System translated (Gemini)

🤖 AI 速览

Today’s main theme is agents entering professional workflows: OpenAI, Atlassian, and Ironclad are advancing enterprise knowledge execution and contract workflow agents, while reliability assessment is simultaneously being scrutinized; DeepMind open-sourced EmbeddingGemma 2, with multimodal …
📋 文章元数据
发布时间
2026-10-07
类型
ai-daily
字数
5896
阅读时长
28 min

2026-10-07 AI Daily | Agents Delve into Enterprise Contract and Knowledge Workflows, EmbeddingGemma 2 Open-Sources On-Device Multimodality Link to heading

Today’s main theme is agents entering professional workflows: OpenAI, Atlassian, and Ironclad are advancing agents for enterprise knowledge execution and contract processes, while the reliability of evaluations is simultaneously being scrutinized. DeepMind open-sources EmbeddingGemma 2, bringing multimodal retrieval down to the device level, and the announcement of America’s “Super Intelligence Force” sparks discussion on policy centralization.

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

Today’s main theme can be summarized as: agents are entering professional workflows, with efficiency and evaluation being scrutinized in parallel. The updates from OpenAI, Atlassian, and Ironclad are best read together: the former turns enterprise knowledge into action, while the latter trains computers to use agents in contract processes. Progress in mathematical reasoning shows that frontier models continue to extend into scientific problems. Google DeepMind released EmbeddingGemma 2, opening up lightweight multimodal embeddings; research on arXiv, such as SYNLAT’s long CoT compression and fixed input embedding tables, all point towards reducing inference costs and creating lightweight models. On the evaluation side, several papers caution that reference selection, brain alignment scores, and general decision-making benchmarks can all be misinterpreted, making them essential reading for evaluation teams.

🌐 AI Hot Topics on X Link to heading

Topic 1: Trump Launches Super Intelligence Force to Lead U.S. Efforts Link to heading

  • Category: AI · News
  • Overview: Trending for: 2 days ago, Related posts: 111,000
  • What it is: Trump announced the launch of a “Super Intelligence Force” to lead U.S. AI affairs.
  • Why it’s important: This move indicates that the U.S. is strengthening AI strategy coordination at the federal level, potentially reshaping AI policy, regulation, and the global competitive landscape.
  • Discussion summary: Discussions on X focus on the force’s actual authority and effectiveness, whether AI policy will become centralized, and the impact of appointing individuals like Jay Clayton on innovation and regulation. Opinions are divided between accelerating U.S. AI leadership and concerns about politicization or excessive centralization.

Topic 2: Google Launches EmbeddingGemma 2 for On-Device Multimodal AI Link to heading

  • Category: AI · News
  • Overview: Trending for: 7 hours ago, Related posts: 4,000
  • What it is: Google released the open-source multimodal embedding model EmbeddingGemma 2, which can process text, code, images, video, and audio on consumer-grade devices like mobile phones.
  • Why it’s important: The model brings multimodal retrieval capabilities down to the device level, reducing latency and privacy risks. It uses Matryoshka Representation Learning to significantly decrease storage and memory usage, promoting local AI search and lightweight multimodal applications.
  • Discussion summary: Discussions on X primarily focus on its on-device resource consumption, the quality of its multimodal vectors, and competition with recent releases like Mistral Large 4 and Nano Banana 2.1. Others are discussing the significance of the Apache 2.0 license for commercial adoption. The point of contention is whether small-parameter models can achieve the accuracy required for complex cross-modal retrieval.

Topic 3: OpenAI Launches 28-Day Codex Improvement Challenge Amid Claude Rivalry Link to heading

  • Category: AI · News
  • Overview: Trending for: 2 days ago, Related posts: 67,000
  • What it is: OpenAI has launched a 28-day Codex improvement challenge to compete with Claude in the AI programming assistant space.
  • Why it’s important: This move reflects the intensifying competition in the AI code generation sector. OpenAI aims to rapidly iterate on Codex through community feedback to solidify its market position in developer tools.
  • Discussion summary: Current discussions on X are focused on comparing the coding abilities of Codex and Claude, whether the challenge can bring substantial improvements, and developers’ expectations and doubts about the models’ actual performance.

Summary of AI Public Opinion on X Today Link to heading

Today’s main narrative is that AI competition is expanding from model capabilities to multiple fronts, including national policy, on-device open source, and developer ecosystems, with all parties accelerating their efforts to secure a dominant position. The consensus is that the strengthening of AI strategy coordination at the U.S. federal level, the push for models on consumer devices, and the rising importance of the programming assistant track all reflect the intensifying battle for AI dominance. The main points of disagreement revolve around whether Trump’s “Super Intelligence Force” will lead to over-centralization or politicization, and whether small-parameter on-device models like EmbeddingGemma 2 can truly meet the accuracy demands of complex cross-modal retrieval. Potential risks include the possibility that policy centralization could stifle open innovation, that overestimated on-device multimodal capabilities could fail in critical scenarios, and that OpenAI’s reliance on a short-term challenge to iterate Codex might amount to mere community marketing, failing to substantially close the coding capability gap with Claude.

💡 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

Timeframe: Last 3 days; covers 22 sources; 35 updates in total.

Stratechery by Ben Thompson (A_full) Link to heading

  • Game Decompilation, Is This Legal?, A Well-Trodden Path
    • Publication Time: 2026-10-06 18:07 Beijing Time
    • Summary: - Games are being decompiled, but the real risk to gaming is new games and increased personalization.
      • $15/month* or *$150/year.
      • Substantive analysis of the day’s news via three weekly emails or a podcast.
      • Strategy Interviews.
      • Interviews with leading public company CEOs, private company founders, and discussions with fellow analysts.
    • EN Key Points:
      • Games are being decompiled, but the real risk to gaming is new games and increased personalization.

OpenAI Blog (A_full) Link to heading

  • Atlassian and OpenAI expand partnership to turn enterprise knowledge into action

    • Publication Time: 2026-10-07 00:00 Beijing Time
    • Summary: - Atlassian and OpenAI are expanding their partnership to connect frontier models with enterprise knowledge and help teams plan, build, and deliver work.
      • This article from the OpenAI Blog explains how Atlassian and OpenAI are expanding their partnership to turn enterprise knowledge into action, shaping the broader AI and infrastructure landscape.
      • The expansion of the Atlassian and OpenAI partnership to turn enterprise knowledge into action also has practical implications for founders, operators, and investors.
    • EN Key Points:
      • Atlassian and OpenAI are expanding their partnership to connect frontier models with enterprise knowledge and help teams plan, build, and deliver work.

  • Sharing AI progress in mathematics

    • Publication Time: 2026-10-06 20:00 Beijing Time
    • Summary: - OpenAI has published new results on open problems in mathematics from an internal frontier model and has shared the Lean proof formalizations and research details on GitHub….
      • This article from the OpenAI Blog explains how sharing AI progress in mathematics is shaping the broader AI and infrastructure landscape.
      • Sharing AI progress in mathematics also has practical implications for founders, operators, and investors.
    • EN Key Points:
      • OpenAI publishes new results on open problems in mathematics from an internal frontier model and shares Lean proof formalizations and research details on GitHub…

  • Advancing computer use with Ironclad

    • Publication Time: 2026-10-06 18:00 Beijing Time
    • Summary: - Learn how OpenAI and Ironclad are training and evaluating AI agents on complex contracting workflows to advance the use of computers in professional work.
      • This article from the OpenAI Blog explains how advancing computer use with Ironclad shapes the broader AI and infrastructure landscape.
      • Advancing computer use via Ironclad also has practical implications for founders, operators, and investors.
    • EN Key Points:
      • Learn how OpenAI and Ironclad are training and evaluating AI agents on complex contracting workflows to advance computer use for professional work.

Google DeepMind Blog (A_full) Link to heading

  • EmbeddingGemma 2: an open, lightweight multimodal embedding model
    • Published: 2026-10-07 03:57 Beijing Time
    • Abstract: - EmbeddingGemma 2: An open, lightweight multimodal embedding model.
      • This article from the Google DeepMind blog explains how EmbeddingGemma 2, an open, lightweight multimodal embedding model, is shaping the broader AI and infrastructure landscape.
      • It also brings practical implications for the founders, operators, and investors of EmbeddingGemma 2: an open, lightweight multimodal embedding model.
    • EN Key points:
      • EmbeddingGemma 2: an open, lightweight multimodal embedding model

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

  • Training Numerical Intelligence via Auto-Diagnosis and Skill Discovery

    • Published: 2026-10-06 12:00 Beijing Time
    • Abstract: - arXiv:2610.03872v1 Announce Type: new.
      • Abstract: AI agents are becoming increasingly capable of generating scientific code, but generating code is not the same as improving the algorithms behind it.
      • For numerical solvers, execution feedback can expose poor performance, but rarely reveals its underlying cause and how to address it.
      • We introduce Auto-Diagnosis and Skill Discovery (ADSD), a framework that links numerical diagnosis to reusable solver self-improvement.
    • EN Key points:
      • arXiv:2610.03872v1 Announce Type: new
      • Abstract: AI agents are becoming increasingly capable of generating scientific code, but generating code is not the same as improving the algorithms behind it
      • For numerical solvers, execution feedback can expose poor performance, but rarely reveals its underlying cause and how to address it
      • We introduce Auto-Diagnosis and Skill Discovery (ADSD), a framework that links numerical diagnosis to reusable solver self-improvement

  • REACT: Physically and Chemically Consistent Reconstruction of Marine Active Tracers

    • Published: 2026-10-06 12:00 Beijing Time
    • Abstract: - arXiv:2610.03888v1 Announce Type: new.
      • Abstract: Reconstructing global sea surface pH from sparse observations is critical for monitoring ocean acidification and understanding the marine carbon cycle.
      • Traditional assimilation and inverse models are physically grounded but are costly for large-scale reconstruction.
      • Recent black-box and physics-guided AI models have improved efficiency but are primarily designed for passive tracers, where the reconstructed variable is also the transported inventory.
    • EN Key points:
      • arXiv:2610.03888v1 Announce Type: new
      • Abstract: Reconstructing global sea surface pH from sparse observations is critical for monitoring ocean acidification and understanding marine carbon cycling
      • Traditional assimilation and inverse models are physically grounded but costly for large-scale reconstruction

  • Recent black-box and physics-guided AI models improve efficiency, but are mainly designed for passive tracers, where the reconstructed variable is also the tran…

  • Proxy Confidence: Auditing Black-Box LLM Agents with a Surrogate’s Log-Probabilities

    • Publication Time: 2026-10-06 12:00 Beijing Time
    • Abstract: - arXiv:2610.03894v1 Announcement Type: new.
      • Abstract: Deployed LLM agents emit tool calls, queries, and code that can be silently wrong—by the time the error surfaces, the action has already run.
      • Frontier chat APIs hide the model’s token probabilities; the agent’s stated confidence barely beats chance on the mistakes that matter; and resampling does not help because frontier models are highly repetitive, copying the same calls between samples.
      • We recover the missing signal from a low-cost open-weight surrogate run in parallel.
    • EN Key Points:
      • arXiv:2610.03894v1 Announce Type: new
      • Abstract: A deployed LLM agent emits tool calls, queries, and code that can be silently wrong – by the time the error surfaces, the action has run
      • Frontier chat APIs hide the model’s token probabilities; the agent’s stated confidence barely beats chance on the mistakes that matter; and resampling does not…
      • We recover the missing signal from a low-cost open-weight surrogate run in parallel

  • SGAnalog: An End-to-End Circuit Benchmark from Open-Source Silicon Tapeouts

    • Publication Time: 2026-10-06 12:00 Beijing Time
    • Abstract: - arXiv:2610.03934v1 Announcement Type: new.
      • Abstract: Existing analog integrated circuit design benchmarks make two questions hard to answer: whether a model has learned transferable circuit skills rather than recalling familiar examples, and whether its output works under defined processes and test conditions.
      • We introduce a benchmark built from human-designed, open-source circuits associated with Tiny Tapeout manufacturing shuttles.
      • The collection contains 273 topologically distinct top-level designs.
    • EN Key Points:
      • arXiv:2610.03934v1 Announce Type: new
      • Abstract: Existing analog integrated circuit design benchmarks make two questions hard to answer: whether a model has learned transferable circuit skills rather…
      • We introduce a benchmark built from human-designed, open-source circuits associated with Tiny Tapeout manufacturing shuttles
      • The collection contains 273 topologically distinct top-level designs

  • MLLMs Fail to Refuse when Using Tools Agentically

    • Publication Time: 2026-10-06 12:00 Beijing Time
    • Abstract: - arXiv:2610.03938v1 Announcement Type: new.
      • Abstract: Agentic Multimodal Large Language Models (MLLMs) have recently pushed the frontiers of visual reasoning by invoking tools such as scaling and labeling.

  • Despite the recent great success of agentic MLLMs, this work uncovers a critical safety failure in the tool-use paradigm: the ability of MLLMs using agentic tools to refuse harmful requests is becoming progressively weaker.

  • Our experiments confirm that, across three popular safety benchmarks, the safety of all the top open- and closed-weight MLLMs we tested is significantly lower in tool-using settings than in non-tool settings, with the relative refusal failure rate increasing by up to 68.7%.

  • EN Key Points:

    • arXiv:2610.03938v1 Announce Type: new
    • Abstract: Agentic multimodal large language models (MLLMs) have recently pushed the frontier of visual reasoning by calling tools such as zooming and tagging
    • Despite the recent strong success of agentic MLLMs, this work uncovers a critical safety failure in the tool-use paradigm: agentic tool-using MLLMs become less…
    • Our experiments confirm that, across three popular safety benchmarks, all the top open- and closed-weight MLLMs we test exhibit significantly lower safety in to…

  • Retrieval-Augmented Large Language Model Decision-Making for Autonomous Driving Guided by Chinese Philosophical Wisdom

    • Release Time: 2026-10-06 12:00 Beijing Time
    • Abstract: - arXiv:2610.03948v1 Announcement Type: new.
    • Abstract: Autonomous driving decision systems must balance safety, efficiency, and social norms in complex traffic interactions.
    • In existing autonomous driving decision-making approaches based on numerical optimization, sequence prediction, and large language models (LLMs), philosophical and ethical considerations have received limited attention.
    • We propose Chinese Philosophical Wisdom-Guided Driving (CPW-Drive), a closed-loop retrieval-augmented generation (RAG) framework that incorporates value guidance from Chinese philosophy into autonomous driving decision-making.
    • EN Key Points:
      • arXiv:2610.03948v1 Announce Type: new
      • Abstract: Autonomous driving decision systems must balance safety, efficiency, and social norms in complex traffic interactions
      • Philosophical and ethical considerations have received limited attention in existing autonomous driving decision-making approaches based on numerical optimizati…
      • We propose Chinese Philosophical Wisdom-Guided Driving (CPW-Drive), a closed-loop retrieval-augmented generation (RAG) framework that incorporates value guidanc…

  • LatentQuant: Preserving the Policy-Facing Latent Contract under NVFP4 VAE Quantization

    • Release Time: 2026-10-06 12:00 Beijing Time
    • Abstract: - arXiv:2610.03959v1 Announcement Type: new.
    • Abstract: Recent World Action Models (WAMs) reuse pretrained video VAEs, whose encoder latents directly condition downstream action policies.
    • Consequently, quantization must not only preserve reconstruction fidelity but also the policy-facing latent contract expected by a frozen policy.
    • Direct NVFP4 leaves W4A4 quantization error uncompensated, whereas joint quantization-aware training (QAT) can recover reconstruction by shifting this representation.

  • ROAR: Unifying Runs across Heterogeneous AI-Driven Research Systems

    • Publication Time: 2026-10-06 12:00 Beijing Time
    • Abstract: - arXiv:2610.03966v1 Announcement Type: new.
      • Abstract: Each run of an AI-driven research system (ADRS) is an expensive search in a vast solution space, and reliable evaluation requires multiple runs, making run data costly to generate and valuable to preserve for large-scale analysis.
      • However, this data remains fragmented: teams operate independently, ADRS frameworks emit results in different formats, and no shared infrastructure exists for aggregating or comparing runs across problems and systems.
      • We present ROAR, a solution for systematically unifying and analyzing heterogeneous ADRS outputs.
    • EN Highlights:
      • arXiv:2610.03966v1 Announce Type: new
      • Abstract: Each run of an AI-driven research system (ADRS) is an expensive search over a vast solution space, and dependable evaluation requires many runs, makin…
      • Yet this data remains fragmented: teams operate in isolation, ADRS frameworks emit results in different formats, and no shared infrastructure exists to aggregat…
      • We present ROAR, a solution for systematically unifying and analyzing heterogeneous ADRS outputs

  • Teaching Agents to Code Reliably

    • Publication Time: 2026-10-06 12:00 Beijing Time
    • Abstract: - arXiv:2610.03984v1 Announcement Type: new.
      • Abstract: Autonomous coding agents solve repository issues by reading code, running commands, editing files, and submitting patches.
      • Extra inference-time compute only yields gains when it produces a useful repair and provides reliable evidence for choosing a fix.
      • Three behaviors determine both, which we argue are teachable rather than byproducts of scale, so policies can embody them instead of scaffolding.
    • EN Highlights:
      • arXiv:2610.03984v1 Announce Type: new
      • Abstract: Autonomous coding agents solve repository issues by reading code, running commands, editing files, and submitting patches
      • Extra inference-time compute yields gains only when it produces a useful repair and supplies reliable evidence for choosing one

  • Three behaviors decide both, and we argue they are teachable rather than byproducts of scale, so a policy can carry them instead of a scaffold

  • Behavioral History Outperforms Descriptions of the Person for LLM Synthetic Personas

    • Posted: 2026-10-06 12:00 Beijing time
    • Abstract: - arXiv:2610.03998v1 Announce Type: new.
      • Abstract: Large language models (LLMs) are increasingly used as synthetic personas representing survey respondents.
      • Their validity as substitutes for particular respondents depends on whether they reproduce individuals’ decisions.
      • We examine what information helps synthetic respondents predict each individual’s later choices, using five conditions that add progressively richer information: no personal information, demographics, personality traits, cognitive scores, and finally, the respondent’s earlier survey choices as behavioral history.
    • EN Highlights:
      • arXiv:2610.03998v1 Announce Type: new
      • Abstract: Large language models (LLMs) are increasingly used as synthetic personas representing survey respondents
      • Their validity as substitutes for particular respondents depends on whether they reproduce individuals’ decisions
      • We examine what information helps synthetic respondents predict each individual’s later choices, using five conditions that add progressively richer information…

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

  • Fine-Grained Emotion Classification from Mobile App Reviews: An Empirical Study with Large Language Models

    • Posted: 2026-10-06 12:00 Beijing time
    • Abstract: - arXiv:2610.03802v1 Announce Type: new.
      • Abstract: Context: Fine-grained emotion classification of mobile app reviews enables requirements engineering activities that go beyond polarity-based opinion mining, including emotion-based issue prioritization and feature-oriented feedback analysis.
      • However, automatic fine-grained emotion extraction from app reviews remains understudied.
      • Objectives: Building on a previously published annotation framework and human-labelled ground truth adapted from Plutchik’s taxonomy, this paper investigates how to leverage large language models for automatic multi-label emotion classification in the presence of severe class imbalance.
    • EN Highlights:
      • arXiv:2610.03802v1 Announce Type: new
      • Abstract: Context: Fine-grained emotion classification of mobile app reviews enables requirements engineering activities that go beyond polarity-based opinion m…
      • However, automatic fine-grained emotion extraction from app reviews remains understudied
      • Objectives: Building on a previously published annotation framework and human-labelled ground truth adapted from Plutchik’s taxonomy, this paper investigates ho…

  • Same Output, Different Gold: Measuring How Reference Choice Moves a Multilingual Benchmark Score

    • Publication Time: 2026-10-06 12:00 Beijing Time
    • Abstract: - arXiv:2610.03825v1 Announcement Type: new.
      • Abstract: Benchmark scores compare system output to a reference, and methodological attention falls almost entirely on the first term.
      • The retained annotation record of a six-language benchmark for personally identifiable information contains two independent annotator labels, an aggregate delivered as gold, and a reviewer’s gold from a re-annotation of a sample by an independent expert.
      • Using this, we can fix the scored output and exchange the reference.
    • EN Key Points:
      • arXiv:2610.03825v1 Announce Type: new
      • Abstract: A benchmark score compares a system output against a reference, and methodological attention falls almost entirely on the first term
      • We measure the second
      • The retained annotation record of a six-language benchmark for personally identifiable information contains two independent annotator labellings, the aggregate…

  • The Score Is Not the Structure: Brain Alignment and Cross-Lingual Transfer

    • Publication Time: 2026-10-06 12:00 Beijing Time
    • Abstract: - arXiv:2610.03827v1 Announcement Type: new.
      • Abstract: Researchers often support the claim that a model shares structure with the brain, or across languages, by reporting a similarity score.
      • We ask what such a score reads when the shared structure is absent, or when the tool that measures it does not work.
      • We examine two settings, and in both, the score is not what it appears to be.
    • EN Key Points:
      • arXiv:2610.03827v1 Announce Type: new
      • Abstract: Researchers often support the claim that a model shares structure with the brain, or across languages, by reporting a similarity score
      • We ask what such a score reads when the shared structure is absent, or when the tool that measures it does not work
      • We check two settings, and in both the score is not what it appears

  • OncoNoteBERT: A Foundation Representation Model for Natural Language Processing of Real-World Outpatient Oncology Notes

    • Publication Time: 2026-10-06 12:00 Beijing Time
    • Abstract: - arXiv:2610.03829v1 Announcement Type: new.
      • Abstract: Real-world outpatient oncology notes contain specialized terminology, tumor staging expressions, treatment names, toxicity descriptions, and institution-specific de-identification markers, which may not be effectively represented by general biomedical or adjacent clinical language models.
      • We developed and evaluated an oncology-specific BERT-style encoder using a regulated UK outpatient oncology corpus, containing 290,026 notes from 21,564 patients treated for lung and head-and-neck cancer.
      • We compared RadBERT and PathologyBERT with two local strategies: OncoNote-RadBERT (generated through continued masked language model pre-training) and OncoNoteBERT (trained from scratch using an oncology WordPiece tokenizer).
    • EN Key Points:
      • arXiv:2610.03829v1 Announce Type: new

  • Abstract: Real-world outpatient oncology notes contain specialised terminology, tumour staging expressions, treatment names, toxicity descriptions, and institut…

  • We developed and evaluated oncology-specific BERT-style encoders using a governed UK outpatient oncology corpus comprising 290,026 notes from 21,564 patients tr…

  • We compared RadBERT and PathologyBERT with two local strategies: OncoNote-RadBERT, produced by continued masked language model pretraining, and OncoNoteBERT, tr…

  • SYNLAT: Syntax-Aligned Text-Latent Compression for Chain-of-Thought Reasoning

    • Publication Time: 2026-10-06 12:00 Beijing Time
    • Abstract: - arXiv:2610.03839v1 Announcement Type: New.
      • Abstract: Long chain-of-thought (CoT) traces incur significant output token costs.
      • With a limited budget, compression must preserve critical answer information, making boundary placement central.
      • Token-level and fixed-length boundaries can segment coherent spans, such as phrases, formulas, and local derivations, while step-level boundaries can bind content requiring different compression operations.
    • EN Highlights:
      • arXiv:2610.03839v1 Announce Type: new
      • Abstract: Long chain-of-thought (CoT) traces impose substantial output-token costs
      • Under constrained budgets, compression must preserve answer-critical information, making boundary placement central
      • Token-level and fixed-length boundaries can fragment coherent spans such as phrases, formulas, and local derivations, whereas step-level boundaries can bind con…

  • When Evidence Changes: Evaluating Memory Repair and Re-reading in Language-Model Agents

    • Publication Time: 2026-10-06 12:00 Beijing Time
    • Abstract: - arXiv:2610.03902v1 Announcement Type: New.
      • Abstract: When documents supporting facts derived by an agent are revoked or replaced, should memory be repaired or should the current evidence be re-read?
      • We introduce an evidence revision evaluation for medication and problem list tasks in public ICU records.
      • Under revocation, replacement, and control events, we compare full re-reading and source-filtered re-reading of two 7B models with caching, reconstruction, and graph-local repair.
    • EN Highlights:
      • arXiv:2610.03902v1 Announce Type: new
      • Abstract: When documents supporting an agent’s derived facts are revoked or replaced, should it repair memory or re-read current evidence
      • We introduce an evidence-revision evaluation on medication- and problem-list tasks from public ICU records

  • Under revocation, replacement and control events, we compare full and source-filtered re-reading with caching, rebuilding and graph-local repair across two 7B m…

  • General Decision Models: Benchmarking and Insights Beyond Jev

    • Publication Time: 2026-10-06 12:00 Beijing Time
    • Summary: - arXiv:2610.03935v1 Announcement Type: New.
      • Abstract: General decision models (e.g., Jev) have recently emerged as effective alternatives to LLMs for structured judgment and selection.
      • However, what types of decisions can these models reliably make, and how does their behavior change when individual decisions are composed into larger systems?
      • To study this, we introduce JEVal, a bilingual benchmark comprising 11,257 instances from 36 datasets across 10 application domains, and evaluate 25 model configurations covering general decision models and generative LLMs.
    • EN Key Points:
      • arXiv:2610.03935v1 Announce Type: new
      • Abstract: General decision models, such as Jev, have recently emerged as efficient alternatives to LLMs for structured judgment and selection
      • But what kinds of decisions can these models reliably make, and how does their behavior change when individual decisions are composed into larger systems
      • To study this, we introduce JEVal, a bilingual benchmark comprising 11,257 instances from 36 datasets across 10 application domains, and evaluate 25 model confi…

  • A Step Towards Forgetting: Optimiser History and the Loss of Answer Mass

    • Publication Time: 2026-10-06 12:00 Beijing Time
    • Summary: - arXiv:2610.03940v1 Announcement Type: New.
      • Abstract: During the fine-tuning process, language models can assign a smaller probability to previously learned answers, even when the current gradient is intended to preserve that probability.
      • With momentum, each update also carries gradients computed at earlier model states, and these stored contributions can push the model in the opposite direction.
      • We investigate how this optimizer memory contributes to forgetting by separating the old-task loss into confusion among its answers and the leakage of probability outside the answer set.
    • EN Key Points:
      • arXiv:2610.03940v1 Announce Type: new
      • Abstract: During fine-tuning, a language model can assign less probability to previously learned answers even when the current gradient acts to preserve that pr…
      • With momentum, each update also carries gradients computed at earlier model states, and these stored contributions can push the model in the opposite direction
      • We investigate how this optimiser memory contributes to forgetting by separating old-task loss into confusion among its answers and leakage of probability outsi…

  • BAIBAICHUCHU at the NTCIR-19 FinArg-3 Task: When Is Maximum Possible Profit Predictable from Investor Text?

    • Publication Time: 2026-10-06 12:00 Beijing Time
    • Abstract: - arXiv:2610.03962v1 Announce Type: new.
      • Abstract: The BAIBAICHUCHU team participated in the Social Media Subtask of NTCIR-19 FinArg-3, ranking Chinese investor posts by Maximum Possible Profit (MPP).
      • A three-track ensemble of lexical features, a FinArg-2 pre-finetuned MacBERT ranker, and an LLM judge reached 0.734 in post-grouped development evaluation, but our best official run score was 0.517.
      • All twelve submitted runs scored between 0.4598 and 0.5402, and our 26 unanimous three-track pairs scored 0.500.
    • EN Key Points:
      • arXiv:2610.03962v1 Announce Type: new
      • Abstract: The BAIBAICHUCHU team participated in the Social Media Subtask of NTCIR-19 FinArg-3, ranking Chinese investor posts by Maximum Possible Profit (MPP)
      • A three-track ensemble of lexical features, a FinArg-2-pre-finetuned MacBERT ranker, and an LLM judge reaches 0.734 in post-grouped development evaluation, but…
      • All twelve submitted runs lie between 0.4598 and 0.5402, and our 26 unanimous three-track pairs score 0.500

  • Do Language Models Need a Trainable Input Embedding Table? Fixed Minimal Token Codes at 1.7B-Class Scale

    • Publication Time: 2026-10-06 12:00 Beijing Time
    • Abstract: - arXiv:2610.04002v1 Announce Type: new.
      • Abstract: A trainable input embedding table assigns each vocabulary item an independently adjustable vector.
      • We investigate whether this token-specific parameterization is necessary for substantial language modeling capability, or whether a shared Transformer can learn from fixed token identities.
      • We compare three decoder-only language models trained from scratch, using the same tokenizer, contextual backbone, untied output-head architecture, and training recipe, each targeting a budget of 100 billion predicted tokens.
    • EN Key Points:
      • arXiv:2610.04002v1 Announce Type: new
      • Abstract: A trainable input embedding table assigns each vocabulary item an independently adjustable vector
      • We investigate whether this token-specific parameterization is required for substantial language-modeling capability, or whether a shared Transformer can learn…
      • We compare three decoder-only language models trained from scratch with the same tokenizer, contextual backbone, untied output-head architecture, and training r…

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

  • Bayes-Sufficient Compression Is Not Enough: How Does Communication Help Multi-Agent Systems?

  • Published: 2026-10-06 12:00 Beijing Time

    • Abstract: - arXiv:2610.03769v1 Announcement Type: new.
      • Abstract: Multi-agent LLM systems pair a sender with broad context and an executor with a limited local view.
      • We study when a short message can improve the executor’s next decision, when raw context is preferable, and when a stronger sender would be helpful.
      • Our framework, \emph{receiver-relative bounded coordination}, expresses message utility as receiver gain minus protocol tax.
    • EN Highlights:
      • arXiv:2610.03769v1 Announce Type: new
      • Abstract: Multi-agent LLM systems pair a sender with broad context and an executor with a limited local view
      • We study when a short message improves the executor’s next decision, when raw context is preferable, and when a stronger sender helps
      • Our framework, \emph{receiver-relative bounded coordination}, expresses message utility as receiver gain minus protocol tax

  • Least Squares for Time Series Forecasting

    • Published: 2026-10-06 12:00 Beijing Time
    • Abstract: - arXiv:2610.03812v1 Announcement Type: new.
      • Abstract: A time series forecast is scored based on the future values of that series.
      • The representation loss of regressing the next latent variable (as in LeNEPA) is a different least-squares problem on the same bottleneck.
      • The forecasting program minimizes the decoded latent error on the coordinates to be reported.
    • EN Highlights:
      • arXiv:2610.03812v1 Announce Type: new
      • Abstract: A time-series forecast is scored on a future value of the series
      • A representation loss that regresses the next latent, as in LeNEPA, is a different least-squares problem on the same bottleneck
      • We write both programs down

  • Memory-State Critic for Asymmetric Actor-Critic with Application to Vision-Based Pursuit-Evasion

    • Published: 2026-10-06 12:00 Beijing Time
    • Abstract: - arXiv:2610.03830v1 Announcement Type: new.
      • Abstract: In partially observable Markov decision processes, the optimal policy usually depends on the history of observations and past actions.
      • When additional information (e.g., the true state of the environment) is available during training, asymmetric actor-critic methods have become a popular approach for learning such policies.
      • The critic is not needed at execution but can access the state.
    • EN Highlights:
      • arXiv:2610.03830v1 Announce Type: new
      • Abstract: In partially observable Markov decision processes, the optimal policy generally depends on the history of observations and past actions
      • Asymmetric actor-critic methods have become popular to learn such policies when additional information, such as the true state of the environment, is available…
      • The critic, which is not needed at execution, is given access to the state

  • Where Does Jagged Competence Come From?

    • Publish Time: 2026-10-06 12:00 Beijing Time
    • Abstract: - arXiv:2610.03831v1 Announce Type: new.
      • Abstract: Capable systems often exhibit jagged competence: low average error alongside failures on specific inputs.
      • We ask where it comes from in a task constructed from two known layers.
      • A lower layer A computes five per-slot sums from records; an upper layer B uses the slot-1 sum and a mode inherited from earlier boards to predict the next board’s category, so B cannot be computed solely based on the current A: B is a strict extension of A.
    • EN 要点:
      • arXiv:2610.03831v1 Announce Type: new
      • Abstract: Capable systems often show jagged competence: low average error alongside failures on particular inputs
      • We ask where it comes from in a task built from two known layers
      • A lower layer A computes five per-slot sums from records; an upper layer B uses the slot-1 sum and a mode carried over from earlier boards to predict the next b…

  • LLM-enhanced spatio-temporal learning for grid-level docked bike sharing demand prediction

    • Publish Time: 2026-10-06 12:00 Beijing Time
    • Abstract: - arXiv:2610.03834v1 Announce Type: new.
      • Abstract: Short-term bike-sharing demand prediction is complicated by spatio-temporal non-stationarity and the practical difficulty of incorporating unstructured external text into numerical pipelines.
      • Traditional methods rely on historical flow sequences and fixed graph structures, thereby limiting their accuracy when anomalous social events disrupt normal travel patterns.
      • We propose a forecasting framework where a Large Language Model (LLM) drives a semantic shockwave mechanism, converting free-form urban text (e.g., municipal event schedules, local news, and traffic announcements) into a quantified spatio-temporal perturbation field.
    • EN 要点:
      • arXiv:2610.03834v1 Announce Type: new
      • Abstract: Short-term bike-sharing demand forecasting is complicated by spatial-temporal non-stationarity and the practical difficulty of incorporating unstructu…
      • Conventional approaches rely on historical flow sequences and fixed graph structures, thereby constraining their accuracy when anomalous social events perturb n…
      • We propose a forecasting framework in which a Large Language Model (LLM) drives a semantic shockwave mechanism that converts free-form urban text, such as munic…

  • KVE-KD: Key Visual Evidence-Guided Knowledge Distillation for Vision-Language Models

    • Publish Time: 2026-10-06 12:00 Beijing Time
    • Abstract: - arXiv:2610.03842v1 Announce Type: new.
      • Abstract: Knowledge distillation is crucial for deploying vision-language models on resource-constrained devices.
      • However, existing methods often impose uniform supervision on visual tokens or rely on static token selection, which can confuse task-relevant cues with background noise and degrade cross-modal reasoning.

  • To address this limitation, we propose Key Visual Evidence-guided Knowledge Distillation (KVE-KD), a framework that dynamically focuses feature distillation on task-relevant visual tokens identified by the teacher model.

    • EN Highlights:
      • arXiv:2610.03842v1 Announce Type: new
      • Abstract: Knowledge distillation is crucial for deploying vision-language models on resource-constrained devices
      • However, existing methods typically impose uniform supervision across visual tokens or rely on static token selection, which confuses task-relevant cues with ba…
      • To address this limitation, we propose Key Visual Evidence-guided Knowledge Distillation (KVE-KD), a framework that dynamically focuses feature distillation on…

  • BAT-NO: A Boundary-Condition-Aware Transformer Neural Operator for Crashworthiness Prediction of Vehicle Components

    • Published: 2026-10-06 12:00 Beijing Time
    • Abstract: - arXiv:2610.03854v1 Announce Type: new.
      • Abstract: High-fidelity finite-element simulations provide accurate crashworthiness predictions, but their cost limits iterative design exploration.
      • Deep learning surrogates can reduce this cost, but many component-level models are developed under a single prescribed boundary condition, limiting generalization to boundary variations.
      • This work proposes a Boundary-Condition-Aware Transformer Neural Operator (BAT-NO) for autoregressive prediction of transient displacement fields and scalar crashworthiness responses under varying geometries and boundary conditions.
    • EN Highlights:
      • arXiv:2610.03854v1 Announce Type: new
      • Abstract: High-fidelity finite-element simulations provide accurate crashworthiness predictions, but their cost limits iterative design exploration
      • Deep learning surrogates can reduce this cost, but many component-level models are developed under a single prescribed boundary condition, limiting generalisati…
      • This work proposes a Boundary-Condition-Aware Transformer Neural Operator (BAT-NO) for autoregressive prediction of transient displacement fields and scalar cra…

  • Retrieval-Centric Deep Learning in Growing Nonparametric Neural Networks

    • Published: 2026-10-06 12:00 Beijing Time
    • Abstract: - arXiv:2610.03858v1 Announce Type: new.
      • Abstract: We study universal layers for deep learning that, instead of compressing arbitrarily large training data into a fixed-size weight matrix, store a new pair of key-value representations for each data point during training and retrieve and recombine these representations via an attention mechanism at inference – resulting in growing neural networks (NNs).
      • (arXiv:2202.05798) proposed this perspective starting from classical duality expressing any linear layer in a deep NN trained via gradient descent as a linear attention (LA) over the training data points, replacing LA with more powerful attention functions, as they suggested, turns out to be non-trivial: We show that naively applying the learning rules from the LA case to advanced kernels does not lead to principled optimization.

  • Here, we fill this gap and, based on radial basis function (RBF) and softmax-like kernels, develop function-gradient-based learning rules for kernelized attention layers - establishing a principled “retrieval-centric deep learning” (RCDL) paradigm.

    • EN Highlights:
      • arXiv:2610.03858v1 Announce Type: new
      • Abstract: We investigate a general-purpose layer for deep learning that, instead of compressing arbitrary-size training data into fixed-size weight matrices, st…
      • While Irie et al
      • (arXiv:2202.05798) have put forward this perspective from the classic duality expressing any linear layer in a deep NN trained by gradient descent as linear att…

  • Reinforcement Learning on the Discrete Composition Channel of a Crystal Generator: Validated Gains and Reward Hacking

    • Release Time: 2026-10-06 12:00 Beijing Time
    • Abstract: - arXiv:2610.03880v1 Announce Type: new.
      • Abstract: Inverse materials design is a long-standing goal of computational materials discovery.
      • Generative models for crystalline materials are typically trained to match the distribution of a structure database, while nothing in their training objective points to a specific design goal, such as a target property.
      • We use group-relative policy optimization (GRPO) to align a generative model based on stochastic interpolants and discrete flow matching with general black-box reward functions via reinforcement learning.
    • EN Highlights:
      • arXiv:2610.03880v1 Announce Type: new
      • Abstract: Inverse materials design is a long-standing goal of computational materials discovery
      • Generative models for crystalline materials are typically trained to match the distribution of a structure database, while nothing in their training objective p…
      • We use group-relative policy optimization (GRPO) to align a generative model based on stochastic interpolants and discrete flow matching with general black-box…

  • AdaEva: Accelerating LLM-Driven Algorithm Design with Adaptive Partial Evaluation

    • Release Time: 2026-10-06 12:00 Beijing Time
    • Abstract: - arXiv:2610.03896v1 Announce Type: new.
      • Abstract: Large Language Models (LLMs) are increasingly used for automated algorithm design.
      • However, the computational cost of evaluating the generated algorithms can be excessive.
      • We consider the common LLM-driven automated algorithm design (LLM4AD) setup, where candidate algorithms are evaluated by aggregating performance on a shared set of training instances.
    • EN Highlights:
      • arXiv:2610.03896v1 Announce Type: new
      • Abstract: Large Language Models (LLMs) are increasingly used for automated algorithm design
      • However the computational cost of evaluating the generated algorithms can be excessive

  • We consider the common LLM-driven automated algorithm design (LLM4AD) setting in which a candidate algorithm is evaluated by aggregating its performance over a…