System translated (Gemini)

🤖 AI 速览

Today’s main theme shifts from model demonstration to system implementation: agents are beginning to enter real-world environments such as workflows, clinical settings, customer service, and large codebases, with validators, routing, and backend writing becoming key capabilities. Meanwhile, …
📋 文章元数据
发布时间
2026-07-04
类型
ai-daily
字数
7337
阅读时长
35 min

2026-07-04 AI Daily | Agents Move into Business Backends: Constraints, Validation, and Write Capabilities Become the New Focus Link to heading

Today’s main theme shifts from model demonstrations to system implementation: agents are beginning to enter real-world environments like workflows, clinical settings, customer service, and large codebases, where validators, routing, and backend write capabilities are becoming crucial. Meanwhile, approaches like small models, diffusion language models, and weight guidance continue to diverge. The compute supply chain, AI sovereignty, and collaborations in film production are also expanding the industry’s boundaries.

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

The most noteworthy topic for today’s Watch List is the research cluster on “agents entering real business systems.” From the RLVR tool-calling in Atlassian workflows and feedback in FHIR clinical environments to difficulty-based routing for customer service agents and multi-agent summarization of large codebases, the focus is no longer on conversational abilities. Instead, it’s on stable execution in environments with constraints, validators, and backend write capabilities.

The second main theme is the continued divergence of model forms. Wiola explores new architectures for small language models, diffusion language models are being used to draft radiology reports, and CreativityNeuro is experimenting with weight guidance to mitigate “mode collapse.” These papers are worth watching because they all challenge the single-path, next-token paradigm.

On the industry side, the Palantir-Nvidia partnership, discussions on AI sovereignty and employment, continue the themes of infrastructure and policy maneuvering. Meanwhile, the collaboration between Google DeepMind and A24 suggests that generative AI is becoming more deeply integrated into the filmmaking process.

🌐 AI Hot Topics on X Link to heading

Topic 1: Naval Ravikant Explores AI’s Rapid Future in New Podcast Link to heading

  • Category: AI · News
  • Overview: Trending for 12 hours, 490 related posts
  • What it is: Investor and thinker Naval Ravikant discussed the rapidly advancing future of AI and its potential societal, business, and personal impacts in a new podcast.
  • Why it matters: Naval is highly influential in the tech and startup communities. His assessment of AI trends could shape how entrepreneurs, investors, and developers perceive the opportunities and risks associated with AI.
  • Discussion summary: The discussion on X centers on whether the pace of AI advancement is underestimated, how AI will transform work and entrepreneurship, and the divide between optimistic visions and potential societal risks.

Topic 2: AXT Secures $22.3 Million Prepayment Deal with Coherent for AI Wafers Link to heading

  • Category: AI · News
  • Overview: Trending for 23 hours, 481 related posts
  • What it is: AXT and Coherent have entered into a $22.3 million prepayment agreement to support the supply of wafers for AI-related applications.
  • Why it matters: This indicates that demand from AI data centers and high-speed optical communications continues to propagate to the upstream semiconductor materials and wafer supply chain, highlighting the pull effect of AI infrastructure expansion on critical hardware segments.
  • Discussion summary: The discussion on X focuses on whether this order signifies sustained growth in demand for AI optical communications and compound semiconductors, whether AXT will benefit, and whether the market is overestimating the actual impact of a $22.3 million deal on the company’s performance and industry trends.

Topic 3: Anthropic’s Claude Fable 5 Returns Online with Top AI Performance Link to heading

  • Category: AI · News
  • Overview: Trending for 1 day, 35,000 related posts
  • What it is: Anthropic’s Claude Fable 5 is back online and is reported to have leading performance in several AI benchmarks.
  • Why it matters: This shows that the competition among frontier large models continues to accelerate, especially in capabilities like reasoning, programming, and safety alignment, which could influence model selection by businesses and governments.
  • Discussion summary: Discussions on X are centered on whether its performance is genuinely superior, whether its return is related to government access or regulatory arrangements, and how Anthropic balances its safety commitments with commercial competition.

Topic 4: South Korean Chip Stocks Surge 5.8% on Meta AI Clarifications Link to heading

  • Category: AI · News
  • Overview: Trending for 19 hours, 5,300 related posts
  • What it is: After Meta provided clarifications on its AI infrastructure and chip procurement plans, market concerns about AI chip demand eased, leading to a surge of about 5.8% in South Korean semiconductor stocks at one point.
  • Why it matters: This demonstrates that the AI capital expenditure expectations of large tech companies continue to directly impact the global chip supply chain, especially the position of South Korean manufacturers in HBM, high-end memory, and AI server-related segments.
  • Discussion summary: The discussion on X primarily focuses on whether Meta’s clarification implies that demand for AI chips and HBM remains strong. Optimists believe South Korean chip stocks will continue to benefit from AI compute investment, while cautious voices worry about overheated valuations, customer concentration, and the possibility that in-house AI chips could weaken long-term orders.

Summary of Today’s AI Public Opinion on X Link to heading

The main narrative today is that AI is still seen as the core driving force behind the next wave of technology, capital expenditure, and industry chain revaluation. From Naval’s optimistic outlook on the future to optical communication wafer orders, the return of Claude models, and Meta’s clarification on chip procurement, everything reinforces the consensus that “AI demand is still expanding.” The consensus is that competition in large model capabilities and computing infrastructure is accelerating, and this continues to ripple through to sectors like semiconductor materials, HBM, high-speed communications, and enterprise model selection. Disagreements, however, are concentrated on two points: first, whether AI’s progress and commercialization opportunities are underestimated or already overpriced by the market; and second, whether companies like Anthropic can maintain a credible balance between performance leadership, safety commitments, and regulatory relationships. Potential risks include overheated valuations in the hardware supply chain, excessive concentration of orders and customers, tech giants’ in-house chip development weakening long-term demand for external suppliers, and the underestimation of the safety, governance, and societal impacts brought by frontier models in a rapidly competitive commercial landscape.

💡 Influencer Insights Link to heading

Alright, based on the activity of AI thought leaders on the X platform over the past 24 hours, here is an analysis of industry intelligence compiled for you.

Main Hotspot: The Return of Claude Fable 5 and Its Engineering Practices

Without a doubt, the relaunch of Fable 5 is the most significant event to rock the community today.

  • Return Dynamics: @dotey and @Pluvio9yte were among the first to follow this event. @AnthropicAI announced that after receiving permission from the U.S. Department of Commerce, it has restored access to Fable 5 and Mythos 5 (@zhixianio confirmed via retweet). Users can experience it with a 50% subscription credit until July 7th, after which it will switch to a pay-per-use points system (@dotey).
  • Pricing Controversy and Application Strategies: After testing, @Pluvio9yte pointed out that despite the stunning results, its API pricing is as high as $50 per million output tokens, leading to constant complaints from the community. For this reason, @Pluvio9yte shared a detailed “Best Way to Work with Fable 5,” proposing to use it as a “main coordinator,” paired with Opus (for deep reasoning) and Sonnet (for mechanical execution) to form an Agent team, in order to save precious Fable 5 credits. His comparative review shows that Fable 5 is significantly superior to GPT-5.5 and Deepseek V4 Pro on complex tasks.
  • Developer Tool Integration: Along with the return of Fable 5, @ClaudeDevs reset everyone’s Claude Code rate limits. @dotey humorously remarked, “I lost out, it was going to be reset anyway.” This shows the deep integration between Fable 5’s return and developer tools.

Secondary Focus: In-depth Use and Pitfalls of OpenAI Codex

Outside the spotlight of Fable 5, OpenAI Codex remains a frequently mentioned term, but the discussion has shifted from initial frenzy to refined operation and problem-solving.

  • Practical Tip Sharing: @Pluvio9yte systematically shared seven major tips for using Codex, including how to utilize the 5-hour rolling window, enable the memory feature, use the /side command to check the progress of long tasks, and the efficient wrap-up method of the Goal mode.
  • Bugs and Limitations: @Pluvio9yte discovered an “infinite credit” bug in the /goal mode, while @vista8 criticized Codex for its flaw of “using templates to spout nonsense” when generating large amounts of content, and pointed out that forcing the use of a Subagent or memory feature is needed to mitigate this.

2. Noteworthy Unique Perspectives or Industry Foresight Link to heading

1. Technological Autonomy vs. Political Dependence: OpenAI’s “Public Wealth Fund” Concept @dotey provided a detailed interpretation of OpenAI’s proposal to the U.S. government to offer a 5% stake (as revealed by @AndrewCurran_). The plan aims to create a “public wealth fund” to allow the entire population to share in the dividends of AI development. This move is not only intended to alleviate public concerns about the impact of AI but is also seen as a strategic move by OpenAI to clear political hurdles and seek a regulatory “talisman.” This indicates that top AI companies are attempting to become deeply involved in national governance structures through interest alignment.

2. The Distinction Between Security Agents and Models: The Deep Advantage in Specific Tasks @dotey cited an in-depth article by the security expert “Atuin AI.” The team discovered a critical curl vulnerability that the Mythos model missed, but they cautiously pointed out: “This is an Agent designed for vulnerability mining, whereas Mythos is a general-purpose model.” Winning on a specific task does not mean overall superiority. This provides a calmer, more professional perspective for evaluating AI capabilities, distinguishing the capability boundaries between “specialized Agents” and “general-purpose models.”

3. The Development Direction of On-Device Models: “Model Cartridges” and Future Interaction @zhixianio strongly agrees with the “Model-Pak” (when large models become like cartridges) concept proposed by @geekbb. This suggests that in the future, on-device AI could be plug-and-play, much like game cartridges, allowing different tasks to call upon physically isolated and functionally specialized models. Combined with @zhixianio’s own continuous testing of on-device models like MiniCPM-o 4.5, this vision is gradually gaining a technical foundation.

4. The Infrastructure for “One-Person Companies” is Maturing: The AI Agent Marketplace @Pluvio9yte keenly spotted the AI Agent marketplace launched on OKX and commented, “This whole one-person company thing is really happening.” In this market, Agents can autonomously accept jobs, receive payments on-chain, and hire each other. This marks a shift for AI from being a “tool” to an “independent economic entity,” providing the foundational infrastructure for ordinary people to achieve large-scale commercialization using AI.

  • API Aggregation and Evaluation Platform:
    • Zenmux: Recommended by @Pluvio9yte, it allows calls to over 200 models, including Claude Fable 5, with no rate limits. There’s currently a bonus for top-ups, making it suitable for multi-model benchmarking and long-term development.
  • Front-End Design Skill Evaluation:
    • @vista8 is conducting a parallel evaluation of the 5 most popular front-end page design Skills and has published all test cases and comparison results (link in their tweet). This is a valuable reference for web developers to choose the best design Skill.
  • Research and Media Models:
    • Claude Science: This news was shared by both @vista8 and @dotey. It’s an AI workbench for researchers, now available for download, and it supports various Science Skills.
    • New Google Gemini Models: @vista8 shared API information for Gemini Nano Banana 2 Lite (for ultra-fast image generation) and Gemini Omni Flash (billed by the second), offering new options for multimodal applications.
  • Office Automation:
    • @ruanyf discovered that several domestic office platforms have released open-source CLI toolkits for Agents to call. Among them, Feishu’s project has the highest number of stars and the most comprehensive features, representing a significant entry point for smart offices.
  • Useful Chrome Extension:
    • Released by @Pluvio9yte, this free and open-source extension can batch delete or archive ChatGPT web chat history, solving a major pain point for many users.
  • Developer Tools:
    • Windows Codex Dynamic Island: Developed by @mooyuking (recommended by @AI_Jasonyu), it allows for intuitive viewing of Codex status and quotas directly on the Windows desktop.
    • China User Environment Detection Skill: Developed and open-sourced by @vista8, it aims to help developers deal with model access issues caused by complex network environments.

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

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

All-In Podcast (A_full) Link to heading

  • AI Sovereignty Wars, Palantir-Nvidia Deal, SCOTUS Birthright Ruling, Newsom’s CA Budget Lie
    • Release Time: 2026-07-04 06:12 Beijing Time
    • Abstract: - (0:00) Bestie intros: Happy Fourth of July.
      • (0:21) Palantir-Nvidia open source deal, Alex Karp’s CNBC “Crashout”.
      • (33:52) Update on the AI jobs debate.
      • (50:24) Anthropic’s Fable 5 available after export restrictions lifted.
    • EN Highlights:
      • (0:00) Bestie intros: Happy Fourth of July
      • (0:21) Palantir-Nvidia open source deal, Alex Karp’s CNBC “Crashout”
      • (33:52) Update on the AI jobs debate
      • (50:24) Anthropic’s Fable 5 available after export restrictions lifted

Google DeepMind Blog (A_full) Link to heading

  • Google DeepMind and A24 announce first-of-its-kind research partnership
    • Published: 2026-07-03 22:25 Beijing Time
    • Summary: - Google DeepMind and A24 announce their first research partnership.
      • This article from the Google DeepMind blog explains how the first-of-its-kind research partnership announced by Google DeepMind and A24 will shape the broader artificial intelligence and infrastructure landscape.
      • Following the announcement of this first-of-its-kind research partnership by Google DeepMind and A24, it also brings practical implications for founders, operators, and investors.
    • EN Highlights:
      • Google DeepMind and A24 announce first-of-its-kind research partnership

Two Minute Papers (B_intro+search) Link to heading

  • They Said This Will Never Run In Real Time
    • Published: 2026-07-04 01:19 Beijing Time
    • Summary: - ❤️ Check out Weights & Biases and sign up for a free demo here:.
      • 📝 The paper is available here:.
      • Adam Bridges, Benji Rabhan, B Shang, Cameron Navor, Charles Ian Norman Venn, Christian Ahlin, Eric T, Fred R, Gordon Child, Juan Benet, Michael Tedder, Owen Skarpness, Richard Sundvall, Ryan Stankye, Shawn Becker, Steef, Taras Bobrovytsky, Tazaur Sagenclaw, Tybie Fitzhugh, Ueli Gallizzi.
      • They said this will never run in real time.
    • EN Highlights:
      • ❤️ Check out Weights & Biases and sign up for a free demo here:
      • 📝 The paper is available here:
      • 🙏 We would like to thank our generous Patreon supporters who make Two Minute Papers possible:
      • Adam Bridges, Benji Rabhan, B Shang, Cameron Navor, Charles Ian Norman Venn, Christian Ahlin, Eric T, Fred R, Gordon Child, Juan Benet, Michael Tedder, Owen Ska…

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

  • PACE: A Neuro-Symbolic Framework for Plausible and Actionable Counterfactual Explanations

    • Published: 2026-07-03 12:00 Beijing Time
    • Summary: - arXiv:2607.01306v1 Announcement Type: new.
      • Abstract: Counterfactual explanations explain machine learning predictions by identifying minimal input changes that would alter a model’s decision.
      • While many existing methods successfully generate alternatives that change the prediction, they often produce unrealistic or infeasible recommendations due to the lack of an explicit mechanism to integrate domain knowledge and intervention constraints.
      • Neuro-symbolic AI offers a promising direction by combining data-driven predictive models with symbolic reasoning capable of representing human-understandable rules and feasible actions.
    • EN Highlights:
      • arXiv:2607.01306v1 Announce Type: new
      • Abstract: Counterfactual explanations explain machine learning predictions by identifying minimal input changes that would alter a model’s decision

  • Although many existing methods successfully generate prediction-changing alternatives, they often produce unrealistic or infeasible recommendations due to a lac…

  • Neuro-symbolic AI offers a promising direction by combining data-driven predictive models with symbolic reasoning capable of representing human-understandable r…

  • Auto-FL-Research: Agentic Search for Federated Learning Algorithms

    • Publication Time: 2026-07-03 12:00 Beijing Time
    • Abstract: - arXiv:2607.01366v1 Announcement Type: New.
      • Abstract: Federated Learning (FL) research often relies on many small but important algorithmic choices: optimizer variants, server aggregation rules, local training schedules, normalization, regularization, and model architecture.
      • Manually exploring these choices is costly, and it is difficult to make fair comparisons when a candidate’s changes can also alter the FL training or evaluation path.
      • In this work, we present Auto-FL-Research (AFR), a constrained coding-agent workflow for FL algorithm recipe search.
    • EN Highlights:
      • arXiv:2607.01366v1 Announce Type: new
      • Abstract: Federated learning (FL) research often depends on many small but consequential algorithmic choices: optimizer variants, server aggregation rules, loca…
      • These choices are expensive to explore manually and difficult to compare fairly when candidate changes can also alter the FL training or evaluation path
      • In this work, we present Auto-FL-Research (AFR), a constrained coding-agent workflow for FL algorithmic recipe search

  • The Wiola Architecture for Efficient Small Language Models

    • Publication Time: 2026-07-03 12:00 Beijing Time
    • Abstract: - arXiv:2607.01394v1 Announcement Type: New.
      • Abstract: We propose Wiola, a completely original small language model (SLM) architecture built from first principles that shares no structural lineage with any existing model series, including GPT, LLaMA, Mistral, or Falcon.
      • Wiola introduces five independent novel components: (i) Spiral Rotational Position Encoding (SRPE), which embeds token positions on a 3D helical manifold combining absolute, relative, and hierarchical position signals; (ii) Gated Cross-Layer Attention (GCLA), which provides each decoder layer with soft cross-attention access to a compressed summary of the two preceding layers for inter-layer coherence; (iii) Adaptive Token Merging (ATM), which dynamically merges semantically redundant adjacent tokens in intermediate network layers to reduce attention complexity without information loss; (iv) Dual-Stream Feed-Forward (DSFF), which replaces the traditional MLP with two parallel streams fused by a learned per-dimension gate; and (v) WiolaRMSNorm, an improved normalization that introduces a learned per-dimension offset vector to prevent representational collapse.
      • We provide a full mathematical derivation, architectural diagrams, complexity analysis, and a systematic comparison with GPT-2, LLaMA-2, and Mistral.
    • EN Highlights:
      • arXiv:2607.01394v1 Announce Type: new

  • Abstract: We present Wiola, a fully original Small Language Model (SLM) architecture built from first principles, sharing no structural lineage with any existin…

  • Wiola introduces five independently novel components: (i) Spiral Rotary Positional Encoding (SRPE), which embeds token positions on a three-dimensional helical…

  • We provide complete mathematical derivations, architectural block diagrams, complexity analyses, and systematic comparisons against GPT-2, LLaMA-2, and Mistral

  • Agent4cs: A Multi-agent System for Code Summarization in Large Hierarchical Codebases

    • Release Time: 2026-07-03 12:00 Beijing Time
    • Abstract: - arXiv:2607.01425v1 Announcement Type: New.
      • Abstract: Understanding large, complex codebases, especially those with obfuscated structures and incomplete documentation, remains a significant challenge.
      • Existing code summarization solutions often rely on a single language model or coding assistants like Claude Code, and treat source code as flat text, underutilizing the rich interdependencies and hierarchical information stored in the repository.
      • To address these shortcomings, we propose Agent4cs - a multi-agent framework that summarizes large codebases in a bottom-up fashion, where a summarization agent focuses on generating reliable summaries; a keyword extraction agent proactively identifies key information in subfolders; and a quality assurance agent iteratively improves the readability, coherence, and completeness of the output.
    • EN Highlights:
      • arXiv:2607.01425v1 Announce Type: new
      • Abstract: Understanding large, complex codebases, especially those with obfuscated structures and incomplete documentation, remains a significant challenge
      • Existing code summarization solutions often rely on a single language model or coding assistant like Claude Code, and treat source code as flat text, underutili…
      • To address these shortcomings, we propose Agent4cs - a multi-agent framework that summarizes large codebases in a bottom-up fashion, where a summarization agent…

  • When Should Service Agents Reconsider? Difficulty-Routed Control in Customer-Service Operations

    • Release Time: 2026-07-03 12:00 Beijing Time
    • Abstract: - arXiv:2607.01426v1 Announcement Type: New.
      • Abstract: Autonomous customer service agents are transitioning from conversational interfaces to operational execution roles: they retrieve company records, apply service policies, and perform backend writes, such as refunds, cancellations, exchanges, order modifications, and booking changes.
      • This transition creates a service control problem: companies must maintain fast and low-friction daily service while preventing operational errors arising from the interaction of customer instructions, policy constraints, company records, and backend writes.
      • We propose a difficulty-routed service control architecture, asking when service agents should reconsider before taking action.
    • EN Highlights:
      • arXiv:2607.01426v1 Announce Type: new

  • CreativityNeuro: Steering Language Model Weights to Improve Divergent Thinking and Reduce Mode Collapse

    • Posted: 2026-07-03 12:00 Beijing Time
    • Abstract: - arXiv:2607.01433v1 Announce Type: new.
      • Abstract: Divergent thinking is a crucial aspect of creativity, yet Large Language Models (LLMs) often consistently produce similar answers to open-ended questions, a phenomenon known as the artificial hive mind effect.
      • Here, we introduce CreativityNeuro, a data-free method to enhance divergent thinking in LLMs through contrastive weight steering.
      • We evaluate our method using multiple creativity assessments and report several key findings.
    • EN Key Points:
      • arXiv:2607.01433v1 Announce Type: new
      • Abstract: Divergent thinking is a crucial aspect of creativity, yet large language models (LLMs) tend to consistently generate similar responses to open-ended q…
      • Here, we introduce CreativityNeuro, a data-free method for enhancing divergent thinking in LLMs via contrastive weight steering
      • We evaluate our method across multiple creativity assessments and report several main findings

  • Discrete Diffusion Language Models for Interactive Radiology Report Drafting

    • Posted: 2026-07-03 12:00 Beijing Time
    • Abstract: - arXiv:2607.01436v1 Announce Type: new.
      • Abstract: Diffusion language models, which generate text by bidirectionally denoising a token canvas instead of emitting tokens from left to right, have become competitive with autoregressive (AR) generation.
      • However, medical foundation models remain almost entirely autoregressive.
      • We adapted the Mixture-of-Experts diffusion language model, DiffusionGemma-26B, and benchmarked it against its same-sized AR sibling, Gemma-4-26B, under an identical LoRA recipe on a medical visual question-answering dataset, scored by a verbose LLM judge.
    • EN Key Points:
      • arXiv:2607.01436v1 Announce Type: new
      • Abstract: Diffusion language models, which generate text by denoising a token canvas bidirectionally instead of emitting tokens left to right, have become compe…
      • Medical foundation models, however, remain almost entirely autoregressive

  • We adapt a mixture-of-experts diffusion language model, DiffusionGemma-26B, and benchmark it against its same-size AR sibling Gemma-4-26B under an identical LoR…

  • Beyond Next-Token Prediction: An RLVR Proof of Concept for Tool-Use Agents on Atlassian Workflows

    • Publication Time: 2026-07-03 12:00 Beijing Time
    • Abstract:- arXiv:2607.01465v1 Announce Type: new.
      • Abstract: Large language models are trained to predict the next token, not to perform actions within a specific API.
      • In niche enterprise SaaS workflows, where success means hitting the right endpoint with the right nested parameters in the right order, this objective mismatch manifests as silent failures: deleting required fields, hallucinating tools, or prematurely stopping after a single read.
      • We ask whether Reinforcement Learning with Verifiable Rewards (RLVR), applied directly to the target environment, can close the gap.
    • EN Key Points:
      • arXiv:2607.01465v1 Announce Type: new
      • Abstract: Large language models are trained to predict the next token, not to act inside a specific API
      • In niche enterprise SaaS workflows – where success means hitting the right endpoint with the right nested arguments in the right order – this objective mismat…
      • We ask whether Reinforcement Learning with Verifiable Rewards (RLVR), applied directly in the target environment, closes the gap

  • World Feedback for Clinical Agents: Diagnosing RL in FHIR Environments

    • Publication Time: 2026-07-03 12:00 Beijing Time
    • Abstract:- arXiv:2607.01470v1 Announce Type: new.
      • Abstract: Clinical protocol-execution tasks (checking lab values, applying thresholds, placing correctly structured FHIR orders) are natural candidates for RL with world feedback: once clinical SMEs encode decision logic into a validator, the validator scores infinite rollouts without per-episode annotation.
      • But applying reinforcement learning requires sound feedback channels and sufficient foundational capability.
      • We audit MedAgentBench v1/v2, find a 41.7% silent-finish ceiling that makes inaction the RL dominant strategy, and construct \textbf{MedAgentBench-v3 (MAB-v3)} (508 tasks, 8.9% ceiling).
    • EN Key Points:
      • arXiv:2607.01470v1 Announce Type: new
      • Abstract: Clinical protocol-execution tasks – checking a lab value, applying a threshold, placing a correctly structured FHIR order – are natural candidates f…
      • But applying RL requires a sound feedback channel and sufficient base capability
      • We audit MedAgentBench v1/v2, find a 41.7% silent-finish ceiling that makes inaction the RL dominant strategy, and construct \textbf{MedAgentBench-v3 (MAB-v3)}…

  • Procedural Memory Distillation: Online Reflection for Self-Improving Language Models

    • Published: 2026-07-03 12:00 Beijing Time
    • Abstract: - arXiv:2607.01480v1 Announcement Type: new.
      • Abstract: Reinforcement learning with verifiable rewards (RLVR), along with recent self-distillation variants such as SDPO, evaluates each deployment against a validator and updates the policy based on this episode-level signal.
      • However, the richer procedural information in the deployment is rarely retained or reused.
      • Across episodes and epochs, the model repeatedly encounters related problems under a changing policy, producing cross-episode signals that episode-local updates cannot capture: which strategies consistently pass verification, which failure modes persist, and which patterns recur.
    • EN Highlights:
      • arXiv:2607.01480v1 Announce Type: new
      • Abstract: Reinforcement learning with verifiable rewards (RLVR), along with recent selfdistillation variants such as SDPO, evaluates each rollout against a veri…
      • However, the richer procedural information in the rollout is rarely retained or reused
      • Across episodes and epochs, the model repeatedly encounters related problems under a changing policy, producing cross-episode signals that episode-local updates…

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

  • TokenScope: Token-Level Explainability and Interpretability for Code-Oriented Tasks in Large Language Models

    • Published: 2026-07-03 12:00 Beijing Time
    • Abstract: - arXiv:2607.01235v1 Announcement Type: new.
      • Abstract: Understanding how Large Language Models (LLMs) make token-level decisions during code generation remains a significant challenge for both researchers and practitioners.
      • While recent tools provide insights into model internals or generation outcomes, they often lack decoding-time signals, fine-grained uncertainty measurements, and interactive mechanisms for exploring alternative generation paths.
      • We introduce TokenScope, an interactive interpretation and analysis tool for decoder-based LLMs that exposes token-level metrics, attention patterns, and structural information during the generation process.
    • EN Highlights:
      • arXiv:2607.01235v1 Announce Type: new
      • Abstract: Understanding how Large Language Models (LLMs) make token-level decisions during code generation remains a major challenge for both researchers and pr…
      • While recent tools provide insights into model internals or generation outcomes, they often lack decoding-time signals, fine-grained uncertainty measures, and i…
      • We present TokenScope, an interactive interpretability and analysis tool for decoder-based LLMs that exposes token-level metrics, attention patterns, and struct…

  • Safeguarding LLM Agents from Misalignment through Provenance Analysis

    • Published: 2026-07-03 12:00 Beijing Time

  • Abstract: - arXiv:2607.01236v1 Announce Type: new.

    • Abstract: As LLM agents increasingly use powerful tools, ensuring their actions remain aligned with the user’s intent becomes crucial.
    • When an agent’s proposed tool call deviates from the user’s intent (a phenomenon known as misalignment), it can lead to harmful consequences that are difficult to reverse.
    • Existing runtime guardrails rely on an LLM-as-a-judge paradigm, which lacks a systematic framework for reasoning about alignment and often produces inconsistent or difficult-to-audit judgments.

  • EN Highlights:

    • arXiv:2607.01236v1 Announce Type: new
    • Abstract: As LLM agents gain increasing access to powerful tools, ensuring that their actions are aligned with the user’s intent becomes critical
    • When an agent’s proposed tool invocation deviates from the user’s intent – a phenomenon called misalignment – it may lead to harmful consequences that are dif…
    • Existing runtime guardrails rely on an LLM-as-a-judge paradigm that lacks a systematic framework for reasoning about alignment, often producing judgments that a…

  • Kara: Efficient Reasoning LLM Serving via Sliding-Window KV Cache Compression

    • Publication Time: 2026-07-03 12:00 Beijing Time
    • Abstract: - arXiv:2607.01237v1 Announce Type: new.
    • Abstract: Reasoning language models often generate long chains of thought (CoT), accumulating a large KV cache during the decoding phase, which leads to high decoding latency and limited throughput.
    • To address these issues, KV cache compression has emerged as a promising technique to reduce memory overhead by selectively removing unimportant KV pairs while retaining useful ones for subsequent decoding.
    • Nevertheless, we identify two key limitations in existing KV cache compression methods: 1) their threshold-triggered compression strategy may offer limited throughput improvements or even reduce throughput, and can completely eliminate KV pairs from certain blocks of the sequence, potentially exacerbating information loss.
    • EN Highlights:
      • arXiv:2607.01237v1 Announce Type: new
      • Abstract: Reasoning language models often generate long chain-of-thought (CoT), which accumulates a massive KV cache during the decoding phase and incurs high d…
      • To address these issues, KV cache compression has emerged as a promising technique for reducing memory overhead by selectively removing unimportant KV pairs whi…
      • Nevertheless, we identify two key limitations in existing KV cache compression methods: 1) their threshold-triggered compression policy may provide limited thro…

  • SPARCLE: SPeaker-aware Aligned Representations via Contrastive Language Embeddings

    • Publication Time: 2026-07-03 12:00 Beijing Time
    • Abstract: - arXiv:2607.01238v1 Announce Type: new.
    • Abstract: Recent advancements in speech synthesis have shifted from phoneme-based representations to direct grapheme modeling.

  • While phonemes address the one-to-many mapping between text and acoustics, they rely on grapheme-to-phoneme (G2P) systems that fail to capture speaker-specific acoustic variations.

    • Previous work shows that grapheme-based models outperform phoneme-based systems at scale, but not in low-resource settings.
    • EN 要点:
      • arXiv:2607.01238v1 Announce Type: new
      • Abstract: Recent advances in speech synthesis have shifted from phoneme representations to direct grapheme modeling
      • While phonemes address the one-to-many mapping between text and acoustics, they rely on grapheme-to-phoneme (G2P) systems that fail to capture speaker-specific…
      • Prior work demonstrates that grapheme-based models outperform phoneme-based systems at scale, but not in low-resource settings

  • Breaking Safety at the Token Boundary: How BPE Tokenization Creates Exploitable Gaps in LLM Alignment

    • Published: 2026-07-03 12:00 Beijing Time
    • Abstract: - arXiv:2607.01239v1 Announcement Type: new.
      • Abstract: Character-level perturbations bypass safety alignment in modern LLMs despite leaving prompts human-readable.
      • We identify and test a central structural mechanism: BPE tokenization fragments safety-critical words into sub-word pieces, and the three public alignment datasets we investigated do not contain intentionally fragmented inputs.
      • The mechanism is a chain, tested end-to-end on five model families (Qwen-3-4B, Qwen-2.5-7B, Gemma-3-4B, Llama-3.1-8B, Mistral-7B).
    • EN 要点:
      • arXiv:2607.01239v1 Announce Type: new
      • Abstract: Character-level perturbations bypass safety alignment in modern LLMs despite leaving prompts human-readable
      • We identify and test a central structural mechanism: BPE tokenization fragments safety-critical words into sub-word pieces, and the three public alignment datas…
      • The mechanism is a chain, tested end-to-end on five model families (Qwen-3-4B, Qwen-2.5-7B, Gemma-3-4B, Llama-3.1-8B, Mistral-7B)

  • Prompt Framing Distorts Count-Based Evaluation of LLM Error Detection: Evidence from Numeric Anchoring

    • Published: 2026-07-03 12:00 Beijing Time
    • Abstract: - arXiv:2607.01240v1 Announcement Type: new.
      • Abstract: Count-based F1 is widely used as a proxy for LLM error detection quality, but this paper shows that it can sharply increase without corresponding improvement in span localization, a gap referred to as F1 inflation.
      • This paper introduces ErrorBench, a controlled stress test protocol for prompt-induced count distortion.
      • ErrorBench evaluated six contemporary LLMs on 4,290 responses from 143 CoNLL-2014 paragraphs under five prompting conditions.
    • EN 要点:
      • arXiv:2607.01240v1 Announce Type: new

  • Abstract: Count-based F1 is widely used as a proxy for LLM error-detection quality, but this paper shows that it can rise dramatically without a corresponding i…

  • The paper introduces ErrorBench, a controlled stress-test protocol for prompt-induced count distortion

  • ErrorBench evaluates six contemporary LLMs under five prompt conditions over 4,290 responses from 143 CoNLL-2014 passages

  • Mapping Text to Multiplex Graph: Prompt Compression as L'evy Walk-Guided Graph Pruning

    • Publication Time: 2026-07-03 12:00 Beijing Time
    • Abstract: - arXiv:2607.01241v1 Announcement Type: New.
      • Abstract: Existing prompt compression methods treat text as flat token sequences, failing to capture the distributed nature of important information, which is often spread across multiple locations and connected by local syntactic dependencies and global semantic relationships.
      • This relational structure is naturally represented as a graph, where tokens or sentences become nodes and their dependencies become edges.
      • To this end, we propose RAGP, which formulates prompt compression as Redundancy-Aware Graph Pruning on a multiplex graph that jointly models fine-grained attention-based dependencies and coarse-grained semantic relationships.
    • EN Key Points:
      • arXiv:2607.01241v1 Announce Type: new
      • Abstract: Existing prompt compression methods treat text as flat token sequences, failing to capture the distributed nature of important information, which is o…
      • Such relational structure is naturally represented as a graph, where tokens or sentences become nodes and their dependencies become edges
      • To this end, we propose RAGP, which formulates prompt compression as Redundancy-Aware Graph Pruning on a multiplex graph that jointly models fine-grained attent…

  • Office Comprehension Benchmark

    • Publication Time: 2026-07-03 12:00 Beijing Time
    • Abstract: - arXiv:2607.01245v1 Announcement Type: New.
      • Abstract: We introduce the Office Comprehension Benchmark (OCB), the first public benchmark to jointly evaluate LLM systems on their understanding of native file formats (.docx, .xlsx, .pptx) and their variants for Word, Excel, and PowerPoint.
      • File-fidelity question answering tests the structural and visual perception of office artifacts—tables, charts, embedded images, formulas, and application-specific elements such as titles, speaker notes, and named ranges.
      • Domain-specific question answering tests expert-level reasoning based on real industry documents from 12 professional fields, with queries requiring multi-step analysis and synthesis across documents.
    • EN Key Points:
      • arXiv:2607.01245v1 Announce Type: new
      • Abstract: We introduce Office Comprehension Bench (OCB), the first public benchmark to jointly evaluate LLM systems on Word, Excel, and PowerPoint comprehension…
      • OCB consists of two tracks

  • File Fidelity Q&A tests structural and visual perception of office artifacts - tables, charts, embedded images, formulas, and app-specific elements such as head…

  • RuleChef: Grounding LLM Task Knowledge in Human-Editable Rules

    • Publish Time: 2026-07-03 12:00 Beijing Time
    • Abstract: - arXiv:2607.01293v1 Announcement Type: New.
      • Abstract: We present RuleChef, a framework that uses large language models (LLMs) to generate executable rules for NLP tasks such as text classification, Named Entity Recognition (NER), or relation extraction.
      • Rules are generated based on a task description and a set of labeled examples, then iteratively improved based on additional examples and human feedback on existing rules.
      • RuleChef can also be used to bootstrap rules using observed input-output pairs from any existing model for a given task.
    • EN Highlights:
      • arXiv:2607.01293v1 Announce Type: new
      • Abstract: We present RuleChef, a framework that uses large language models (LLMs) to generate executable rules for NLP tasks such as text classification, Named…
      • Rules are generated based on a task description and a set of labeled examples, then they are iteratively improved based both on additional examples and on human…
      • RuleChef can also be used to bootstrap rules using the observed input-output pairs from any existing model for a given task

  • TurnNat: Automatic Evaluation of Turn-Taking Naturalness in Dyadic Spoken Dialogue

    • Publish Time: 2026-07-03 12:00 Beijing Time
    • Abstract: - arXiv:2607.01345v1 Announcement Type: New.
      • Abstract: Turn-taking naturalness is central to full-duplex spoken dialogue systems, yet its automatic evaluation remains limited.
      • Existing evaluations often rely on human judgments or behavior-specific timing metrics, making it difficult to compare heterogeneous timing failures within a unified framework.
      • We propose TurnNat, a likelihood-based framework for automatic turn-taking naturalness evaluation in dyadic spoken dialogue.
    • EN Highlights:
      • arXiv:2607.01345v1 Announce Type: new
      • Abstract: Turn-taking naturalness is central to full-duplex spoken dialogue systems, yet its automatic evaluation remains limited
      • Existing evaluations often rely on human judgments or behavior-specific timing metrics, making it difficult to compare heterogeneous timing failures within a un…
      • We propose TurnNat, a likelihood-based framework for automatic turn-taking naturalness evaluation in two-channel spoken dialogue

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

  • Multilayer Q-Matrix-Embedded Neural Network for Cognitive Diagnosis (M-QCDNet): Structure-Aware Deep Learning Architecture for Psychometric Interpretability

    • Publication Time: 2026-07-03 12:00 Beijing Time
    • Summary: - arXiv:2607.01278v1 Announcement Type: new.
      • Abstract: The study proposes a multilayer Q-matrix-embedded neural network (M-QCDNet) for cognitive diagnosis, which combines the structural interpretability of cognitive diagnosis models (CDMs) with deep learning neural networks (NNs).
      • M-QCDNet uses the Q-matrix as a structural prior to construct item-skill relationships, ensuring that latent mastery profiles remain interpretable and consistent with cognitive theory. It then proposes a loss function with an L2 penalty to penalize skills that do not align with the Q-matrix and to balance predictive performance with structural alignment.
      • Furthermore, corresponding evaluation matrices, which are interpretable alignment-based metrics, were developed to quantify the degree to which predicted skill activations correspond to item-level skills.
    • EN Key Points:
      • arXiv:2607.01278v1 Announce Type: new
      • Abstract: The research proposes a multilayer Q-matrix-embedded neural network for cognitive diagnosis (M-QCDNet), which integrates the structural interpretabili…
      • M-QCDNet structures the item-skill relationship using the Q-matrix as a structural prior, ensuring latent mastery profiles remain interpretable and consistent w…
      • Corresponding evaluation matrices, the interpretable alignment-based metrics that quantify the degree to which predicted skill activations correspond to item-le…

  • I\textsuperscript{2}RiMA: Spectral Riemannian Representation with Temporal Attention for Mental Stress Detection based on EEG Signals

    • Publication Time: 2026-07-03 12:00 Beijing Time
    • Summary: - arXiv:2607.01279v1 Announcement Type: new.
      • Abstract: Cross-subject EEG stress detection remains challenging because discriminative stress-related patterns are both subject-dependent and frequency-dependent.
      • Traditional Riemannian methods primarily model spatial covariance in the time domain, overlooking neural oscillations that are crucial for decoding high-level cognitive states, while standard temporal tokenization often disrupts inter-segment temporal coherence.
      • To address these limitations, we propose \method{}, an inner Riemannian manifold attention network for EEG-based stress detection.
    • EN Key Points:
      • arXiv:2607.01279v1 Announce Type: new
      • Abstract: Cross-subject EEG stress detection remains challenging because discriminative stress-related patterns are both subject-dependent and frequency-specifi…
      • Conventional Riemannian methods model spatial covariance mainly in the time domain, overlooking neural oscillations that are critical for high-level cognitive s…

  • To address these limitations, we propose \method{}, an Intra-Inter Riemannian Manifold Attention Network for EEG-based stress detection

  • Fixed-Set Robustness in Programming by Example: Example Corruption and Semantic Partition Recovery

    • Publication Time: 2026-07-03 12:00 Beijing Time
    • Abstract: - arXiv:2607.01280v1 Announce Type: new.
      • Programming-by-example systems infer programs from a small set of input-output examples.
      • Robust PBE work usually models wrong examples as samples from a stochastic noise process and then minimizes an expected or empirical loss.
      • This paper studies a different failure mode: an adversary who sees the synthesizer and chooses the examples whose corruption most damages the returned program.
    • EN Highlights:
      • arXiv:2607.01280v1 Announce Type: new
      • Abstract: Programming-by-example systems infer programs from a small set of input-output examples
      • Robust PBE work usually models wrong examples as samples from a stochastic noise process and then minimizes an expected or empirical loss
      • This paper studies a different failure mode: an adversary who sees the synthesizer and chooses the examples whose corruption most damages the returned program

  • Domain Knowledge Based Temporal-Spatial Graph Convolution Network for ECG Recognition

    • Publication Time: 2026-07-03 12:00 Beijing Time
    • Abstract: - arXiv:2607.01282v1 Announce Type: new.
      • Given the advancements in Artificial Intelligence (AI) and its widespread applications, challenges persist in the interpretability of AI models, particularly in specialized domains like healthcare, such as electrocardiogram (ECG) recognition.
      • This paper introduces a novel approach for ECG recognition using a domain knowledge-based graph convolutional network, rather than relying solely on end-to-end convolutional neural networks.
      • Key landmark points of PRQST, vital to ECG interpretation, are incorporated as domain knowledge.
    • EN Highlights:
      • arXiv:2607.01282v1 Announce Type: new
      • Abstract: In light of strides in Arti cial Intelligence (AI) and its wide spread application, challenges persist in the interpretability of AI models, particula…
      • Rather than relying solely on end-to-end convolutional neural networks, this paper introduces a novel approach using a domain knowledge-based graph convolution…
      • Key landmarks points of PRQST, vital to ECG interpreta tion, are incorporated as domain knowledge

  • Scaling Laws for Grid-Based Approximate Nearest Neighbor Search in High Dimensions

    • Publication Time: 2026-07-03 12:00 Beijing Time
    • Abstract: - arXiv:2607.01283v1 Announce Type: new.

  • Abstract: Grid-based approaches to approximate nearest neighbor (ANN) search have been absent from modern scaling analyses.

  • We present a systematic characterization of a multiprobe grid algorithm with respect to dataset size $N$ and dimensionality $d$.

  • Our experiments reveal a previously unreported $d$-scaling crossover on the GloVe embedding family, in which multiprobe grid search maintains an approximately constant dimensionality scaling exponent, while other graph, tree, and partition-based methods exhibit decreasing throughput.

    • EN Highlights:
      • arXiv:2607.01283v1 Announce Type: new
      • Abstract: Grid-based approaches to approximate nearest neighbor (ANN) search have been absent from modern scaling analyses
      • We present a systematic characterization of a multiprobe grid algorithm with respect to dataset size $N$ and dimensionality $d$
      • Our experiments reveal a previously unreported $d$-scaling crossover on the GloVe embedding family, in which multiprobe grid search maintains an approximately c…

  • IonSense-QKG: A Quantum-Readiness Metadata Framework for Lithium-Ion Battery Dataset Discovery

    • Publication Time: 2026-07-03 12:00 Beijing Time
    • Abstract: - arXiv:2607.01286v1 Announce Type: new.
      • Abstract: Public lithium-ion battery datasets are increasingly used for state-of-health estimation, remaining-useful-life prediction, anomaly detection, electrochemical diagnostics, second-life analysis, and battery safety research.
      • However, these datasets vary substantially in chemistry, modality, scale, label quality, sequence structure, access status, and preprocessing complexity.
      • These differences directly affect whether a dataset is feasible for near-term hybrid quantum-classical machine-learning workflows.
    • EN Highlights:
      • arXiv:2607.01286v1 Announce Type: new
      • Abstract: Public lithium-ion battery datasets are increasingly used for state-of-health estimation, remaining-useful-life prediction, anomaly detection, electro…
      • However, these datasets vary substantially in chemistry, modality, scale, label quality, sequence structure, access status, and preprocessing complexity
      • These differences directly affect whether a dataset is feasible for near-term hybrid quantum-classical machine-learning workflows

  • A Novel Machine Learning Approach for Central Nervous System Tumor Classification from DNA Methylation

    • Publication Time: 2026-07-03 12:00 Beijing Time
    • Abstract: - arXiv:2607.01307v1 Announce Type: new.
      • Abstract: DNA methylation analysis has emerged as a powerful approach for central nervous system (CNS) tumor classification, yet important challenges remain regarding cross-cohort transferability, methodological correctness, and robust multi-class evaluation.
      • In this work, we propose a novel and methodologically rigorous machine learning approach for methylation-based CNS tumor classification, which combines sparse random projections for dimensionality reduction with multinomial logistic regression for classification.
      • We evaluate the proposed method within the same general experimental setup established by a widely used reference classifier.
    • EN Highlights:
      • arXiv:2607.01307v1 Announce Type: new

  • Abstract: NA methylation profiling has become a powerful approach for central nervous system (CNS) tumor classification, yet important challenges remain regardi…

  • In this work, we propose a novel and methodologically rigorous machine-learning approach for methylation-based CNS tumor classification that combines Sparse Ran…

  • We evaluate the proposed approach in the same general experimental setting established by a widely used reference classifier

  • From Approximation to Emergence: A Theory of Deep Learning

    • Publication Time: 2026-07-03 12:00 Beijing Time
    • Abstract: - arXiv:2607.01311v1 Announcement Type: new.
      • Abstract: Deep learning has outgrown any single mathematical explanation.
      • “From Approximation to Emergence” develops a unified, proof-oriented account of modern deep learning theory, tracing a path from the classical foundations of approximation, optimization, and generalization to contemporary mechanisms of hyperparameterization, robustness, generative modeling, transformers, in-context learning, scaling laws, interpretability, alignment, and emergence.
      • Rather than presenting isolated results, the book organizes a broad literature into a coherent research narrative: each theory is examined through the object it governs, the assumptions that make it effective, and the phenomena it fails to explain.
    • EN Key Points:
      • arXiv:2607.01311v1 Announce Type: new
      • Abstract: Deep learning has outgrown any single mathematical explanation
      • From Approximation to Emergence develops a unified, proof-oriented account of modern deep learning theory, tracing a path from the classical foundations of appr…
      • Rather than presenting isolated results, the book organizes a broad literature into a coherent research narrative: each theory is examined through the object it…

  • Black-Box Inference of LLM Architectural Properties with Restrictive API Access

    • Publication Time: 2026-07-03 12:00 Beijing Time
    • Abstract: - arXiv:2607.01313v1 Announcement Type: new.
      • Abstract: In practice, most commercial LLM providers do not publicly release details of the underlying LLM architecture.
      • However, previous work has shown that given limited API access to an LLM (i.e., top-$k$ logits and/or logit bias functions), one can recover certain architectural details of the LLM, such as the hidden dimension of the feed-forward network.
      • Perhaps in response to these findings, most commercial LLM providers have restricted their APIs to expose only a single logit per decoded token, and they no longer provide users with the ability to bias logits.
    • EN Key Points:
      • arXiv:2607.01313v1 Announce Type: new
      • Abstract: In practice, most commercial LLM providers do not publicly release details of underlying LLM architectures

  • However, prior work has shown that given limited API access to an LLM (namely, top-$k$ logits and/or a logit bias function), one can recover certain architectur…

  • Perhaps in response to these results, most commercial LLM providers have restricted their APIs to expose only the single logit for each decoded token, and they…

  • Multi-modal Rail Crossing Safety Analysis

    • Publication Time: 2026-07-03 12:00 Beijing Time
    • Abstract:
      • arXiv:2607.01365v1 Announcement Type: new.
      • Abstract: Given one or more images of a railway crossing, can we leverage visual cues to reliably estimate its safety?
      • Can we improve our ability to do so by introducing structured data (such as official accident reports) about the accident history of that crossing into our model?
      • In this work, we explore how to best answer these questions to build an AI system that can ingest multi-modal data for railway crossings and provide safety assessments and scores consistent with expert opinions and safety ratings used by the Federal Railroad Administration (FRA).
    • EN Key Points:
      • arXiv:2607.01365v1 Announce Type: new
      • Abstract: Given one or more images of a railway crossing, can we leverage visual cues that allow us to robustly estimate how safe it is
      • Can we improve our ability to do so by introducing structured data (such as official accident reports) about the accident history of that crossing into our mode…
      • In this work, we explore how to best answer those questions towards building an AI system that can ingest multi-modal data for railway crossings and provide saf…