude from competing on model capabilities to expanding its ecosystem. The plugin ecosystem could be a significant watershed moment for the platformization of AI assistants.
  • Discussion summary: Discussions on X are mainly focused on comparing this move with OpenAI’s plugin ecosystem, the standards for plugin review, the entry barriers for developers, and whether plugins can significantly enhance Claude’s practicality and commercial value.

Topic 3: Zuckerberg vs Musk Meme Fuels Muse and Grok Bot Debate Link to heading

  • Category: AI · News
  • Overview: Trending for: 23 hours ago, Related posts: 400
  • What happened: A comparison meme featuring Zuckerberg and Musk has sparked discussions about the Muse and Grok chatbots.
  • Why it matters: This event highlights issues of stance, bias, and platform competition when AI bots handle controversial content about tech leaders, which affects the credibility of AI-generated content.
  • Discussion summary: The debate on X primarily revolves around which bot, Muse or Grok, provides a more objective response, whether the meme was amplified or distorted by AI, and how AI should respond to sensitive topics involving public figures.

Summary of AI Public Opinion on X Today Link to heading

The main theme in public opinion today is that AI’s influence has expanded from model capability competition to geopolitics, platform ecosystems, and content credibility: superintelligence is entering high-level U.S.-China talks, Claude is opening its plugin ecosystem, and Muse and Grok are being compared over a controversial meme. The consensus is that AI is now seen as a core variable in the competition between major powers and platforms, requiring clearer governance and rules. Points of contention include whether the U.S. and China should cooperate or compete, whether plugin access should prioritize efficiency or review, and whether AI should remain neutral or reflect a platform’s stance when dealing with controversial figures. Potential risks include an escalating superintelligence race and chip rivalry, security vulnerabilities from third-party plugins, and AI bias amplifying misleading content and damaging public trust.

💡 Influencer Insights Link to heading

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

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

Time frame: Last 3 days; covers 22 sources; 32 updates in total

Stratechery by Ben Thompson (A_full) Link to heading

  • 2026.39: Begun, the Aggregator Wars Have
    • Publication time: 2026-09-26 01:00 Beijing Time
    • Summary: - Welcome back to this week’s Stratechery
  • As a reminder, every Friday, we send an overview of the content in the Stratechery bundle; highlighted links are free for everyone.
  • Additionally, you have complete control over the content we send you.
  • With that said, here are a few of our favorites from this week.
  • Amazon. ** One of the best use cases for agents is handling necessary but annoying daily tasks.

OpenAI Blog (A_full) Link to heading

  • Proaction boosts sales 60% and saves 75+ hours with Codex
    • Published: 2026-09-26 03:00 Beijing Time
    • Summary: - With Codex, GPT-Live-1, and GPT-6 Astra, Proaction can build, operate, and sell modern fleet management faster.
      • This article from the OpenAI Blog explains how Proaction is shaping the broader AI and infrastructure landscape with Codex, boosting sales by 60% and saving over 75 hours.
      • Proaction has increased sales by 60% and saved over 75 hours with Codex, which also has real-world implications for founders, operators, and investors.

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

  • When Should Forecasting Agents Reason? Behavioral Stress Tests for Reliability Routing

    • Published: 2026-09-25 12:00 Beijing Time
    • Summary: - arXiv:2609.28475v1 Announcement Type: new.
      • Abstract: Forecasting agents are increasingly combining language model reasoning, retrieval, ensembling, and calibration, but it remains unclear when each behavior should be trusted.
      • We study this issue on ForecastBench-style binary forecasting tasks, treating the choices of retrieval, reasoning, and deferring to market priors as observable agent behaviors rather than hidden implementation details.
      • Our main finding is that mechanism selection depends on the source: structured analogs are dominant in certain data generation processes, while market/crowd styles and conservative baselines are better for others.

  • Our central finding is that mechanism choice is source-dependent: structured analogs dominate for some data-generating processes, while market/crowd-style and c…

  • TW3Cast: A Frozen Router of Lightly Fine-Tuned Foundation Models for Time-Series Forecasting on GIFT-Eval, Selected Entirely on the Training Split

    • Release Time: 2026-09-25 12:00 Beijing Time
    • Abstract: - arXiv:2609.28506v1 Announcement Type: new.
      • Abstract: TW3Cast is a time-series forecasting system that ranks 3rd out of 130 entries on the GIFT-Eval benchmark by mean MASE rank as of September 14, 2026.
      • The two entries above it belong to the leaderboard’s agentic category, which are multi-step systems that use agents or language models to reason about, generate, or select forecasts.
      • TW3Cast runs no agent and no language model.
    • EN Highlights:
      • arXiv:2609.28506v1 Announce Type: new
      • Abstract: TW3Cast is a time-series forecasting system that reaches position 3 of 130 entries on the GIFT-Eval benchmark by mean MASE rank, as of 2026-09-14
      • The two entries above it belong to the leaderboard’s agentic category, multi-step systems that use agents or language models to reason about, generate or select…
      • TW3Cast runs no agent and no language model

  • PAWS: Policy-driven Agentic World Simulation

    • Release Time: 2026-09-25 12:00 Beijing Time
    • Abstract: - arXiv:2609.28547v1 Announcement Type: new.
      • Abstract: Policy interventions propagate through public communication, institutional decisions, and stakeholder responses, yet datasets for financial multi-agent simulations rarely link these processes with temporally consistent historical evidence.
      • We introduce PAWS, a Policy-driven Agentic World Simulation dataset, covering 36 verified U.S.
      • financial and economic policy episodes, 12,727 policy-related news records, and 65,291 source-based stakeholder actions.
    • EN Highlights:
      • arXiv:2609.28547v1 Announce Type: new
      • Abstract: Policy interventions propagate through public communication, institutional decisions, and stakeholder responses, yet datasets for financial multi-agen…
      • We introduce PAWS, a Policy-driven Agentic World Simulation dataset covering 36 verified U.S
      • financial and economic policy episodes, 12,727 policy-linked news records, and 65,291 source-grounded stakeholder actions

  • Pistis Technical Report

    • Release Time: 2026-09-25 12:00 Beijing Time
    • Abstract: - arXiv:2609.28554v1 Announcement Type: new.
      • Abstract: We introduce the Pistis model series, which includes 27B and 9B parameter multimodal large language models built on Qwen3.6 and Qwen3.5 respectively, and developed through a universal and scalable post-training framework.

  • The framework first establishes a strong foundation through large-scale multimodal supervised fine-tuning (SFT).

  • Building on this SFT foundation, we propose Interleaved Distillation and Reinforcement Learning (IDRL), a novel post-training paradigm that tightly integrates policy distillation and reinforcement learning within a single training loop.

    • EN Highlights:
      • arXiv:2609.28554v1 Announce Type: new
      • Abstract: We introduce the Pistis model family, comprising 27B- and 9B-parameter multimodal large language models built on Qwen3.6 and Qwen3.5, respectively, an…
      • The framework first establishes a strong foundation through large-scale multimodal supervised fine-tuning (SFT)
      • Building on this SFT foundation, we propose Interleaved Distillation and Reinforcement Learning (IDRL), a novel post-training paradigm that tightly integrates o…

  • BaseCamp — An Agentic AI Framework for Automating DNA Sequencing Data Pipelines

    • Publication Time: 2026-09-25 12:00 Beijing Time
    • Abstract: - arXiv:2609.28557v1 Announce Type: new.
      • Abstract: DNA sequencing pipelines, spanning quality control, alignment, variant calling, and annotation, are now reliably executed by workflow management systems that can coordinate established bioinformatics tools at scale.
      • What remains manual is the decision layer surrounding that execution: selecting quality thresholds appropriate to a sample and platform, adjudicating borderline variant calls, diagnosing anomalies, and determining which results merit expert review.
      • These decisions are repetitive, judgment-intensive, inconsistent across operators, and frequently undocumented.
    • EN Highlights:
      • arXiv:2609.28557v1 Announce Type: new
      • Abstract: DNA sequencing pipelines, spanning quality control, alignment, variant calling, and annotation, are now reliably executed by workflow management syste…
      • What remains manual is the decision layer surrounding that execution: selecting quality thresholds appropriate to a sample and platform, adjudicating borderline…
      • These decisions are repetitive, judgment-intensive, inconsistent across operators, and frequently undocumented

  • DEEPO: Dual-Entropy Enhanced Policy Optimization for Hallucination in MLLMs

    • Publication Time: 2026-09-25 12:00 Beijing Time
    • Abstract: - arXiv:2609.28570v1 Announce Type: new.
      • Abstract: Reinforcement Learning (RL) is widely used to improve the reasoning capabilities of Multimodal Large Language Models (MLLMs), but its effectiveness on hallucination is not uniform.
      • We trace this back to two weaknesses in the \emph{correction chain} from reward to parameter update.
      • At the rollout level, hard queries (queries with high semantic entropy) often produce consistently erroneous sample groups, causing a relative collapse of the group.
    • EN Highlights:
      • arXiv:2609.28570v1 Announce Type: new

  • TWIST: A Proposed Benchmark for Intervention Quality in Conversational Memory, with a Human-Validated Draft-Alignment

    • Published: 2026-09-25 12:00 Beijing Time
    • Abstract: - arXiv:2609.28575v1 Announcement Type: new.
      • Abstract: Long-conversation memory benchmarks increasingly test recall and prompted knowledge updates, and recent work studies evolving user beliefs and memory states.
      • TWIST is a proposed benchmark suite for a complementary, unmeasured property: intervention quality – whether a deployed memory system, exercised through its own intake/recall/review surfaces, operates correctly at points of belief change.
      • Four tracks cover unprompted tension detection, vetting outgoing drafts against the record, answering with current beliefs while preserving supersession history, and managing sensitive recall.
    • EN Key Points:
      • arXiv:2609.28575v1 Announce Type: new
      • Abstract: Long-conversation memory benchmarks increasingly test recall and prompted knowledge updates, and recent work studies evolving user beliefs and memory…
      • TWIST is a proposed benchmark suite for a complementary, unmeasured property: intervention quality – whether a deployed memory system, exercised through its ow…
      • Four tracks cover unprompted tension detection, vetting outgoing drafts against the record, answering with current beliefs while preserving supersession history…

  • Adversarial Closed-Loop Curriculum for Evolving Role-Playing Agents

    • Published: 2026-09-25 12:00 Beijing Time
    • Abstract: - arXiv:2609.28609v1 Announcement Type: new.
      • Abstract: Role-playing agents based on large language models have been widely applied in areas such as personalized assistance and social simulation.
      • Recent reinforcement learning methods typically train on a fixed scenario pool collected before learning begins.
      • This creates a distribution bottleneck: as the agent improves, the scenarios where it underperforms also change, while the training distribution remains static.
    • EN Key Points:
      • arXiv:2609.28609v1 Announce Type: new
      • Abstract: Role-playing agents based on large language models have been widely applied in areas such as personalized assistance and social simulation
      • Recent RL methods typically train on a fixed scenario pool collected before learning begins

  • This creates a distributional bottleneck: as the agent improves, the scenarios where it performs poorly also change, while the training distribution remains sta…

  • Training Object Permanence in World Models

    • Posted: 2026-09-25 12:00 Beijing Time
    • Abstract: - arXiv:2609.28654v1 Announcement Type: New.
      • Abstract: The persistence and solidity of objects are hallmarks of human cognitive priors.
      • Recent research shows that video generation models, a paradigm of current world models, have begun to show emerging reasoning abilities, making them ideal candidates for building human-like physical intelligence.
      • Has object permanence already emerged in video models?
    • EN Highlights:
      • arXiv:2609.28654v1 Announce Type: new
      • Abstract: Object permanence and solidity are hallmarks of human cognitive priors
      • Recent studies show that video generation models, a paradigmatic class of current world models, have begun to show emerged reasoning abilities, making them idea…
      • Do video models have emerged object permanence in them

  • Beyond Surface Style: Aligning Multi-Turn User Simulators with Behavioral Consistency

    • Posted: 2026-09-25 12:00 Beijing Time
    • Abstract: - arXiv:2609.28690v1 Announcement Type: New.
      • Abstract: Faithful user simulation is the foundation for building, evaluating, and improving interactive artificial intelligence at scale.
      • However, plausible individual responses do not ensure that simulated users reproduce the intent evolution and outcomes observed in real interactions.
      • We propose TRACER, a multi-turn user simulator that explicitly models the user’s changing intent and learns to align simulated behavior with real interaction trajectories.
    • EN Highlights:
      • arXiv:2609.28690v1 Announce Type: new
      • Abstract: Faithful user simulation is fundamental to building, evaluating, and improving interactive AI at scale
      • However, plausible individual responses do not ensure that simulated users reproduce the intent evolution and outcomes observed in real interactions
      • We propose TRACER, a multi-turn user simulator that explicitly models users’ evolving intent and learns to align simulated behavior with real interaction trajec…

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

  • Framing by Wording, Framing by Selection: A Large-Scale Two-Dimensional Audit of French News Headlines, 2022-2025

    • Posted: 2026-09-25 12:00 Beijing Time
    • Abstract: - arXiv:2609.28487v1 Announcement Type: New.
      • Abstract: News headlines frame public issues through their choice of content and wording, but computational framing work often breaks these operations down into a single score.

  • We introduce a two-dimensional framework that separates the salience framework from the selection framework. The salience framework is measured through four wording devices (loaded vocabulary, blame attribution, threat framing, rhetorical questions), while the selection framework is measured through outlet-level story formats and high-charge distribution.

  • We build a 10,000-headline French supervision set using three LLM annotators with majority-vote resolution and human arbitration, validate the labels against two independent blind studies, and apply the strongest classifier to 902,111 deduplicated headlines from 25 French media outlets (2022-2025).

    • EN Highlights:
      • arXiv:2609.28487v1 Announce Type: new
      • Abstract: News headlines frame public issues both by what they select and by how they word it, yet computational framing work typically collapses these operatio…
      • We introduce a two-dimensional framework that separates salience framing, measured through four wording devices (loaded vocabulary, blame attribution, threat fr…
      • We build a 10,000-headline French supervision set using three LLM annotators with majority-vote resolution and human arbitration, validate the labels against tw…

  • Reward Hacking Challenges Oversight of Autonomous Research Agents

    • Release Time: 2026-09-25 12:00 Beijing Time
    • Abstract: - arXiv:2609.28614v1 Announce Type: new.
      • Abstract: Autonomous research agents can design experiments, evaluate results, and write reports, giving them control over both scientific results and the evidence that supports them.
      • This creates a risk of reward hacking: meeting the reward criteria without achieving the intended goal.
      • We study (1) how often models reward-hack without instructions to do so, (2) how effective and detectable their methods are when hacking is allowed, and (3) how they adapt when an LLM review panel returns its decision and reasoning.
    • EN Highlights:
      • arXiv:2609.28614v1 Announce Type: new
      • Abstract: Autonomous research agents can design experiments, evaluate results, and write reports, giving them control over both a scientific result and the evid…
      • This creates a risk of reward hacking: meeting the reward criteria without achieving the intended goal
      • We study (1) how often models reward-hack without instructions to do so, (2) how effective and detectable their methods are when hacking is allowed, and (3) how…

  • Benchmarking Argumentative Behaviour of LLMs: A Study of Defences Against Character Attacks

    • Release Time: 2026-09-25 12:00 Beijing Time
    • Abstract: - arXiv:2609.28673v1 Announce Type: new.
      • Abstract: Large Language Models (LLMs) are increasingly used as argumentative agents in persuasive dialogues, necessitating a rigorous evaluation of their debating capabilities relative to human interlocutors.
      • In this study, we focus on character attacks (ad hominem), which are traditionally seen as fallacies, yet they play a key role in persuasive political dialogues where ethos often rivals propositional content.

  • Specifically, we investigate whether modern LLMs can replicate human competence to strategically use and respond to such attacks.

    • EN 要点:
      • arXiv:2609.28673v1 Announce Type: new
      • Abstract: Large Language Models (LLMs) are increasingly deployed as argumentative agents in persuasive dialogues, necessitating rigorous evaluation of their deb…
      • In this study, we focus on character attacks (ad hominem arguments), traditionally dismissed as fallacies, which play a pivotal role in political persuasive dia…
      • Specifically, we investigate whether modern LLMs can replicate human competence to strategically use and respond to such attacks

  • An Explainable DistilBERT-BiLSTM-Attention Framework for Binary and Multi-Class Hate Speech Detection

    • Publication Time: 2026-09-25 12:00 Beijing Time
    • Abstract: - arXiv:2609.28703v1 Announce Type: new.
      • Abstract: Hate speech on social media poses serious risks to social harmony, mental well-being, and public safety, making timely and accurate detection crucial for content moderation systems.
      • Most existing studies focus on binary classification, evaluate their frameworks on a single dataset, and provide limited insight into how decisions are made, which limits their real-world applicability.
      • Furthermore, limited work has been done on the explainability of their predictive inference.
    • EN 要点:
      • arXiv:2609.28703v1 Announce Type: new
      • Abstract: Hate speech on social media poses serious risks to social harmony, mental well-being, and public safety, making its timely and accurate detection esse…
      • Most existing studies focus on binary classification, evaluated their frameworks on a single dataset, and provide limited insight into how decisions are made, w…
      • In addition, limited work is done on the explainability of their predictive inference

  • PTC-Bias: Phoneme-Level Temporal Competition for Bias Retrieval and Post-Decoding Correction in Speech LLMs

    • Publication Time: 2026-09-25 12:00 Beijing Time
    • Abstract: - arXiv:2609.28727v1 Announce Type: new.
      • Abstract: Contextual bias can improve rare word recognition in Speech Large Language Models (SpeechLLM), but effectively utilizing large bias lists remains challenging.
      • We propose PTC-Bias, a two-stage framework based on phoneme-level temporal competition.
      • In the pre-filling stage, PTC Retrieval performs temporal competition between frame-synchronous phoneme decoding and candidate pronunciations, generating a compact list of biased word candidates and corresponding speech intervals.
    • EN 要点:
      • arXiv:2609.28727v1 Announce Type: new

  • Abstract: Contextual biasing improves rare-word recognition in speech large language models (SpeechLLMs), but efficiently exploiting large bias lists remains ch…

  • We propose PTC-Bias, a two-stage framework based on phoneme-level temporal competition

  • At the prefill stage, PTC Retrieval performs frame-synchronous phoneme decoding and temporal competition among candidate pronunciations, producing a compact bia…

  • Temporal Taxation Compounds Under Post-Training Compression of Whisper Models

    • Published: 2026-09-25 12:00 Beijing Time
    • Abstract: - arXiv:2609.28739v1 Announce Type: new.
      • Abstract: Automatic speech recognition models are audited for demographic fairness at full precision, yet the models that ship to production have been quantized, pruned, and distilled.
      • We ask whether post-training weight compression, which alters model weights rather than the audio signal or its feature representation, redistributes the error burden across demographic groups.
      • Across the Whisper family on Fair-Speech, Common Voice 25, and AfriSpeech-200, 50% Wanda pruning of Whisper-large-v3 sharply widens the Black/AA-vs-Asian temporal taxation gap on Fair-Speech: the absolute word error rate gap between the worst- and best-served groups more than doubles; assuming 5 seconds of correction time per transcription error, this increases correction time per minute of speech from 30 to 64 seconds.
    • EN Highlights:
      • arXiv:2609.28739v1 Announce Type: new
      • Abstract: Automatic speech recognition models are audited for demographic fairness at full precision, yet the models that ship to production have been quantized…
      • We ask whether post-training weight compression, which alters model weights rather than the audio signal or its feature representation, redistributes error burd…
      • Across the Whisper family on Fair-Speech, Common Voice 25, and AfriSpeech-200, 50% Wanda pruning of Whisper-large-v3 sharply widens the Black/AA-vs-Asian tempor…

  • Technical Manual for Toolkit for Confidence-Corpus Consistency via Fine-Tuning on a Fabricated Corpus

    • Published: 2026-09-25 12:00 Beijing Time
    • Abstract: - arXiv:2609.28747v1 Announce Type: new.
      • Abstract: A language model’s confidence in an answer is often interpreted as a proxy for its knowledge of the corresponding fact.
      • This manual documents an open toolkit built for a direct test of reading-in: fine-tuning a small causal language model on a corpus that consistently asserts a fabricated arithmetic answer for each of 81 single-digit addition pairs, and comparing its post-fine-tuning confidence in each fabricated answer to its pre-fine-tuning confidence in the corresponding true answer, using an invariant measurement procedure throughout.
      • We describe and justify each pipeline stage, fact-space generation, tokenization-length-aware confidence measurement, baseline validation, corpus construction, fine-tuning, and paired before-and-after comparison, and each confound to be ruled out, which include tokenization asymmetries between one- and two-digit answers and the difference between answers that merely lose salience and answers that are actively suppressed.
    • EN Highlights:

  • arXiv:2609.28747v1 Announce Type: new

  • Abstract: A language model’s confidence in an answer is often read as a proxy for how well it knows the corresponding fact

  • This manual documents an open toolkit built to test that reading directly: a small causal language model is fine-tuned on a corpus that consistently asserts one…

  • We describe and justify every pipeline stage, fact-space generation, token-length-aware confidence measurement, baseline validation, corpus construction, fine-t…

  • Script Choice in LLMs: Evidence for Late-Layer Commitment

    • Publication Time: 2026-09-25 12:00 Beijing Time
    • Abstract: - arXiv:2609.28784v1 Announce Type: new.
      • Abstract: In this paper, we use two complementary interpretability methods, logistic regression probing and logit-lens analysis, to study how script knowledge is distributed across the layers of LLMs.
      • Our probing experiments reveal a clear asymmetry: both the input script and the instructed output script are encoded in the earliest layers of the network, while in contrast, the commitment to the actual output script only appears in the final layers, with the model’s intermediate representations defaulting to Latin in most layers.
      • This two-stage process is confirmed by logit-lens analysis, which shows that script commitment consistently occurs in the final layers of the LLMs.
    • EN Highlights:
      • arXiv:2609.28784v1 Announce Type: new
      • Abstract: In this paper, we investigate how script knowledge is distributed across the layers of LLMs using two complementary interpretability methods: logistic…
      • Our probing experiments reveal a clear asymmetry: both the input script and the instructed output script are encoded in the earliest layers of the network, whil…
      • This two-stage process is confirmed by logit-lens analyses, which show that script commitment consistently occurs at the very last layers of the LLMs

  • COILD: An Indic-Centric Parallel Corpus and Benchmark for Machine Translation Across Indian Languages

    • Publication Time: 2026-09-25 12:00 Beijing Time
    • Abstract: - arXiv:2609.28826v1 Announce Type: new.
      • Abstract: Machine Translation (MT) for Indian languages is still limited by the finite availability of high-quality, Indic-centric parallel corpora and evaluation benchmarks.
      • Existing multilingual resources are primarily built from English-centric content and often fail to capture the linguistic diversity, cultural complexities, and domain-specific characteristics of Indian languages.
      • We introduce COILD, an Indic-centric parallel corpus containing over 1.16 million manually translated and human-verified sentence pairs, covering 20 Indian language pairs from the Indo-Aryan, Dravidian, Tibeto-Burman, and Austroasiatic language families.
    • EN Highlights:
      • arXiv:2609.28826v1 Announce Type: new

  • Abstract: Machine translation (MT) for Indian languages remains constrained by the limited availability of high-quality, Indic-centric parallel corpora and eval…

  • Existing multilingual resources are largely constructed from English-pivot content and often fail to capture the linguistic diversity, cultural complexity, and…

  • We present COILD, an Indic-centric parallel corpus comprising over 1.16 million human-translated and human-verified sentence pairs, covering 20 Indian language…

  • Persuaded, Not Informed: Incentive-Misaligned Witnesses Defeat In-Context Grounding

    • Publication Time: 2026-09-25 12:00 Beijing Time
    • Abstract: - arXiv:2609.28854v1 Announcement Type: new.
      • Abstract: Language model agents are increasingly answering questions over customer-relationship management (CRM) records, such as whether to qualify a sales lead.
      • We identify a failure mode not addressed by a stronger model: when the context contains an assertion by a party with an incentive toward optimism - here the sales representative, a witness recorded in the CRM - the model treats the assertion as evidence and clears deals that the company’s own records would deem unacceptable.
      • Across 100 lead-qualification tasks from CRMArena-Pro, the representative asserts an acceptable timeline in every call and an acceptable budget in 76; on the 31 tasks where such assertions contradicted the price list and installation policies, a model that only reads transcripts cleared the deal in 29 of the 31 tasks.
    • EN Key Points:
      • arXiv:2609.28854v1 Announce Type: new
      • Abstract: Language-model agents increasingly answer questions over customer-relationship management (CRM) records, such as whether to qualify a sales lead
      • We identify a failure mode not addressed by a stronger model: when the context contains an assertion by a party with an incentive toward optimism - here the sal…
      • Across 100 lead-qualification tasks from CRMArena-Pro, the representative asserts an acceptable timeline in every call and an acceptable budget in 76; on the 31…

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

  • Stable and Faithful Explanations for Knowledge Tracing

    • Publication Time: 2026-09-25 12:00 Beijing Time
    • Abstract: - arXiv:2609.28502v1 Announcement Type: new.
      • Abstract: Knowledge Tracing (KT) models opaquely predict student performance, limiting instructional actions.
      • This study provides a validation protocol to jointly test predictive competitiveness (RQ1), explanation stability (RQ2), and retraining-based faithfulness (RQ3).
      • Thirteen behavioral features spanning 5 instructional topics were designed from ASSISTments 2009 and 2012, with historical features computed from prior interactions and current response latency reserved for retrospective analysis only.
    • EN Key Points:
      • arXiv:2609.28502v1 Announce Type: new

  • Abstract: Knowledge tracing (KT) models predict student performance opaquely, limiting pedagogical action

  • This study contributes a validation protocol testing predictive competitiveness (RQ1), explanation stability (RQ2) and retraining-based faithfulness (RQ3) toget…

  • Thirteen behavioral features across five pedagogical themes were engineered from ASSISTments 2009 and 2012, with history features computed from temporally prece…

  • SMILESGNN: Interpretable Clinical Toxicity Prediction via SMILES-Graph Cross-Attention Fusion

    • Publication Time: 2026-09-25 12:00 Beijing Time
    • Abstract:- arXiv:2609.28553v1 Announce Type: new.
      • Abstract: Drug toxicity prediction is critical for reducing late-stage attrition in drug discovery, yet remains challenging due to severe class imbalance, scaffold-based generalization, and the clinical need for interpretable predictions.
      • Single-modality approaches - SMILES Transformers or graph neural networks capture complementary aspects of molecular structure, while sequence-only models cannot directly provide graph attribute explanations.
      • We propose SMILESGNN, a multimodal architecture that fuses a SMILES Transformer encoder and a GATv2 graph encoder via cross-attention, and SMILESGNN-PT, a variant using a ChemBERTa-2 pre-trained backbone.
    • EN 要点:
      • arXiv:2609.28553v1 Announce Type: new
      • Abstract: Drug toxicity prediction is critical for reducing late-stage attrition in drug discovery, yet remains challenging due to severe class imbalance, scaff…
      • Single-modality approaches-SMILES Transformers or graph neural networks capture complementary aspects of molecular structure, while sequence-only models cannot…
      • We present SMILESGNN, a multimodal architecture that fuses a SMILES Transformer encoder and a GATv2 graph encoder via cross-attention, and SMILESGNN-PT, a varia…

  • CFD Correction of Open Tip Clearance Flow in a Compressor Cascade Using VAE Latent Space Adaptation

    • Publication Time: 2026-09-25 12:00 Beijing Time
    • Abstract:- arXiv:2609.28558v1 Announce Type: new.
      • Abstract: CFD predictions of open tip clearance flow in a compressor cascade show discrepancies with experiments, while experimental observations are scarce and high-resolution experimental ground truth is unavailable.
      • This study proposes a non-invasive correction method based on variational autoencoders (VAE) and latent space adaptation.
      • A VAE is first trained using a dataset of 166 parametrically sampled CFD total pressure loss fields to learn a low-dimensional statistical representation of these fields.
    • EN 要点:
      • arXiv:2609.28558v1 Announce Type: new

  • Abstract: CFD predictions of open tip clearance flow in compressor cascades are subject to discrepancies relative to experiments, while experimental observation…

  • This study proposes a non-intrusive correction method based on a variational autoencoder (VAE) and latent-space adaptation

  • A VAE is first trained using a dataset of 166 parametrically sampled CFD total pressure loss fields to learn a low-dimensional statistical representation of the…

  • CARE: Condition-Aware Representation Regularization for Diffusion Models

    • Published: 2026-09-25 12:00 Beijing Time
    • Abstract: - arXiv:2609.28561v1 Announcement Type: New.
      • Abstract: Recent advances in diffusion models highlight the importance of representation regularization for improving sample quality and training efficiency.
      • However, commonly used regularization methods often overlook the built-in conditions (such as labels or text) that directly determine the generation target.
      • In this work, we demonstrate how conditioning signals affect the feature distribution and introduce CARE (Condition-Aware Representation Regularization).
    • EN 要点:
      • arXiv:2609.28561v1 Announce Type: new
      • Abstract: Recent advances in diffusion models highlight the importance of representation regularization for improving sample quality and training efficiency
      • However, commonly used regularization methods often overlook the built-in conditions (such as labels or texts) which directly determine the generation target
      • In this work, we demonstrate how conditioning signals affect the feature distribution and introduce the CARE (Condition-Aware REpresentation regularization)

  • SpaFactor: Lightweight Spatial Context-Aware Gene Program Modeling for Histology-to-Transcriptomics Inference

    • Published: 2026-09-25 12:00 Beijing Time
    • Abstract: - arXiv:2609.28563v1 Announcement Type: New.
      • Abstract: Spatial transcriptomics (ST) profiles gene expression within tissue architecture, but its cost and experimental complexity limit routine use.
      • Therefore, predicting spatial expression from routinely available hematoxylin and eosin (H&E) images provides a scalable alternative.
      • However, traditional methods often treat high-dimensional gene outputs as independent targets, overlooking the biological coordination between genes while remaining susceptible to high-dimensional noise and overfitting.
    • EN 要点:
      • arXiv:2609.28563v1 Announce Type: new
      • Abstract: Spatial transcriptomics (ST) profiles gene expression within tissue architecture, but its cost and experimental complexity limit routine use

  • Predicting spatial expression from routinely available hematoxylin and eosin (HE) images therefore offers a scalable alternative

  • However, conventional methods often fit high-dimensional gene outputs as independent targets, overlooking the biological coordination among genes while remainin…

  • When Explanations Cannot Be Read: Measuring and Correcting SHAP and LIME Rendering for Right-to-Left Languages

    • Published: 2026-09-25 12:00 Beijing Time
    • Abstract: - arXiv:2609.28565v1 Announcement type: new.
      • Abstract: Post hoc explanation methods such as SHAP and LIME are widely used to interpret text classifiers, but their visualizations are mainly designed for left-to-right languages.
      • When applied to right-to-left (RTL) languages (e.g., Urdu, Arabic, Persian, and Hebrew), the attribution values remain mathematically valid, but their visual presentation fails.
      • Tokens appear out of sequence, connected letterforms break apart, and plot layouts do not follow the natural reading direction.
    • EN Highlights:
      • arXiv:2609.28565v1 Announce Type: new
      • Abstract: Post hoc explanation methods such as SHAP and LIME are widely used to interpret text classifiers, but their visualizations are mainly designed for lef…
      • When applied to right-to-left (RTL) languages such as Urdu, Arabic, Persian, and Hebrew, the attribution values remain mathematically valid, while their visual…
      • Tokens appear out of sequence, connected letterforms break apart, and plot layouts do not follow the natural reading direction

  • Leakage-Safe Machine Learning for Hydrogen Embrittlement Detection in 316L Stainless Steel: A Region-Held-Out Evaluation of Texture and Deep Features in SEM Micrographs

    • Published: 2026-09-25 12:00 Beijing Time
    • Abstract: - arXiv:2609.28567v1 Announcement type: new.
      • Abstract: Scanning electron microscopy (SEM) is commonly used to characterize the microstructural changes in structural steel caused by hydrogen embrittlement (HE).
      • Machine learning can automate this characterization, but models are often evaluated using image-level splits.
      • When multiple images originate from the same sample area, this split between the training and test sets leaks information.
    • EN Highlights:
      • arXiv:2609.28567v1 Announce Type: new
      • Abstract: Scanning electron microscopy (SEM) is routinely used to characterize the microstructural changes caused by hydrogen embrittlement (HE) in structural s…
      • Machine learning can automate this characterization, but models are often evaluated using image-level splits

  • When several images come from the same specimen region, such splits leak information between the training and test sets

  • Time-Series Foundation Models That Understand Data Revisions

    • Publication Time: 2026-09-25 12:00 Beijing Time
    • Abstract: - arXiv:2609.28576v1 Announce Type: new.
      • Abstract: Historical observations are not always fixed: statistical agencies revise previously published values as new evidence arrives.
      • Therefore, forecasting from a contemporary download can expose a model to information unavailable at the date it purportedly made a prediction.
      • We propose VINTAGE-TS, a revision-aware adaptation of a time-series foundation model that distinguishes observation time from information-availability time.
    • EN Key Points:
      • arXiv:2609.28576v1 Announce Type: new
      • Abstract: Historical observations are not always fixed: statistical agencies revise previously published values as new evidence arrives
      • Forecasting from a contemporary download can therefore expose a model to information unavailable at the date it purportedly made a prediction
      • We propose VINTAGE-TS, a revision-aware adaptation of a time-series foundation model that distinguishes observation time from information-availability time

  • Uncovering Residential PV-EV Co-Adoption from Smart-Meter Data: Load Archetypes and Detection for Demand-Side Planning

    • Publication Time: 2026-09-25 12:00 Beijing Time
    • Abstract: - arXiv:2609.28578v1 Announce Type: new.
      • Abstract: The increasing adoption of electric vehicles (EVs) and rooftop photovoltaic (PV) systems is reshaping residential electricity demand and creating new challenges for demand-side management (DSM), tariff design, and low-voltage network planning.
      • Much of the existing literature examines EV charging or PV generation in isolation, leaving the behavioral dynamics of household co-adoption less understood.
      • We develop an integrated, two-part workflow to analyze advanced metering infrastructure (AMI) data.
    • EN Key Points:
      • arXiv:2609.28578v1 Announce Type: new
      • Abstract: The increasing adoption of electric vehicles (EVs) and rooftop photovoltaic (PV) systems is reshaping residential electricity demand and creating new…
      • Much of the existing literature examines EV charging or PV generation in isolation, leaving the behavioral dynamics of household co-adoption less understood
      • We develop an integrated, two-part workflow to analyze advanced metering infrastructure (AMI) data

  • Auditability Is Not One Property: Rule Overlap, Behavioural Agreement, and Composition in Reinforcement Learning

    • Publication Time: 2026-09-25 12:00 Beijing Time

  • Summary: - arXiv:2609.28581v1 Announce Type: new.

    • Abstract: Reinforcement-learning (RL) policies are often distributed as opaque neural checkpoints, while training logs show that a run occurred without explaining what the policy learned.
    • We study whether independently trained policies can be represented and composed through auditable discrete behavioral rules.
    • We define auditability as six separately testable predicates: trace integrity, lossless coding, rule coverage, behavioral agreement, composition quality, and value model reliability.