System translated (Gemini)

🤖 AI 速览

Today’s main theme is not just about models becoming stronger, but about agent engineering entering a battle of details: Are feedback mechanisms truly effective, how can evaluations be more actionable, and how do cost and compliance constrain deployment? Anthropic’s product updates, …
📋 文章元数据
发布时间
2026-07-02
类型
ai-daily
字数
7211
阅读时长
34 min

2026-07-02 AI Daily | Behind Anthropic’s Full-Line Update: Agents Enter a Race for Cost, Compliance, and Reproducibility Link to heading

Today’s main theme isn’t just about models getting stronger, but about agent engineering entering a battle of details: whether feedback is authentic and effective, how to make evaluation more actionable, and how cost and compliance constrain deployment. Anthropic’s product updates, Claude Science, and the controversy over hidden watermarks, all point to the next stage of competition: controllable, reproducible, and trustworthy AI workflows.

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

The most noteworthy topic today is the detailed challenges of “agent engineering”: from optimizing skill descriptions in production environments and the actual gains from natural language feedback, to iterative prompt debugging and test-time verification for Text-to-SQL. Several papers are addressing the same question—it’s not enough for an agent to just run. The key is how to achieve stable routing, controllable improvements, and reduced misjudgments.

The second main theme is the evolution of evaluation paradigms. Works like BayesBench, calibrated fair comparisons, and the mental health-focused TheraJudge are shifting evaluation from static scoring to multi-round belief updating, accuracy control, and actionable quality signals. This is essential reading for teams working on model assessment and safety governance.

Additionally, multilingual and real-world noisy scenarios are gaining prominence. Romanized Indic-English code-mixing, Bengali noisy event detection, and Arabic-Russian scientific corpora all remind us that the next gap in model capabilities may not be found on clean English leaderboards, but in complex, low-resource, and cross-cultural real-world environments.

🌐 AI Hot Topics on X Link to heading

Topic 1: Claude Fable 5 Returns Worldwide After U.S. Lifts AI Export Restrictions Link to heading

  • Category: AI · News
  • Overview: Trending for: 1 day ago, Related posts: 121,000
  • What it is: Anthropic’s Claude Fable 5 has reportedly resumed service globally after the U.S. lifted relevant AI export restrictions.
  • Why it’s important: This highlights how the availability of advanced AI models remains heavily influenced by export controls and geopolitical policies, potentially altering the ability of developers, companies, and research institutions worldwide to access cutting-edge models.
  • Discussion summary: Discussions on X are focused on whether open access will accelerate global AI innovation, whether it undermines U.S. technology controls, and whether the model will face new regulatory pressures regarding performance, security, and compliance upon its return.

Topic 2: Anthropic Launches Claude Sonnet 5 with Strong Agentic Skills Link to heading

  • Category: AI · News
  • Overview: Trending for: 1 day ago, Related posts: 42,000
  • What it is: Anthropic has released Claude Sonnet 5, which features enhanced agentic task execution, programming, and multi-step reasoning capabilities.
  • Why it’s important: This indicates that the competition among major large models is shifting from single-turn Q&A abilities to sustained, agentic task-execution capabilities, which could impact developer tools, enterprise automation, and AI application architecture.
  • Discussion summary: Discussions on X center on whether Sonnet 5 significantly closes the gap with Claude Opus, whether agentic frameworks are becoming the new competitive core, and its cost-effectiveness, programming performance, and competitive relationship with Google and open-source agent tools.

Topic 3: ByteDance’s Seedance 2.0 Delivers 4K AI Video Magic for Creators Link to heading

  • Category: AI · News
  • Overview: Trending for: 16 hours ago, Related posts: 1,400
  • What it is: ByteDance has launched or showcased Seedance 2.0, described as a 4K AI video generation tool for creators.
  • Why it’s important: This suggests that large model video generation is advancing towards higher resolutions and integration into creative workflows, which could accelerate competition in the use of AI video for short films, advertising, and social content production.
  • Discussion summary: Discussions on X mainly focus on its 4K generation quality, stability, and practical value for creators. There is also interest in its competition with video models from OpenAI, Runway, Google, as well as concerns about copyright, cost, and content authenticity risks.

Topic 4: Tesla Brings FSD v14 Lite to Older Hardware 3 Vehicles Link to heading

  • Category: AI · News
  • Overview: Trending for: 2 days ago, Related posts: 48,000
  • What it is: Tesla has begun rolling out FSD v14 Lite to older vehicles equipped with Hardware 3, bringing some of the v14 self-driving capabilities to earlier models.
  • Why it’s important: This demonstrates that AI self-driving models can be distilled and optimized to run on older hardware, which impacts the scalability of in-car AI, user rights, and the commercialization timeline for autonomous driving.
  • Discussion summary: Discussions on X focus on whether HW3 owners are finally receiving the features they were promised, the performance gap between v14 Lite and the full v14, the safety limits of older hardware, and whether Tesla should offer upgrades or compensation for vehicles unable to achieve unsupervised FSD.

AI Public Opinion Summary on X Today Link to heading

The main narrative today is that AI capabilities are rapidly evolving from “usable” to “executable, production-ready, and delegable,” while being continuously shaped by geopolitical policies, hardware constraints, and regulatory risks. The consensus is that advancements in frontier models, agents, 4K video generation, and in-car autonomous driving all indicate that the commercialization of AI is accelerating, potentially benefiting developers, creators, and vehicle owners directly. Disagreements primarily center on whether these developments bring about a truly qualitative change: whether the restoration of Claude services weakens export controls, whether Sonnet 5 is sufficient to challenge higher-end models, whether Seedance 2.0 offers stable creative value, and whether FSD v14 Lite fulfills its promises to owners of older vehicles. Potential risks include diminished cross-border controllability of advanced models, security and liability issues arising from agent execution, the impact of AI video on copyright and authenticity, and safety, compensation, and regulatory disputes caused by the unclear capability boundaries of autonomous driving on older hardware.

💡 Influencer Insights Link to heading

AI Industry Influencer Insights Daily: July 1, 2026 Link to heading

Based on the last 24 hours of activity on the X platform, yesterday was an extremely information-dense day for the AI industry: Anthropic delivered what was almost a “full-family update,” accompanied by privacy controversies, deep evolution in the developer tool ecosystem, and the implementation of Web3 x AI. Here are the key takeaways.

1.1 The Eye of the Storm: Anthropic’s Full Product Line Unleashed and Updated Link to heading

Anthropic’s major moves have captured the attention of all developers, with discussions about its models dominating the conversation.

  • 🚀 Fable 5 Fully Unbanned and Returns: After negotiations with the government, export controls on Claude Fable 5 and Mythos 5 have been lifted. Service was restored on July 1st, with a credit-based billing system to be implemented in 7 days (@dotey). @zhixianio previously marveled at Fable 5’s ultra-high productivity in building demos (completing 70% of the work and identifying design flaws in 40 minutes), which made him willing to “shut up and take my money.” @dotey also mentioned that this lifting of restrictions is accompanied by stronger interception classifiers for cybersecurity tasks (New classifiers), and Anthropic has committed to working with the government to establish risk standards (@dotey).
  • 🧬 Release of Vertical-Specific Tool: Claude Science: Positioned as the “Claude Code for the scientific community,” it integrates over 60 specialized scientific databases to address the pain point of “frequent tool switching” in life sciences. The focus is on supporting reproducibility (code attached to charts) and local computation (sensitive data remains on-premise) (@dotey). Its predecessor and competitor, OpenAI GPT-Rosalind, focused more on domain-specific reasoning models (@dotey).
  • ⚡️ Sonnet 5 Officially Takes Over: Replacing Sonnet 4.6, it features significantly improved agent capabilities, approaching the more expensive Opus 4.8 but at only 40% of its API price. A key detail is the switch to a new tokenizer, which may increase actual token consumption by 35% (the cost is offset by promotional pricing), a point developers should note for cost management. (@dotey).
  • 🔧 Open-Source IDE Orca Emerges: As Claude Code enjoys the spotlight, the community is turning its attention to the open-source competitor IDE Orca (an open-source implementation similar to Codex/CC), which supports multiple platforms and mainstream models (@dotey referencing @LinearUncle).

1.2 Deep Dives and “Vulnerability” Mining in Codex/Claude Code Link to heading

  • Hidden Mechanism “Watermark-gate” Erupts: Claude Code v2.1.196 was exposed through reverse engineering, revealing that it uses Unicode zero-width/similar characters to covertly embed markers (a covert channel) in the system prompts for users on Chinese proxies (domains like Baidu, Alibaba, ByteDance) or in specific time zones. This incident could become a catalyst for industry-wide compliance discussions (@dotey). Subsequently, Anthropic researcher @trq212 responded that this was an old anti-abuse mechanism already scheduled for removal (@dotey).
  • Credit Bugs and Tricks: A blogger discovered a new bug in Codex’s /goal mode where credits are not consumed (@Pluvio9yte); meanwhile, micro-tricks are circulating, such as using a bypass /side to check progress and using Goal mode to wrap up long tasks (@Pluvio9yte referencing @afei_AI).
  • The Asset Flow of Vibe Coding: The “Volcano Engine Coding Plan” officially went on sale for 9.9 RMB/month, signaling that AI programming APIs have become extremely cheap, fast-moving consumer goods (@Pluvio9yte).

1.3 The Foundation for ‘One-Person Companies’ Derived from Web3 + AI Agents is Now in Place Link to heading

  • OKX Launches AI Agent Marketplace: A marketplace where AI agents can autonomously accept tasks, perform work, receive on-chain payments, and hire each other has been launched. The Agent Trade Kit has been open-sourced. Combined with tools like Claude Code, a single person can configure a few AI agents (for analysis, market monitoring, and trading) to match the efficiency of a professional institutional team. This narrative has been validated (@Pluvio9yte, citing @star_okx).

2. Noteworthy Perspectives or Industry Foresight Link to heading

  • @dotey’s “Maker vs. Perceiver” Theory: While commenting on Spotify being mocked by users for promoting its use of Claude Code, he pointed out that the standard for measuring AI’s value should not just be “engineering efficiency metrics”—like 4,500 daily deployments or 73% of PRs being AI-assisted—while ignoring the user’s actual perception. Producing more code without users feeling the product has improved reflects the industry’s misguided governance, indulging in production metrics while neglecting product experience.
  • @zhixianio’s “10B Parameter Bottleneck” Theory: He tested the Gemma 4 12B Coder and found that while fine-tuning enables faster convergence, the model’s 12B parameter size creates an insurmountable ceiling when handling long, stateful, and complex programs generated in a single pass (e.g., it fails to code Tetris correctly). He emphasizes that “sweet spot” models with large parameter counts and sparse activation, like the Qwen 35B MoE, still reign supreme for on-device and consumer-grade development.
  • @gefei55’s “Trojan Horse Pricing”: This strategy uses the anchoring effect in tiered pricing (e.g., 50 items for ¥34.9, with an option to get 50 more for an additional ¥10). He maps this strategy to SaaS monthly/annual plan design, effectively boosting average customer value by nudging users toward the “second-cheapest, high-value package.”
  • @ruanyf’s Questioning of “Moats”: He relayed the story of a Cloudflare engineer who replicated Next.js using AI for only $1,100 in token fees. Under the disruptive pressure of AI, traditional code barriers are disappearing. The only remaining moats to prevent competitors from devouring the market may be “test cases” and “integration experience with complex scenarios.”
  • @lijigang’s “Token Heterogeneity” Theory: He issues an economic warning that tokens are not a homogeneous commodity like electricity. Tokens from different models possess drastically different values and cognitive capabilities. Top-tier model tokens will resist commoditization and instead become a supply-bottlenecked monopoly, akin to a form of tax.
  • @vista8 on 3Blue1Brown’s Summary of AI’s Ceiling: He argues that AI’s most underestimated strength is its parallel computation capability, not its intelligence. It excels at rapidly connecting existing human knowledge but remains weak at low-probability events like “creating entirely new conceptual frameworks for the world” (paradigm shifts).

🧠 Development & Frontier Tools Link to heading

  • Claude Science: An agent workbench for life science computation (@dotey). It addresses the challenge computational biologists face by eliminating the need to constantly switch between documentation, code, and clusters. It comes with 60+ built-in databases and reproducible code.
  • Orca (Coding IDE): A cross-platform, open-source code editor frequently recommended by experts. It serves as a desktop open-source alternative to Claude Code/Codex, ideal for developers who prefer not to use “black box” tools (@dotey/@LinearUncle).
  • DevSpace (MCP Tunnel): A fascinating new open-source project that exposes a local MCP server to the ChatGPT web interface, granting it read/write access to your local code. This enables you to use the GPT-5.5 Pro model within ChatGPT, effectively doubling your quota (@gefei55).
  • Frontend Animation Lexicon Skill: Includes three open-source Skill libraries: animation-vocabulary (look up terms), emil-design-eng (refine textures), and review-animations (review motion effects). It helps you avoid the awkward situation of being unable to implement advanced animations simply because you don’t know the correct professional terminology (@vista8).

🤖 Efficiency & Business Link to heading

  • X Agent Flow Miner: A small open-source tool from @gefei55. It calls a low-cost X API to scan trending links and reverse-trace their SEO traffic sources. It can identify popular domains that are just beginning to trend before they hit Google Trends and turn them into traffic-generating sites. It also includes recommendations for affordable and reliable Twitter API providers.
  • OKX MCP Plugin / Agent Trade Kit: Fully open-source and compatible with Claude Code/Cursor. The AI can autonomously execute trades using 164 professional tools, including spot, grid, and futures trading (@Pluvio9yte).
  • ChatGPT Conversation Cleaner Plugin: A completely free, local Chrome extension developed by @Pluvio9yte for bulk deleting or archiving conversations in the ChatGPT sidebar—a lifesaver for the organizationally obsessed.

🎨 Creation & Infrastructure Link to heading

  • Open-Source AI Video Skill Library: @Pluvio9yte has open-sourced a complete suite of skills to replicate HyperFrames-style video generation, a “video fine-tuning” benefit for those unfamiliar with AE/PR.
  • ARGUS Cluster Monitoring Solution: The Tencent team has open-sourced a solution for automatically locating faults in ten-thousand-GPU clusters. The core finding is that 70% of training interruptions are caused by network packet loss and routing failures, not faulty cards. Suitable for large model infra teams doing private deployments (@vista8).

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

Time frame: Last 3 days; covers 22 sources; 31 updates in total

Two Minute Papers (B_intro+search) Link to heading

  • AI Just Entered A New Era
    • Published: 2026-07-01 13:23 Beijing Time
    • Abstract: - ❤️ Check out Lambda here and sign up for their GPU Cloud:.
      • 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.
    • EN Highlights:
      • ❤️ Check out Lambda here and sign up for their GPU Cloud:
      • 🙏 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

  • What Drives Interactive Improvement from Feedback?

    • Published: 2026-07-01 12:00 Beijing Time
    • Abstract: - arXiv:2606.30774v1 Announcement Type: New.
      • Abstract: We study when natural language feedback produces improvements beyond the gains obtainable from repeated attempts alone.
      • In a multi-turn language agent setting, higher final accuracy can reflect useful feedback, but it can also arise from resampling, format correction, or additional test-time computation.
      • To distinguish these effects, we introduce a controlled student-teacher protocol across Omni-MATH, Codeforces, BBEH Linguini, and ARC-AGI1, evaluating thirteen open-weight models in student and teacher roles.
    • EN Highlights:
      • arXiv:2606.30774v1 Announce Type: new
      • Abstract: We study when natural-language feedback produces improvement beyond the gains obtainable from repeated attempts alone
      • In multi-turn language agent setting, higher final accuracy can reflect useful feedback, but it can also arise from resampling, format correction, or additional…
      • To separate these effects, we introduce a controlled student-teacher protocol across Omni-MATH, Codeforces, BBEH Linguini, and ARC-AGI1, evaluating thirteen ope…

  • Contrastive Reflection for Iterative Prompt Optimization

    • Publication Time: 2026-07-01 12:00 Beijing Time
    • Abstract: - arXiv:2606.30840v1 Announcement Type: new.
      • Abstract: LLM agents are becoming central to information retrieval: they issue retrieval queries, synthesize answers, and increasingly serve as judges for IR evaluation.
      • Improving the prompts that control these agents is an optimization problem, but in applied IR settings, it often looks less like blind search and more like debugging.
      • Engineers need to know which behavior failed, which nearby behavior still worked, what distinguishes the two, and whether a prompt edit can improve held-out quality without introducing regressions.
    • EN Key Points:
      • arXiv:2606.30840v1 Announce Type: new
      • Abstract: LLM agents are becoming central to information retrieval: they issue retrieval queries, synthesize answers, and increasingly serve as judges for IR ev…
      • Improving the prompts that control these agents is an optimization problem, but in applied IR settings it often looks less like blind search and more like debug…
      • Engineers need to know which behavior failed, which nearby behavior still worked, what distinguishes the two, and whether a prompt edit improves held-out qualit…

  • How Can AI Find My Model? A Model-Finding Experimental Study Considering Data Formats, Embeddings, and Retrieval Strategies

    • Publication Time: 2026-07-01 12:00 Beijing Time
    • Abstract: - arXiv:2606.30846v1 Announcement Type: new.
      • Abstract: Discovering reusable simulation models remains a fundamental challenge in the Modeling and Simulation (M&S) field.
      • When many models coexist, identifying those that align with a given modeling intent remains difficult.
      • Recent advances in Artificial Intelligence (AI), particularly retrieval-based approaches, offer a promising pathway for operating on this semantic layer.
    • EN Key Points:
      • arXiv:2606.30846v1 Announce Type: new
      • Abstract: Discovering simulation models for reuse remains a fundamental challenge in Modeling and Simulation (M&S)
      • When many models coexist, identifying those that align with a given modeling intent remains difficult
      • Recent advances in Artificial Intelligence (AI), particularly retrieval-based approaches, offer a promising pathway to operate at this semantic layer

  • BayesBench: Evaluating LLM Belief Trajectories Under Multi-Turn Evidence Accumulation

    • Publication Time: 2026-07-01 12:00 Beijing Time
    • Abstract: - arXiv:2606.30850v1 Announcement Type: new.
      • Abstract: Large Language Models (LLMs) are often deployed in multi-turn dialogues, where each turn provides new evidence that can reduce cognitive uncertainty about the environment.
      • Rational action requires inferring unobserved quantities that control it and updating beliefs about them as evidence accumulates.

  • However, most evaluations only score the model’s final-turn answer in a single-turn format, leaving this process unexamined.

    • EN Highlights:
      • arXiv:2606.30850v1 Announce Type: new
      • Abstract: Large language models (LLMs) are typically deployed in multi-turn conversations, where each turn provides new evidence that should reduce epistemic un…
      • Acting rationally then requires inferring the unobserved quantities that govern it and updating beliefs about them as evidence accumulates
      • Yet most evaluations only score the model’s final-turn answer in a single-turn format, leaving this process unexamined

  • When Does Learning to Stop Help? A Cost-Aware Study of Early Exits in Reasoning Models

    • Publication Time: 2026-07-01 12:00 Beijing Time
    • Abstract: - arXiv:2606.30852v1 Announcement Type: New.
      • Abstract: Reasoning models spend different amounts of useful computation across instances, but it remains unclear when a learned stopping rule improves over simple confidence or convergence thresholds.
      • We study this question with LearnStop, a hidden-state-free checkpoint stopper for reasoning language models.
      • At fixed-budget checkpoints, LearnStop probes a short answer from the current reasoning prefix and predicts prefix correctness from online features such as answer confidence, entropy, prefix vote share, answer stability, and backtrack token density.
    • EN Highlights:
      • arXiv:2606.30852v1 Announce Type: new
      • Abstract: Reasoning models spend different amounts of useful computation across instances, but it remains unclear when a learned stopping rule improves over sim…
      • We study this question with LearnStop, a hidden-state-free checkpoint stopper for reasoning language models
      • At fixed budget checkpoints, LearnStop probes a short answer from the current reasoning prefix and predicts prefix correctness from online features such as answ…

  • Beyond expert users: agents should help users construct preferences, not just elicit them

    • Publication Time: 2026-07-01 12:00 Beijing Time
    • Abstract: - arXiv:2606.30863v1 Announcement Type: New.
      • Abstract: Agents often assume an expert user—one with clear preferences for what they want—and default to clarifying questions when tasks are underspecified.
      • We argue this assumption is unrealistic.
      • Users often lack the domain knowledge to have fully-specified preferences; if asked their preference on some feature, a user may be unable to answer if the agent does not help the user learn some of the domain knowledge necessary to form a preference on that feature (e.g., through examples or explanations).
    • EN Highlights:
      • arXiv:2606.30863v1 Announce Type: new

  • Abstract: Agents typically assume an expert user – one with well-formed preferences about what they want – and default to clarifying questions whenever the ta…

  • We argue this assumption is unrealistic

  • Users often lack the domain knowledge to have completely specified preferences; if asked about their preference on some feature, the user may be unable to answe…

  • Investigating Multi-Agent Deliberation in Law

    • Publication Time: 2026-07-01 12:00 Beijing Time
    • Abstract: - arXiv:2606.30906v1 Announcement Type: New.
      • Abstract: Artificial intelligence is increasingly applied to the field of law, and has the potential to increase access to justice.
      • One particular movement that is gaining traction is that of agentic AI, wherein AI agents, based on Large Language Models (LLMs) can take autonomous actions.
      • In particular, multi-agent approaches in the legal domain remain largely unexplored.
    • EN Key Points:
      • arXiv:2606.30906v1 Announce Type: new
      • Abstract: Artificial Intelligence is increasingly applied to the field of law, and has the potential to increase access to justice
      • One particular movement that is gaining traction is that of agentic AI, wherein AI agents, based on Large Language Models (LLMs) can take autonomous actions
      • In particular, multi-agent approaches in the legal domain remain largely unexplored

  • Why Solve It Twice? Hierarchical Accumulation of Skills for Transfer-Efficient ML Engineering

    • Publication Time: 2026-07-01 12:00 Beijing Time
    • Abstract: - arXiv:2606.30911v1 Announcement Type: New.
      • Abstract: Machine learning engineering agents waste compute rediscovering known techniques because every competition is a cold start.
      • We present HASTE, a hierarchical multi-agent system that organizes cross-competition knowledge into three scope tiers (global, domain, and competition-specific), each coupled to a matching agent level.
      • An orchestrator coordinates domain specialists and promotes learning between tiers via LLM-driven abstraction.
    • EN Key Points:
      • arXiv:2606.30911v1 Announce Type: new
      • Abstract: ML engineering agents waste compute rediscovering known techniques because every competition is a cold start
      • We present HASTE, a hierarchical multi-agent system that organizes cross-competition knowledge into three scope tiers (global, domain, and competition-specific)…
      • An orchestrator coordinates domain specialists and promotes learning between tiers via LLM-driven abstraction

  • RoPoLL: Robust Panel of LLM Judges

  • Publication Time: 2026-07-01 12:00 Beijing Time

    • Abstract: - arXiv:2606.30931v1 Announcement Type: new.
      • Abstract: The LLM Jury, a Panel of LLM Evaluators (PoLL) that reports consensus scores, has become a practical alternative to single-judge LLM evaluation, yet its statistical behavior remains poorly understood.
      • We formalize the LLM Jury under the Huber contamination model and show that PoLL incurs unbounded bias
      • under any positive contamination, regardless of jury size, whenever a single judge fails in a biased, LLM-typical way (mode collapse, sycophancy, safety refusal).
    • EN Highlights:
      • arXiv:2606.30931v1 Announce Type: new
      • Abstract: The LLM Jury, a Panel of LLM Evaluators (PoLL) reporting consensus scores, has become a practical alternative to single-judge LLM evaluation, yet its…
      • We formalize the LLM Jury under the Huber contamination model and show that PoLL incurs unbounded bias
      • under any positive contamination, regardless of jury size, whenever a single judge fails in a biased, LLM-typical way (mode collapse, sycophancy, safety refusal…

  • AgRefactor: Self-Evolving Agentic Workflow for HLS Compatibility and Performance

    • Publication Time: 2026-07-01 12:00 Beijing Time
    • Abstract: - arXiv:2606.30949v1 Announcement Type: new.
      • Abstract: High-Level Synthesis (HLS) provides a fast path from concept to chip, but converting real-world software into synthesizable HLS code remains challenging due to restrictive language support and the gap between software and hardware programming practices.
      • Existing automated and LLM-based refactoring approaches partially address this problem, but they often lack flexibility, struggle to scale, and incur high computational costs.
      • We introduce AgRefactor, an LLM-based multi-agent workflow for refactoring software into HLS-compatible programs.
    • EN Highlights:
      • arXiv:2606.30949v1 Announce Type: new
      • Abstract: High-Level Synthesis (HLS) provides a fast path from concepts to silicon, but converting real-world software into synthesizable HLS code remains chall…
      • Existing automated and LLM-based refactoring approaches partially address this problem, yet they often lack flexibility, struggle to scale, and incur high compu…
      • We introduce AgRefactor, an LLM-based multi-agent workflow for refactoring software into HLS-compatible programs

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

  • A Single Rewrite Suffices: Empirical Lessons from Production Skill Description Optimization

    • Publication Time: 2026-07-01 12:00 Beijing Time
    • Abstract: - arXiv:2606.30775v1 Announcement Type: new.
      • Abstract: Enterprise AI agents route user queries to specialized skills by matching the query against natural language skill descriptions.
      • When two skills share overlapping descriptions, the routing LLM misroutes the query, a failure we term skill conflict.

  • As agents expand to dozens of skills, manually adjusting descriptions to maintain routing accuracy becomes a significant engineering bottleneck.

    • EN 要点:
      • arXiv:2606.30775v1 Announce Type: new
      • Abstract: Enterprise AI agents route user queries to specialized skills by matching queries against natural language skill descriptions
      • When two skills share overlapping descriptions, the routing LLM misroutes queries, a failure we term skill collision
      • As agents scale to dozens of skills, manually tuning descriptions to maintain routing accuracy becomes a significant engineering bottleneck

  • Indi-RomCoM: Code-Mixed Benchmark for Evaluating LLMs on Romanized Indic-English Instructions

    • Published: 2026-07-01 12:00 Beijing Time
    • Abstract:- arXiv:2606.30790v1 Announce Type: new.
      • Abstract: Romanized Code Mixing (RCM), where bilingual speakers fluently blend local languages with English in Roman script, has become a primary form of communication in multilingual communities.
      • While Large Language Models (LLMs) perform strongly in monolingual and native-script benchmarks, their ability to follow instructions and reason over RCM-based content largely remains unexplored.
      • To this end, we introduce the Indi-RomCoM benchmark to facilitate systematic evaluation of Indic Romanized Code-Mixed instructions.
    • EN 要点:
      • arXiv:2606.30790v1 Announce Type: new
      • Abstract: Romanized Code Mixing (RCM), where bilingual speakers fluidly blend local languages with English in Roman script, has emerged as the dominant form of…
      • While Large Language Models (LLMs) perform strongly on monolingual and native-script benchmarks, their ability to follow instructions and reason over RCM-based…
      • To this end, we introduce the Indi-RomCoM benchmark for facilitating systematic evaluation on Indic Romanized Code-Mixed instructions

  • Using AI Agents to Automate Black-Box Audits of Personalization Algorithms at Scale

    • Published: 2026-07-01 12:00 Beijing Time
    • Abstract:- arXiv:2606.30801v1 Announce Type: new.
      • Abstract: Personalization algorithms determine the content users encounter on online platforms.
      • Auditing these systems is difficult because independent auditors only have black-box access to the algorithms, and personalization depends on user attributes, behavior, and evolving interaction history.
      • Existing audit methods face trade-offs: studies with real users capture authentic behavior but are costly and difficult to control, while puppet audits are easier to scale but often rely on scripted behaviors that limit realism.
    • EN 要点:
      • arXiv:2606.30801v1 Announce Type: new
      • Abstract: Personalization algorithms determine what content users encounter on online platforms

  • Auditing these systems is difficult because independent auditors have only black-box access to the algorithms, while personalization depends on users’ attribute…

  • Existing auditing methods face a tradeoff: studies with real users capture realistic behavior but are costly and hard to control, whereas sock-puppet audits sca…

  • When Calibration Rankings Reverse: Accuracy-Controlled Evaluation for Fair Comparison of LLMs

    • Published: 2026-07-01 12:00 Beijing Time
    • Abstract: - arXiv:2606.30814v1 Announce Type: new.
      • Abstract: Calibration evaluates whether a model’s confidence aligns with its empirical accuracy.
      • Existing studies often use global calibration metrics such as Expected Calibration Error and Brier score to compare the calibration of different large language models.
      • We first show, both theoretically and empirically, that such comparisons are confounded by differences in model accuracy.
    • EN Highlights:
      • arXiv:2606.30814v1 Announce Type: new
      • Abstract: Calibration evaluates whether a model confidence aligns with its empirical accuracy
      • Existing studies often compare the calibration of different large language models using global calibration metrics such as Expected Calibration Error and Brier…
      • We begin by showing, both theoretically and empirically, that such comparisons are confounded by differences in model accuracy

  • When transformers learn “impossible” languages, what do they learn?

    • Published: 2026-07-01 12:00 Beijing Time
    • Abstract: - arXiv:2606.30815v1 Announce Type: new.
      • Abstract: Recent research shows that Transformer language models exhibit a bias towards human languages over non-natural (“impossible”) languages considered unattainable by humans.
      • However, this literature is largely based on differences in sample efficiency and test-set perplexity, rather than on direct evaluations of linguistic capabilities that could reasonably explain ungrammaticalities in human languages.
      • We evaluate two theoretically driven hypotheses for this connection: the impossibility arises from defects in either grammatical sensitivity or generation.
    • EN Highlights:
      • arXiv:2606.30815v1 Announce Type: new
      • Abstract: Recent work suggests that transformer language models show a bias towards human languages over unnatural (“impossible”) languages argued to be unacqui…
      • However, this literature has largely based these claims on differences in sample efficiency and test-set perplexity, rather than on direct evaluations of the li…

  • We evaluate two theoretically motivated linking hypotheses: impossibility arising from deficiencies in grammatical sensitivity or generative production

  • Test-Time Verification for Text-to-SQL via Outcome Reward Models

    • Release Time: 2026-07-01 12:00 Beijing Time
    • Abstract: - arXiv:2606.30851v1 Announce Type: new.
      • Abstract: Improving the reliability of large language models (LLMs) at inference time is a central challenge in structured reasoning tasks such as Text-to-SQL.
      • Common test-time inference strategies, including Best-of-N sampling and Majority Voting, rely on heuristic signals such as execution success or output frequency, which provide limited semantic differentiation among candidate outputs.
      • In this work, we study Outcome Reward Models (ORMs) as learned semantic scoring functions for test-time verification in Text-to-SQL.
    • EN Key Points:
      • arXiv:2606.30851v1 Announce Type: new
      • Abstract: Improving the reliability of large language models (LLMs) at inference time is a central challenge in structured reasoning tasks such as Text-to-SQL
      • Common test-time inference strategies, including Best-of-N sampling and Majority Voting, rely on heuristic signals such as execution success or output frequency…
      • In this work, we study Outcome Reward Models (ORMs) as learned semantic scoring functions for test-time verification in Text-to-SQL

  • Multilingual Polarization Detection Using Transformer-Based Models with Class Weighting and Threshold Tuning

    • Release Time: 2026-07-01 12:00 Beijing Time
    • Abstract: - arXiv:2606.30857v1 Announce Type: new.
      • Abstract: This paper describes our submission to SemEval-2026 Task 9 on detecting multilingual, multicultural, and multievent online polarization.
      • We address all three subtasks: binary polarization detection, polarization type classification, and manifestation identification for English and Swahili.
      • Our method leverages transformer-based models (RoBERTa-base for English, AfroXLMR-base for Swahili) with class-weighted loss functions to address severe label imbalance and uses per-label threshold tuning to optimize multi-label classification.
    • EN Key Points:
      • arXiv:2606.30857v1 Announce Type: new
      • Abstract: This paper describes our submission to SemEval-2026 Task 9 on detecting multilingual, multicultural, and multievent online polarization
      • We address all three subtasks: binary polarization detection, polarization type classification, and manifestation identification for English and Swahili
      • Our approach leverages transformer-based models (RoBERTa-base for English, AfroXLMR-base for Swahili) with class-weighted loss functions to address severe label…

  • Training Therapeutic Judges and Multi-Agent Systems for Human-Aligned Mental Health Support

    • Published: 2026-07-01 12:00 Beijing Time
    • Abstract: - arXiv:2606.30887v1 Announce Type: new.
      • Abstract: Large language models show promise for mental health support, yet therapeutic quality improves only when evaluation functions as an actionable control signal rather than a passive metric.
      • We introduce a framework that formulates therapeutic response generation as a decision-refinement problem driven by multi-dimensional, human-aligned evaluation.
      • In Stage I, we introduce TheraJudge, an open-source therapeutic evaluator trained via preference-based optimization on human-annotated data to produce reliable judgments across 7 psychological dimensions.
    • EN Highlights:
      • arXiv:2606.30887v1 Announce Type: new
      • Abstract: Large language models show promise for mental health support, yet therapeutic quality improves only when evaluation functions as an actionable control…
      • We introduce a framework that formulates therapeutic response generation as a decision-refinement problem driven by multi-dimensional, human-aligned evaluation
      • In Stage I, we introduce TheraJudge, an open-source therapeutic evaluator trained via preference-based optimization on human-annotated data to produce reliable…

  • Beyond Clean Text: Evaluating Encoder and Decoder Robustness for Bangla Event Detection in Noisy Text

    • Published: 2026-07-01 12:00 Beijing Time
    • Abstract: - arXiv:2606.30914v1 Announce Type: new.
      • Abstract: Event detection (ED) systems are typically evaluated on clean, well-curated text, while their robustness to real-world noise remains largely unexplored, especially for low-resource languages like Bangla.
      • We introduce a generalized Bangla news event ontology and a benchmark comprising 9,979 annotated sentences across 40 event subtypes, spanning clean news text, real-world Automatic Speech Recognition (ASR) transcripts, and misspelled text.
      • We systematically evaluate fine-tuned encoder-only models (BanglaBERT and XLM-R) alongside instruction-tuned decoder-only large language models (Llama 3 and Gemma 3).
    • EN Highlights:
      • arXiv:2606.30914v1 Announce Type: new
      • Abstract: Event detection (ED) systems are typically evaluated on clean, curated text, leaving their robustness to real-world noise largely unexplored, particul…
      • We introduce a generalized Bangla news event ontology and a benchmark comprising 9,979 annotated sentences across 40 event subtypes, spanning clean news text, r…
      • We systematically evaluate fine-tuned encoder-only models (BanglaBERT and XLM-R) alongside instruction-tuned decoder-only large language models (Llama 3 and Gem…

  • Bridging Scientific Heritage: An Arabic–Russian Parallel Corpus and LLM Benchmark for Sustainable Knowledge Transfer

    • Publication Time: 2026-07-01 12:00 Beijing Time
    • Summary: - arXiv:2606.30943v1 Announcement Type: New.
      • Abstract: Russian and Arabic are among the major languages of scientific communication.
      • Language barriers impede the exchange of research results between these communities, which affects international collaboration and the progress of research related to sustainable development.
      • We present a benchmark for Arabic-Russian scientific translation.
    • EN Highlights:
      • arXiv:2606.30943v1 Announce Type: new
      • Abstract: Russian and Arabic are among the major languages of scientific communication
      • Language barriers impede the exchange of research results between these communities, which affects international collaboration and the progress of sustainabilit…
      • We present a benchmark for Arabic–Russian scientific translation

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

  • Joint discovery of governing partial differential equations from multi-source datasets by competitive optimization

    • Publication Time: 2026-07-01 12:00 Beijing Time
    • Summary: - arXiv:2606.30699v1 Announcement Type: New.
      • Abstract: Discovering governing equations directly from observational data is a key step towards interpretable scientific machine learning.
      • Current data-driven approaches typically operate on a single dataset, inherently limiting their performance when faced with restricted observations.
      • In practice, multiple datasets are often available for the same physical system, distinguished only by distinct initial conditions or boundary configurations.
    • EN Highlights:
      • arXiv:2606.30699v1 Announce Type: new
      • Abstract: Discovering governing equations directly from observational data is a key step towards interpretable scientific machine learning
      • Current data-driven approaches typically operate on a single dataset, inherently limiting their performance when faced with restricted observations
      • In practice, multiple datasets are often available for the same physical system, distinguished only by distinct initial conditions or boundary configurations

  • Accelerometry-Derived Digital Biomarkers for Cardiometabolic Risk: A Population-Representative Tabular Benchmark with Uncertainty Quantification

    • Publication Time: 2026-07-01 12:00 Beijing Time
    • Summary: - arXiv:2606.30702v1 Announcement Type: New.
      • Abstract: Structured tabular data is predominant in clinical medicine, but existing benchmarks fail to reflect real-world characteristics such as complex survey sampling, population oversampling, and subgroup fairness.

  • We introduce the NHANES Accelerometry Cardiometabolic Benchmark, derived from NHANES 2003-2006, comprising 1,381 adults with hip-worn accelerometry, fasting lab biomarkers, dietary intake, and anthropometric data.

  • We evaluate three tabular learning methods (ridge regression, XGBoost, and the foundation model TabPFN v2) to predict glycated hemoglobin (HbA1c), fasting triglycerides, and C-reactive protein (CRP) based on activity phenotypes and lifestyle covariates.

    • EN Highlights:
      • arXiv:2606.30702v1 Announce Type: new
      • Abstract: Structured tabular data dominates clinical medicine, yet existing benchmarks fail to reflect real-world properties like complex survey sampling, demog…
      • We introduce the NHANES Accelerometry Cardiometabolic Benchmark, derived from NHANES 2003-2006, comprising 1,381 adults with hip-worn accelerometry, fasting lab…
      • We evaluate three tabular learning methods – ridge regression, XGBoost, and the foundation model TabPFN v2 – to predict glycated haemoglobin (HbA1c), fasting…

  • From Search to Synthesis: Training LLMs as Zero-Shot Workflow Generators

    • Publication Time: 2026-07-01 12:00 Beijing Time
    • Abstract: - arXiv:2606.30704v1 Announcement Type: new.
      • Abstract: Large Language Models (LLMs) excel in a wide range of tasks, but their instance-specific solutions often lack the structural consistency required for reliable deployment.
      • Workflows that encode recurring algorithmic patterns at the task level provide a principled framework, offering robustness across instance variations, interpretable debugging traces, and reusability across problem instances.
      • However, manually designing such workflows requires significant expertise and effort, limiting their broader application.
    • EN Highlights:
      • arXiv:2606.30704v1 Announce Type: new
      • Abstract: Large language models (LLMs) excel across a wide range of tasks, yet their instance-specific solutions often lack the structural consistency needed fo…
      • Workflows that encode recurring algorithmic patterns at the task level provide a principled framework, offering robustness across instance variations, interpret…
      • However, manually designing such workflows requires significant expertise and effort, limiting their broader application

  • Why Do Few-Step Text Latents Fail When Image Latents Work? Non-Commitment at Sharp Categorical Readouts

    • Publication Time: 2026-07-01 12:00 Beijing Time
    • Abstract: - arXiv:2606.30705v1 Announcement Type: new.
      • Abstract: Deterministic few-step generation is successful on continuous image latents but collapses into incoherent text for continuous text latents. We show that the reason is geometric, rather than a training or scaling flaw: smooth, regularity-limited deterministic maps cannot resolve discrete branching choices before a sharp categorical readout. Therefore, few-step failures are determined by decoder sharpness, not transport accuracy.

  • In the overlapping regime of real text autoencoders, we prove (Theorem 3) that the posterior-mean terminal step flips tokens at the rate of the latent mass in a $O(s(t))$ tube around the decision boundary.

  • Two diagnostics, DABI (readout sharpness) and CCI (categorical commitment), measured on published checkpoints show that four independently built continuous-text decoders amplify boundary-aligned perturbations far beyond standard-matched isotropic perturbations (DABI from $5\times10^{2}$ to $>10^{5}$), while image decoders have a DABI of $\approx 1$.

    • EN Key Points:
      • arXiv:2606.30705v1 Announce Type: new
      • Abstract: Deterministic few-step generation succeeds on continuous image latents but collapses to incoherent text on continuous text latents, and we show the ca…
      • In the overlapping regime of real text autoencoders, we prove (Theorem 3) that the posterior-mean terminal step flips tokens at the rate of the latent mass in a…
      • Two diagnostics, DABI (readout sharpness) and CCI (categorical commitment), measured on published checkpoints show that four independently built continuous-text…

  • Hierarchical Global Attention (HGA)

    • Release Time: 2026-07-01 12:00 Beijing Time
    • Abstract: - arXiv:2606.30709v1 Announce Type: new.
      • Abstract: Hierarchical Global Attention (HGA) is a drop-in replacement for dense causal attention in pretrained long-context transformers.
      • HGA preserves the original checkpoint parameters: the pretrained $W_Q$, $W_K$, $W_V$, and $W_O$ projections remain unchanged, no calibration parameters are introduced, and no retraining is required.
      • Applied to Qwen3-30B-A3B-Instruct-2507-FP8 on a single RTX~5090 (32GB), the patched model runs out-of-the-box in a 64K token context, where token-level K/V storage is not feasible on this hardware.
    • EN Key Points:
      • arXiv:2606.30709v1 Announce Type: new
      • Abstract: Hierarchical Global Attention (HGA) is a drop-in replacement for dense causal attention in pretrained long-context transformers
      • HGA preserves the original checkpoint parameters: the pretrained $W_Q$, $W_K$, $W_V$, and $W_O$ projections remain unchanged, no calibration parameters are intr…
      • Applied to Qwen3-30B-A3B-Instruct-2507-FP8 on a single RTX~5090 (32GB), the patched model runs out of the box at a 64K-token context, where token-level K/V stor…

  • ReactionAtlas: Ab origine exploration of chemical reaction networks with machine learning

    • Release Time: 2026-07-01 12:00 Beijing Time
    • Abstract: - arXiv:2606.30778v1 Announce Type: new.
      • Abstract: Mapping chemical reaction networks, the graphs of minima and transition states (TSs) and the elementary reactions connecting them, is the natural language of chemistry, from catalysis to combustion to the origin of life.

  • Constructing such a reaction network for a given chemical reaction is impractical: it requires finding and characterizing tens of thousands of TS, while traditional methods like Density Functional Theory (DFT) are often prohibitively slow and require reactants and products as input.

  • We introduce ReactionAtlas, which builds a reaction network $\textit{ab origine}$ from a small number of seed molecules and without hand-crafted rules.

    • EN Highlights:
      • arXiv:2606.30778v1 Announce Type: new
      • Abstract: Mapping a chemical reaction network, the graph of minima and transition states (TS) and the elementary reactions connecting them, is the natural langu…
      • Constructing such a reaction network for a given chemistry has been impractical: it requires finding and characterizing tens of thousands of TS, a task for whic…
      • We introduce ReactionAtlas, which builds a reaction network $\textit{ab origine}$ from a handful of seed molecules and without hand-crafted rules

  • Revocable Learned State via Process Sidecars

    • Publication Time: 2026-07-01 12:00 Beijing Time
    • Abstract: - arXiv:2606.30788v1 Announcement Type: New.
      • Abstract: Language models are often adapted in stages: a public skill phase, a private memory phase, and a later safety phase that learns to refuse outputs related to remembered entities.
      • Revoking memory after the safety phase is not the same problem as subtracting the memory update: the later safety optimizer has already transported the memory direction.
      • We introduce process sidecars, a two-coefficient edit family $\hat{\theta}(\lambda,\gamma)=\theta_{\mathrm{AMS}}-\lambda\Delta_{\mathrm{M}}-\gamma\hat{R}{\mathrm{S}\leftarrow\mathrm{M}}$, where $\hat{R}{\mathrm{S}\leftarrow\mathrm{M}}=\hat{J}{\mathrm{S},\varepsilon}(\Delta{\mathrm{M}})-\Delta_{\mathrm{M}}$, and where $\hat{J}_{\mathrm{S},\varepsilon}$ is the central secant safety training process through a realized future AdamW.
    • EN Highlights:
      • arXiv:2606.30788v1 Announce Type: new
      • Abstract: Language models are often adapted in stages: a public skill phase, a private memory phase, and a later safety phase that learns to refuse outputs tied…
      • Revoking the memory after the safety phase is not the same problem as subtracting the memory update: the later safety optimizer has transported the memory direc…
      • We introduce process sidecars, a two-coefficient edit family $\hat{\theta}(\lambda,\gamma)=\theta_{\mathrm{AMS}}-\lambda\Delta_{\mathrm{M}}-\gamma\hat{R}_{\math…

  • Predictable GRPO: A Closed-Form Model of Training Dynamics

    • Publication Time: 2026-07-01 12:00 Beijing Time

  • Abstract:- arXiv:2606.30789v1 Announce Type: New.

    • Abstract: Group Relative Policy Optimization (GRPO) has become a standard tool for improving the reasoning ability of large language models, yet its training dynamics remain empirically described: reward trajectories conform to low-parameter functional forms whose constants lack mechanical meaning, and hyperparameter selection remains a matter of trial and error.
    • We develop a first-principles reduced-order model of these dynamics.
    • The reduction has three consequences.
    • EN Key Points:
      • arXiv:2606.30789v1 Announce Type: new
      • Abstract: Group Relative Policy Optimization (GRPO) has become a standard tool for improving the reasoning ability of large language models, yet its training dy…
      • We develop a first-principles reduced-order model of these dynamics
      • The reduction has three consequences

  • Gradient Smoothing: Coupling Layer-wise Updates for Improved Optimization

    • Publication Time: 2026-07-01 12:00 Beijing Time
    • Abstract:- arXiv:2606.30813v1 Announce Type: New.
      • Abstract: Deep neural networks with repeated architectural blocks (e.g., Transformers) often exhibit structured relationships across layers that emerge during training.
      • Motivated by this observation, we introduce \emph{Depth-wise Gradient Augmentation}, a general optimization paradigm in which the update applied to each layer is obtained by transforming a collection of block-wise optimizer updates along the depth dimension.
      • Within this framework, we study \emph{Gradient Smoothing}, a family of depth-wise smoothing methods, and instantiate it with a simple local \emph{Window Smoothing} operator.
    • EN Key Points:
      • arXiv:2606.30813v1 Announce Type: new
      • Abstract: Deep neural networks with repeated architectural blocks, such as transformers, often exhibit structured relationships across layers that emerge during…
      • Motivated by this observation, we introduce \emph{Depth-wise Gradient Augmentation}, a general optimization paradigm in which the update applied to each layer i…
      • Within this framework, we study \emph{Gradient Smoothing}, a family of depth-wise smoothing methods, and instantiate it with a simple local \emph{Window Smoothi…

  • Mind the Residual Gap: Probabilistic Downscaling under Real-World Bias

    • Publication Time: 2026-07-01 12:00 Beijing Time
    • Abstract:- arXiv:2606.30821v1 Announce Type: New.
      • Abstract: Probabilistic downscaling, the task of modeling the conditional distribution of a high-resolution field given coarse inputs, is a core challenge in atmospheric science, climate modeling, and other multi-scale physical systems.
      • A widely used paradigm decomposes the problem into a deterministic mean predictor followed by a stochastic residual generator.
      • While effective in idealized settings, this mean-residual approach frequently yields biased and under-dispersive ensembles in real-world applications.
    • EN Key Points:
      • arXiv:2606.30821v1 Announce Type: new

  • Abstract: Probabilistic downscaling is the task of modeling the conditional distribution of high-resolution fields given coarse inputs, and is a central challen…

  • A widely used paradigm decomposes the problem into a deterministic mean predictor followed by a stochastic residual generator

  • While effective in idealized settings, this mean–residual approach frequently produces biased and under-dispersive ensembles in real-world applications