2026-07-03 AI Daily | From Code Generation to Agent Orchestration: Testing and Auditing Become the New Moat Link to heading

Today’s main theme shifts to the reliability of AI in production: Agents are now competing not just on capabilities, but on their routing, validation, auditing, and testing systems. The lifting of restrictions on Claude Fable 5 and the strong agentic capabilities of Sonnet 5 are pushing developers from writing code to orchestrating workflows. Furthermore, multi-lingual and real-world noise evaluations remind the industry that clean English benchmarks are no longer sufficient to measure global adoption.

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

The most important topic to follow today is “the reliability of agents moving into production.” Skill description optimization, black-box auditing algorithms, multi-agent systems for mental health, and test-time validation for Text-to-SQL all point to the same issue: agents don’t lack capabilities, but rather controllable routing, verifiable outputs, and actionable evaluation signals. This is a must-read for engineering teams.

The second main theme is multi-lingual and real-world noise scenarios. Indi-RomCoM, Bengali event detection, Arabic-Russian scientific corpora, and multi-lingual polarization detection all remind us that the global evaluation of LLMs cannot remain stuck on clean English benchmarks.

Additionally, topics like calibrated fair comparison, mechanistic interpretability representation layers, and the unified perspective of GRPO/DAPO are worth following for research teams, as they help re-examine “what models have actually learned and whether evaluations are comparable.”

🌐 AI Hotspots on X Link to heading

Topic 1: Z.ai Launches ZCode for GLM-5.2 AI Model Link to heading

• Category: AI · News
• Overview: Trending Time: 23 hours ago, Related Posts: 65
• What it is: Z.ai has released ZCode for the GLM-5.2 AI model, aiming to enhance code generation and development assistance capabilities.
• Why it’s important: This indicates that Chinese AI companies are continuing to accelerate their development of large model programming capabilities and developer toolchains. These advancements could impact code intelligence assistants, enterprise R&D efficiency, and the competitive landscape of open-source vs. closed-source models.
• Discussion Summary: Discussions on X are mainly focused on ZCode’s actual coding capabilities compared to tools like GitHub Copilot and Cursor, whether GLM-5.2’s model performance is competitive, and that its ecosystem, availability, and acceptance by international developers still need to be validated.

Topic 2: Notion Engineer Shares Techniques to Understand AI-Generated Code Link to heading

• Category: AI · News
• Overview: Trending Time: 17 hours ago, Related Posts: 798
• What it is: A Notion engineer shared a set of practical methods for understanding and auditing AI-generated code, drawing attention from the developer community.
• Why it’s important: As AI programming tools are more widely used in production environments, how to quickly understand, validate, and maintain model-generated code has become a critical issue for software engineering reliability and team collaboration.
• Discussion Summary: Discussions on X primarily focus on whether AI-generated code will increase development efficiency, whether engineers need new code review skills, and whether over-reliance on AI will introduce risks related to maintainability, security, and accountability.

Topic 3: U.S. Lifts Export Controls on Anthropic’s Claude Fable 5 Model Link to heading

• Category: AI · News
• Overview: Trending Time: 2 days ago, Related Posts: 155,000
• What it is: The U.S. has reportedly lifted export controls on Anthropic’s Claude Fable 5 model, making it available again to a wider range of overseas users or customers.
• Why it’s important: This shows that frontier AI models are being incorporated into national security and export control frameworks similar to those for advanced chips. The degree of model openness will affect the global commercialization, technology dissemination, international competition, and security governance of AI companies.
• Discussion Summary: Discussions on X center on whether lifting the ban represents a relaxation of regulations. Supporters believe it will help U.S. AI companies maintain their global market and technological lead, while critics worry that high-capability models could be jailbroken and used for high-risk applications like cybersecurity or biosecurity. Another key point of discussion is whether frontier models will move towards a tiered access model, with limited public versions and more powerful versions for trusted users or governments.

Topic 4: Anthropic Launches Claude Sonnet 5 with Strong Agentic Skills Link to heading

• Category: AI · News
• Overview: Trending Time: 2 days ago, Related Posts: 57,000
• What it is: Anthropic has released Claude Sonnet 5, featuring stronger agentic capabilities, and it has shown outstanding performance on benchmarks for coding and terminal tasks like SWE-bench Pro and Terminal-Bench.
• Why it’s important: This indicates that the competition among frontier models is shifting from mere parameters and conversational abilities to “agentic” capabilities for executing complex tasks, which will particularly impact software development, knowledge work, and automated workflows.
• Discussion Summary: Discussions on X focus on whether Sonnet 5 is approaching or surpassing the more expensive Opus 4.8, whether its coding and terminal operation capabilities are sufficient for real-world production scenarios, and whether the view that “toolchains and evaluation frameworks are more critical than the model itself” is becoming a new consensus.

Topic 5: Anthropic’s Claude Fable 5 Returns with Top Coding Prowess Link to heading

• Category: AI · News
• Overview: Trending as of 21 hours ago, Number of related posts: 17,000
• What happened: Anthropic’s Claude Fable 5 has attracted attention after its reopening, reportedly demonstrating leading capabilities in code generation and software engineering tasks.
• Why it matters: Programming ability is a key indicator for the commercial viability of large models. If its performance is validated, it will intensify competition between companies like Anthropic, OpenAI, and Google in the market for AI programming assistants and agent development tools.
• Discussion summary: The discussion on X focuses on whether Claude Fable 5’s benchmark scores are reliable, whether its actual development experience is superior to existing models, and whether Anthropic can strike a balance between capability, security, speed, cost, and the developer ecosystem.

Summary of AI Public Opinion on X Today Link to heading

Today’s main narrative clearly revolves around “AI programming and agent capabilities”: Whether it’s Z.ai launching ZCode or Anthropic releasing/reopening Claude Sonnet 5 and Fable 5, the discussion points to the competition among large models shifting from chat capabilities to real-world software engineering, endpoint operations, and automated workflows. The broad consensus is that AI coding tools have entered production environments and are expected to boost development efficiency. However, beyond model capabilities, factors like code review, toolchains, evaluation frameworks, ecosystem, and maintainability are equally crucial. The main points of contention are whether the benchmark scores of these new models and tools can translate into real-world development experience, and whether leading models like Anthropic’s are truly pulling ahead or are still constrained by cost, speed, security, and ecosystem limitations. Potential risks are concentrated on two fronts: on one hand, the security, maintenance, and accountability issues arising from engineering teams over-relying on AI-generated code; on the other, the increased governance pressure from high-capability models being misused in sensitive areas like cybersecurity and biosecurity after the relaxation of export controls on frontier models.

💡 Influencer Insights Link to heading

Daily AI Industry Observations (Based on X Platform Influencer Tweets) Link to heading

1. Today’s Core Trend: Shifting from a “Model Arms Race” to “Agent Orchestration and Workflow Integration” Link to heading

The focus of today’s discussion has clearly shifted from the release of single models to how powerful models (especially Claude Fable 5) can be used as core schedulers to work in concert with other tools and sub-models to build complex automated workflows.

• Claude Fable 5/The Mythos 5 Restrictions Lifted and Its Return: This was the biggest industry event of the day. The U.S. Department of Commerce lifted export controls on these two flagship models, marking a temporary end to the recent AI regulatory turmoil surrounding national security. @AnthropicAI announced it will gradually restore access. Blogger @dotey analyzed the impact, noting that OpenAI’s GPT-5.6 had also faced similar restrictions, and argued that a licensing system lacking clear safety standards “turns every model release into an impromptu negotiation.”

• Sonnet 5 Release, Focusing on “Democratizing Agent Capabilities”: @AnthropicAI released Claude Sonnet 5 with clear positioning: offering agent capabilities close to the flagship model Opus 4.8, but at only 40% of the price. @dotey’s review concluded that on complex tasks, Sonnet 5 “can complete them in one go, unlike before when it would get stuck halfway,” which signals that top-tier agent capabilities are being cascaded down from expensive flagship models to mainstream ones.

• A “Claude Code Moment” for Scientific Research: Anthropic released Claude Science, an AI workbench for researchers. @dotey analyzed its strategy in detail: Don’t change the model, change the workflow. It connects over 60 scientific databases through a primary AI agent, integrating the research process into a single environment, attempting to become the “operating layer for the field of scientific research.” This is in stark contrast to OpenAI’s approach of training domain-specific models (GPT-Rosalind).

• Growing Interest and Real-world Tests for On-Device Models: On-device models continue to be a hot topic for tech enthusiasts. @zhixianio thoroughly tested the local code generation capabilities of Gemma 4 12B Coder and conducted a comparative evaluation with Qwen3.6-35B-A3B MoE on complex tasks (such as generating a complete, playable Tetris game). The conclusion was that “a 12B model can’t sustain long, stateful, complex programs,” with the bottleneck being the base model’s capability rather than fine-tuning. At the same time, he was satisfied with the full-duplex audio-video capabilities of MiniCPM-o 4.5, stating, “For a 9B model to achieve this effect, Good job.”

2. Unique Perspectives and Industry Outlook Link to heading

• Identity Shift from “Codex User” to “Fable Coordinator”: @Pluvio9yte and @diegocabezas01 shared a representative new workflow: using Claude Fable 5 as the “main architect/coordinator” for planning and task decomposition; using Opus or Codex as “sub-agents” for complex reasoning; and using Sonnet as a “fast worker” for repetitive tasks. This suggests a future where developers may engage in “Agent orchestration” rather than direct “programming.”

• Model Commoditization and “Token” Value Differentiation: @lijigang proposed that “Tokens are the calories of thought” and “Tokens are heterogeneous,” pointing out that the value of tokens varies across different models. This disparity could lead to a “tax-like” situation, rather than tokens becoming a fully homogenized commodity like electricity. This echoes the community’s heated debate sparked by Fable 5’s high API pricing ($50/million output tokens).

• The Future of Software Engineering: The Paradox of Trust and Review: @dotey forwarded @laike9m_’s view that in the era of Agentic Coding, automated testing has become critically important. With the surge in AI-generated code, manual code review is impractical. Behaviors that can be automatically verified should be handled by tests, allowing human effort to concentrate on building complex integration testing environments.

• The Competition Between CLI, API, and MCP: @vista8 summarized @RhysSullivan’s perspective on Agent tool invocation methods, arguing that the CLI is the most user-friendly option today, but it is a dead end in the long run. MCP (Model Context Protocol) is the only solution that considers “Human-in-the-loop” at the protocol level, while APIs, with their rich metadata, are suitable for 90% of scenarios.

• Redefining and Questioning “Open Source”: @ruanyf cited the Anthropic founder’s viewpoint that current AI models can only be described as “open-weight,” not truly “open-source,” because one cannot see their internal workings or participate in their development. This breaks the collaborative model of traditional open-source software.

• Deep Insights into AI’s Limitations: @vista8 cited an interview with Grant Sanderson of 3Blue1Brown, who pointed out that AI’s most underrated advantage is parallelization, not intelligence. It excels at connecting existing knowledge across different domains but is currently unable to create entirely new frameworks of thought. @nishuang, from a design perspective, distinguished between “Gamification,” which leverages dopamine, and “Game-like design,” which stimulates endorphins—a highly insightful distinction for AI product design.

3. Noteworthy Tools & Resources Link to heading

• Claude Science: An AI workbench for researchers, integrating over 60 scientific databases, local computation support (SSH connection to clusters), and reproducible chart generation. Available on MacOS/Linux for Pro users and above. ( @dotey)

• Raycast Glaze: An AI tool for developing desktop software with a single sentence, created by the team behind the well-known productivity tool Raycast. It is now open to everyone and embodies the “Vibe Coding” philosophy on the desktop. ( @vista8)

• DevSpace: An open-source project that connects via MCP, allowing the ChatGPT web version to directly read, write, and run local project code. This effectively gives ChatGPT the capabilities of Codex, with a separate usage quota. ( @gefei55)

• Open-Source Video Creation Skills Library: @Pluvio9yte has open-sourced a skills repository for replicating HyperFrames/Remotion-style videos, enabling users without editing skills to generate high-quality videos through AI.

• Front-End Development Skill Set: @vista8 recommended three front-end development Skills: animation-vocabulary (motion effects dictionary), emil-design-eng (polishing UI motion effects), and review-animations (reviewing animation issues), effectively solving the pain point of “being unable to create professional motion effects due to a lack of professional terminology.”

• A New Paradigm for Code Review: The view of blogger @laike9m_ deserves everyone’s attention—in the AI era, test cases are the new moat. Investing in automated testing, especially for complex integration and environments, is key to ensuring software quality. (Cited by both @ruanyf and @dotey)

• GPU Monitoring Solution for 10K+ Card Clusters: ARGUS: A cluster monitoring solution open-sourced by the Tencent team, specifically designed to address the pain point of fault localization caused by network communication issues in large-scale training. It is highly valuable for AI Infra teams. ( @vista8)

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

Time window: Last 3 days; 22 sources covered; 30 updates in total

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

• Constructive Alignment: Governing Preference Dynamics in Human-AI Interaction

  • Publication Time: 2026-07-02 12:00 Beijing Time
  • Abstract:- arXiv:2607.00001v1 Announcement Type: new.
    • Abstract: Most AI alignment methods treat human preferences as fixed targets that need to be inferred and optimized.
    • This assumption conflicts with extensive empirical evidence showing that preferences are layered, dynamic, and constructed through interaction (especially with adaptive technologies).
    • As AI systems become more persistent, personalized, and integrated into society, they increasingly participate in shaping what people attend to, value, and endorse over time.
  • EN Highlights:
    • arXiv:2607.00001v1 Announce Type: new
    • Abstract: Most approaches to AI alignment treat human preferences as fixed targets to be inferred and optimized
    • This assumption conflicts with extensive empirical evidence showing that preferences are layered, dynamic, and constructed through interaction–particularly wit…
    • As AI systems become more persistent, personalized, and socially embedded, they increasingly participate in shaping what people attend to, value, and endorse ov…

• Bounded Morality: Defining the Space of Moral Computation

  • Publication Time: 2026-07-02 12:00 Beijing Time
  • Abstract:- arXiv:2607.00002v1 Announcement Type: new.
    • Abstract: Traditionally, moral cognition has been modeled as adherence to fixed ethical theories—deontology, consequentialism, virtue ethics—implemented as static rules or value functions.
    • We propose Bounded Morality, a formal framework for analyzing the computational demands of moral problems faced by finite agents.
    • Extending Herbert Simon’s concept of bounded rationality, we formalize moral situations along two orthogonal dimensions: moral breadth, the scope of entities treated as morally relevant, and moral depth, evaluating the inferential integration required for their interaction.
  • EN Highlights:
    • arXiv:2607.00002v1 Announce Type: new
    • Abstract: Moral cognition has traditionally been modeled as adherence to fixed ethical theories–deontology, consequentialism, virtue ethics–implemented as sta…
    • We propose Bounded Morality, a formal framework for analyzing the computational demands of moral problems faced by finite agents
    • Extending Herbert Simon’s notion of bounded rationality, we formalize moral situations along two orthogonal dimensions: moral breadth, the scope of entities tre…

• The MMM Data Model – A Normative Specification for Knowledge Interoperability in a Decentralisable Knowledge Commons

  • Publication Time: 2026-07-02 12:00 Beijing Time
  • Abstract:- arXiv:2607.00032v1 Announcement Type: new.

• Abstract: Many information systems are built around documents: self-contained units optimized for print production and linear reading.

• While effective for large-scale dissemination, document-centric organization limits how knowledge can be structured, updated, shared, and reused.

• Formal methods address some of these limitations, but struggle to achieve widespread contribution and adoption because formal structure is prioritized over other system attributes (such as human usability and scope).

• EN Highlights:

  • arXiv:2607.00032v1 Announce Type: new
  • Abstract: Many information systems are built around documents: self-contained units optimised for print production and linear reading
  • While effective for large-scale dissemination, the document-centric organisation constrains how knowledge can be structured, updated, shared, and reused
  • Formal approaches address some of these limitations but struggle to achieve widespread contribution and adoption due to their prioritisation of formal structure…

• Making Failure Safe: A Constrained, Verifiable Agent Framework for Open-Web Data Collection

  • Published: 2026-07-02 12:00 Beijing Time
  • Abstract: - arXiv:2607.00035v1 Announce Type: new.
    • Abstract: LLMs and agents can generate web scrapers from natural language requirements, but direct generation remains unreliable due to dependency errors, broken selectors, schema mismatches, and heterogeneous page structures.
    • We propose a constrained, verifiable agent framework that shifts LLM output from free-form code to typed JSON collector configurations, combining a six-type collector taxonomy, template and utility function constraints, static Airflow DAG execution, rule-based quality checks, and structured feedback correction.
    • Experiments on 138 tasks show that the taxonomy supports description-based requirement typing, while confirming that stable instantiation requires completing source, field, and execution constraints beyond the initial description.
  • EN Highlights:
    • arXiv:2607.00035v1 Announce Type: new
    • Abstract: LLMs and agents can generate web scrapers from natural-language requirements, but direct generation remains unreliable because of dependency errors, b…
    • We propose a constrained, verifiable agent framework that shifts LLM output from free-form code to typed JSON collector configurations, combining a six-type col…
    • Experiments on 138 tasks show that the taxonomy supports description-based requirement typing, while confirming that stable instantiation requires completing so…

• Solution space path planning for supporting en-route air traffic control

  • Published: 2026-07-02 12:00 Beijing Time
  • Abstract: - arXiv:2607.00064v1 Announce Type: new.
    • Abstract: With technological advancements, many path planning algorithms have been proposed for air traffic management, but their operational application in tactical control remains limited, revealing a discrepancy between algorithmic design priorities and the needs of air traffic controllers.
    • This highlights the need for decision support solutions that are inherently interpretable, computationally efficient, and explicitly designed for human use.

• To address this design challenge, this study develops a conflict-free path-planning algorithm for en-route Air Traffic Control (ATC) designed to be compatible with two guiding considerations: (1) the interpretability and flexibility offered by a solution-space display, which motivates an algorithm that exposes all feasible safe actions and adapts to changing optimization objectives; and (2) the natural suitability of a decision logic controller for executing operational constraints, such as separation criteria, maneuverability limits, waypoint minimization, and route utility.

  • EN Highlights:
    • arXiv:2607.00064v1 Announce Type: new
    • Abstract: As technology advances, many path-planning algorithms have been proposed for Air Traffic Management, yet their operational adoption in tactical contro…
    • This underscores the need for decision-support solutions that are inherently interpretable, computationally efficient, and explicitly designed for human use
    • Focusing on this design challenge, this study develops a conflict-free path-planning algorithm for en-route Air Traffic Control (ATC) designed to be compatible…

• RareDxR1: Autonomous Medical Reasoning for Rare Disease Diagnosis Beyond Human Annotation

  • Release Time: 2026-07-02 12:00 Beijing Time
  • Abstract: - arXiv:2607.00147v1 Announce Type: new.
    • Abstract: Rare disease differential diagnosis is a critical yet arduous clinical task, requiring physicians to identify precise phenotypes from complex, unstructured patient symptoms and perform complex reasoning across a vast search space.
    • However, existing AI approaches typically rely on pipeline-based phenotype extraction or retrieval-augmented generation, which suffer from critical information loss due to predefined ontologies, retrieval bottlenecks, and a lack of diagnostic logic.
    • To address these challenges, we introduce RareDxR1, an end-to-end reasoning-centric large language model designed for open-domain rare disease diagnosis directly from unstructured clinical records.
  • EN Highlights:
    • arXiv:2607.00147v1 Announce Type: new
    • Abstract: Rare disease differential diagnosis is a critical yet arduous clinical task, requiring physicians to identify precise phenotypes from complex, unstruc…
    • However, existing AI approaches typically rely on pipeline-based phenotype extraction or retrieval-augmented generation, which suffer from critical information…
    • To address these challenges, we introduce RareDxR1, an end-to-end reasoning-centric large language model designed for open-domain rare disease diagnosis directl…

• A Contextual-Bandit Oversight Game with Two-Sided Informational Asymmetry

  • Release Time: 2026-07-02 12:00 Beijing Time
  • Abstract: - arXiv:2607.00155v1 Announce Type: new.
    • Abstract: We study the runtime human oversight of an AI agent when private information flows both ways: the human privately knows her reward function, while the AI privately knows the quality of the actions it proposes.
    • This asymmetry arises naturally when an autonomous robot or a software agent examines situations that its human supervisor cannot directly assess.

• Building on Cooperative Inverse Reinforcement Learning (CIRL) and the Oversight Game, we introduce a contextual-bandit team game with two-sided asymmetric information and a play/ask/trust/oversee interface.

  • EN Highlights:
    • arXiv:2607.00155v1 Announce Type: new
    • Abstract: We study runtime human oversight of an AI agent when private information runs in both directions: the human privately knows her reward function, while…
    • This is the kind of asymmetry that arises naturally when an autonomous robot or software agent has inspected a situation its human supervisor cannot directly as…
    • Building on Cooperative Inverse Reinforcement Learning (CIRL) and the Oversight Game, we introduce a contextual-bandit team game with two-sided asymmetric infor…

• Constructing Epistemic AI Literacy: Detecting Epistemic Aims and Processes in Student-AI Co-Programming

  • Published: 2026-07-02 12:00 Beijing Time
  • Abstract: - arXiv:2607.00211v1 Announce Type: new. -Abstract: Epistemic thinking plays a central role in students’ learning processes when applying generative artificial intelligence (GenAI), particularly in programming environments where learners must construct queries, evaluate and validate AI-generated outputs, and regulate problem-solving strategies.
    • This study introduces the conceptual framework of Epistemic AI Literacy (EAIL), reframing AI literacy as a process-oriented epistemic phenomenon that emerges through dynamic human-AI interactions across different domains.
    • Drawing on the AIR (epistemic aims, ideals and reliable epistemic processes) framework, this study examines how epistemic aims and epistemic processes are enacted in GenAI-supported co-programming activities, and explores scalable methods for operationalizing these constructs in interaction data.
  • EN Highlights:
    • arXiv:2607.00211v1 Announce Type: new
    • Abstract: Epistemic thinking plays a central role in students’ learning processes when applying generative artificial intelligence (GenAI), particularly in prog…
    • This study introduces the conceptual framework of Epistemic AI Literacy (EAIL), reframing AI literacy as a process-oriented epistemic phenomenon that emerges th…
    • Drawing on the AIR (epistemic aims, ideals and reliable epistemic processes) framework, this study examines how epistemic aims and epistemic processes are enact…

• From Signals to Structure: How Memory Architecture Drives Language Emergence in LLM Agents

  • Published: 2026-07-02 12:00 Beijing Time
  • Abstract: - arXiv:2607.00233v1 Announce Type: new.
    • Abstract: How can two agents invent a shared language from scratch?
    • In a Lewis signaling game, a sender and a receiver must coordinate on a code using only their interaction history.
    • We study with LLM agents five memory architectures across different channel configurations and find that memory architecture matters more than channel capacity.
  • EN Highlights:

• arXiv:2607.00233v1 Announce Type: new

  • Abstract: How do two agents invent a shared language from scratch
  • In a Lewis signaling game, a sender and receiver must coordinate on a code using only their interaction history
  • We study five memory architectures across varying channel configurations with LLM agents and find that memory architecture matters more than channel capacity

• Seed2.0 Model Card: Towards Intelligence Frontier for Real-World Complexity

  • Published:2026-07-02 12:00 Beijing Time
  • Abstract:- arXiv:2607.00248v1 Announce Type: new.
    • Abstract: We present Seed2.0, a model series that takes a meaningful step toward solving complex, real-world tasks.
    • Our approach begins with identifying users’ genuine needs and constructing a reliable, forward-looking evaluation system by selecting and abstracting benchmarks based on these needs and real, complex scenarios.
    • Guided by this evaluation system, Seed2.0 targets two persistent challenges, long-tail knowledge and complex instruction following, substantially improving the model’s reliability in complex, long-term tasks.
  • EN Highlights:
    • arXiv:2607.00248v1 Announce Type: new
    • Abstract: We present Seed2.0, a model series that takes a meaningful step toward solving complex, real-world tasks
    • Our approach begins with identifying users’ genuine needs and constructing a reliable, forward-looking evaluation system by selecting and abstracting benchmarks…
    • Guided by this evaluation system, Seed2.0 targets two persistent challenges, long-tail knowledge and complex instruction following, substantially improving the…

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

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

  • Published:2026-07-02 12:00 Beijing Time
  • Abstract:- 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.
  • EN Highlights:
    • 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

  • Release Time: 2026-07-02 12:00 Beijing Time
  • Summary: - arXiv:2606.30790v1 Announcement 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 communication in multilingual communities.
    • While Large Language Models (LLMs) perform strongly on monolingual and native-script benchmarks, their ability to follow instructions and reason over RCM-based content remains largely unexplored.
    • To this end, we introduce the Indi-RomCoM benchmark to facilitate systematic evaluation on Indic Romanized Code-Mixed instructions.
  • EN Highlights:
    • 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

  • Release Time: 2026-07-02 12:00 Beijing Time
  • Summary: - arXiv:2606.30801v1 Announcement 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, while personalization depends on users’ attributes, behaviors, and evolving interaction history.
    • Existing auditing methods face a trade-off: studies with real users capture authentic behavior but are costly and difficult to control, whereas sock-puppet audits are more scalable but often rely on scripted behaviors that limit realism.
  • EN Highlights:
    • 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

  • Release Time: 2026-07-02 12:00 Beijing Time
  • Summary: - 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.

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

  • Release Time: 2026-07-02 12:00 Beijing Time
  • Summary: - arXiv:2606.30815v1 Announce Type: new.
    • Abstract: Recent research indicates that Transformer language models exhibit a bias towards human languages over unnatural (“impossible”) languages argued to be unacquirable by humans.
    • However, this literature is largely based on differences in sample efficiency and test-set perplexity, rather than on direct evaluations of linguistic competence, which could reasonably explain non-demonstrability in human languages.
    • We evaluate two theoretically driven 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: 2026-07-02 12:00 Beijing Time
  • Summary: - arXiv:2606.30851v1 Announce Type: new.
    • Abstract: Improving the reliability of large language models (LLMs) during inference is a core challenge for 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 distinction 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 Highlights:
    • 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: 2026-07-02 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 approach leverages transformer-based models (RoBERTa-base for English, AfroXLMR-base for Swahili) with class-weighted loss functions to address severe label imbalance, and utilizes threshold tuning for each label to optimize multi-label classification.
  • EN Highlights:
    • 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

  • Release Time: 2026-07-02 12:00 Beijing Time
  • Abstract: - arXiv:2606.30887v1 Announce Type: new.
    • Abstract: Large language models show promise for mental health support, but therapeutic quality can only improve if assessment serves as an actionable control signal rather than a passive metric.
    • We introduce a framework that formulates the generation of therapeutic responses as a decision refinement problem driven by multi-dimensional, human-aligned evaluation.
    • In the first stage, 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: 2026-07-02 12:00 Beijing Time
  • Abstract: - arXiv:2606.30914v1 Announcement Type: new.
    • Abstract: Event detection (ED) systems are typically evaluated on clean, curated text, while their robustness to real-world noise remains largely unexplored, especially for low-resource languages such as Bangla.
    • We introduce a generalized Bangla news event ontology and a benchmark, comprising 9,979 annotated sentences across 40 event subtypes, covering 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

  • Published: 2026-07-02 12:00 Beijing Time
  • Abstract: - arXiv:2606.30943v1 Announcement Type: new.
    • Abstract: Russian and Arabic are among the major languages for scientific communication.
    • Language barriers hinder the exchange of research findings between these communities, impacting the progress of international cooperation and 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

• Representation as a Bottleneck for Mechanistic Interpretability: The Manifestation Unit Protocol

  • Publication Time: 2026-07-02 12:00 Beijing Time
  • Abstract: - arXiv:2607.00089v1 Announcement Type: New.
    • Abstract: Mechanistic interpretability has produced a rich inventory of component-level analyses that describe what neural network components encode and how they interact.
    • However, their outputs are not easily reusable: selectivity tables, circuit diagrams, and feature lists remain locked in individual research notebooks—non-composable, not queryable with natural language, and not directly usable for downstream auditing or intervention.
    • We study the representation layer that sits between these analyses and downstream use as a bottleneck that can be evaluated independently, and introduce Manifestation Units, a typed-tuple protocol (E, S, R, D, G), extended with attention-head primitives (T) for transformer architectures, that organizes component-wise statistics into auto-populated structured fields, queryable via hybrid retrieval.
  • EN Key Points:
    • arXiv:2607.00089v1 Announce Type: new
    • Abstract: Mechanistic interpretability has produced a rich inventory of component-level analyses that characterise what neural-network components encode and how…
    • Their outputs, however, are not easily reusable: selectivity tables, circuit diagrams, and feature lists remain locked in per-study notebooks - non-composable,…
    • We study the representation layer that sits between these analyses and downstream use as a bottleneck that can be evaluated independently, and introduce Manifes…

• SNAP-FM: Sparse Nonlinear Accelerated Projection for Physics-Constrained Generative Modeling

  • Publication Time: 2026-07-02 12:00 Beijing Time
  • Abstract: - arXiv:2607.00095v1 Announcement Type: New.
    • Abstract: Generative models have become a scalable alternative to physical simulations, but they do not guarantee that their outputs adhere to the conservation laws, boundary conditions, and nonlinear invariants governing the underlying physics.
    • Constrained sampling bridges this gap by precisely enforcing such constraints at inference time without retraining, but it incurs a computational cost: repeated projection, correction, and trajectory optimization steps during sampling, which become expensive for nonlinear constraints.
    • Standard machine learning frameworks exacerbate this situation: their dense tensor algebra and limited sparse solver composability obscure the structure naturally arising from physical constraints, making efficient batched nonlinear optimization difficult to achieve in practice.
  • EN Key Points:
    • arXiv:2607.00095v1 Announce Type: new
    • Abstract: Generative models have emerged as scalable surrogates for physical simulation, yet they offer no guarantee that their outputs respect the conservation…

• Constrained sampling closes this gap, enforcing such constraints exactly at inference time without retraining, but at a computational cost: projection, correcti…

• Standard ML frameworks exacerbate this: their dense tensor algebra and limited sparse solver composability obscure the structure that physical constraints natur…

• SemiScope: Disentangling Classifier Tuning and Joint Optimization in Semi-Supervised Security Classification

  • Publication Time: 2026-07-02 12:00 Beijing Time
  • Abstract: - arXiv:2607.00113v1 Announcement Type: New.
    • Labeled data for security classification is scarce.
    • Semi-supervised learning (SSL) propagates labels from a small labeled pool to a larger unlabeled pool.
    • However, security applications often use SSL as a black box: with default parameters, fixed classifiers, and without handling class imbalance caused by pseudo-labels.
  • EN Highlights:
    • arXiv:2607.00113v1 Announce Type: new
    • Abstract: Background
    • Labeled data for security classification is scarce
    • Semi-supervised learning (SSL) propagates labels from a small labeled pool to larger unlabeled pools

• A Filtered Mixture-of-Generators for Fully Synthetic Survival Training

  • Publication Time: 2026-07-02 12:00 Beijing Time
  • Abstract: - arXiv:2607.00127v1 Announcement Type: New.
    • Abstract: Survival analysis models time-to-event data, but in clinical settings, training data is costly and scarce: events accumulate over years of follow-up, cohort sizes are small, and privacy regulations restrict cross-institutional sharing.
    • Tabular generative models promise augmentation and privacy-preserving cohort sharing, yet they are themselves data-hungry—on the small cohorts typical of survival analysis, a single generator is rarely able to characterize the population well enough for downstream models trained on its output to match real-data performance.
    • FoGS (Filtered Mixture-of-Generators for Survival analysis) reframes synthetic-data construction as sample selection rather than generation.
  • EN Highlights:
    • arXiv:2607.00127v1 Announce Type: new
    • Abstract: Survival analysis models time-to-event data, but in clinical settings training data are costly and scarce: events accrue over years of follow-up, coho…
    • Tabular generative models promise augmentation and privacy-preserving cohort sharing, yet are themselves data-hungry – on the small cohorts typical of survival…
    • FoGS (Filtered Mixture-of-Generators for Survival analysis) reframes synthetic-data construction as sample selection rather than generation

• GRPO, Dr. GRPO, and DAPO Are Three Operations on One Number: The Group-Standard-Deviation Identity

  • Publication Time: 2026-07-02 12:00 Beijing Time

• Abstract: - arXiv:2607.00152v1 Announce Type: new.

  • Abstract: The three most popular methods for training language models to reason look like three different tricks.
  • All three adjust a single number: the standard deviation, reflecting the degree to which a prompt’s sampled answers are inconsistent.
  • When training such a model, it answers each question multiple times, and an automatic checker marks each answer as correct or incorrect.
  • EN Highlights:
    • arXiv:2607.00152v1 Announce Type: new
    • Abstract: Three of the most popular methods for training language models to reason look like three different tricks
    • They are not
    • All three adjust a single number: standard deviation, reflecting how much a prompt’s sampled answers disagree

• EVOTS: Evolutionary Transformer Search for Time Series Forecasting

  • Published: 2026-07-02 12:00 Beijing Time
  • Abstract: - arXiv:2607.00154v1 Announce Type: new.
    • Abstract: Evolutionary neural architecture design for multivariate time-series forecasting remains underexplored, with most approaches relying on a fixed Transformer architecture despite significant variations across tasks and forecasting settings.
    • This paper introduces an evolutionary neural architecture search framework for discovering task-adaptive Transformer-like models for time-series forecasting (EVOTS).
    • Architectures are encoded using a modular genome representation that enables flexible composition of attention, feed-forward, and projection components, while a repair mechanism enforces structural validity throughout the evolutionary process.
  • EN Highlights:
    • arXiv:2607.00154v1 Announce Type: new
    • Abstract: Evolutionary neural architecture design for multivariate time-series forecasting remains underexplored, with most approaches relying on fixed Transfor…
    • This paper introduces an evolutionary neural architecture search framework for discovering task-adaptive Transformer-like models for time-series forecasting (EV…
    • Architectures are encoded using a modular genome representation that enables flexible composition of attention, feed-forward, and projection components, while a…

• FRAME: Learning the Adaptation Domain with a Mixture of Fractional-Fourier Experts

  • Published: 2026-07-02 12:00 Beijing Time
  • Abstract: - arXiv:2607.00162v1 Announce Type: new.
    • Abstract: Parameter-Efficient Fine-Tuning (PEFT) reparameterizes weight updates on a fixed basis: low-rank adapters operate in the spatial domain, while a recent series of spectral methods operate in a fixed Fourier domain.
    • We argue that the choice of domain is itself a design degree of freedom that should be learned, and that no single basis is optimal across tasks, layers, or tokens.
    • We introduce the Mixture of Fractional Fourier Experts, a Mixture-of-Experts adapter where each expert carries a learnable fractional Fourier order, which continuously interpolates between the spatial domain (recovering vanilla LoRA) and the Fourier domain (recovering spectral adapters).
  • EN Highlights:
    • arXiv:2607.00162v1 Announce Type: new

• Abstract: Parameter-efficient fine-tuning (PEFT) reparameterizes weight updates in a fixed basis: low-rank adapters operate in the spatial domain, while a recen…

• We argue that the choice of domain is itself a design degree of freedom that should be learned, and that no single basis is optimal across tasks, layers, or tok…

• We introduce Fractional-Fourier Mixture of Experts, a mixture-of-experts adapter in which every expert carries a learnable fractional-Fourier order that continu…

• Verifiable Rewards for Calibrated Probabilistic Forecasting

  • Published: 2026-07-02 12:00 Beijing Time
  • Abstract: - arXiv:2607.00164v1 Announce Type: new.
    • Abstract: Reinforcement learning with verifiable rewards can in principle train calibrated probabilistic forecasters, since proper scoring rules (e.g., the Brier score) are calculated solely from the outcomes and their expectation is minimized by the true probabilities.
    • In practice, it degrades calibration, and existing remedies address epistemic uncertainty, where a model’s confidence is accompanied by a verifiably correct or incorrect answer.
    • We study aleatoric forecasting, where the forecast itself is the output and the label is one stochastic outcome, using NFL in-game win probability as a testbed and betting markets as a reference.
  • EN Highlights:
    • arXiv:2607.00164v1 Announce Type: new
    • Abstract: Reinforcement learning with verifiable rewards can in principle train calibrated probabilistic forecasters, since a proper scoring rule such as the Br…
    • In practice it degrades calibration, and existing remedies address epistemic uncertainty, where a model’s confidence accompanies a verifiably correct or incorre…
    • We study aleatoric forecasting, where the forecast itself is the output and the label is one stochastic outcome, taking NFL in-game win probability as a testbed…

• Scaling Up Thermodynamic AI Models

  • Published: 2026-07-02 12:00 Beijing Time
  • Abstract: - arXiv:2607.00170v1 Announce Type: new.
    • Abstract: Thermodynamic computing devices based on the Ising model show great promise for low-power AI inference and edge computing, but scalable methods for training large-scale models for such hardware remain limited.
    • Previous theories have shown that the time-averaged behavior of a high-temperature Gibbs sampling Ising system can achieve feed-forward neural inference.
    • We translate this theoretical correspondence into a scalable and purely backpropagation-based algorithm for training deep convolutional networks for thermodynamic inference on Ising machine hardware.
  • EN Highlights:
    • arXiv:2607.00170v1 Announce Type: new
    • Abstract: Thermodynamic computing devices based on the Ising model show great promise for low-power AI inference and edge computing, but scalable methods for tr…

• Prior theory shows that the time-averaged behavior of high-temperature Gibbs-sampled Ising systems can implement feed-forward neural inference

• We turn this theoretical correspondence into a scalable and purely backpropagation-based algorithm for training deep convolutional networks for thermodynamic in…

• TallyTrain: Communication-Efficient Federated Distillation

  • Publication Time: 2026-07-02 12:00 Beijing Time
  • Abstract: - arXiv:2607.00173v1 Announcement Type: New.
    • Abstract: Federated learning is bandwidth-bound on two orthogonal axes: model size (which limits how often parameter-averaging methods can afford to merge) and class count (which makes soft-label distillation per probe difficult to implement in large vocabularies).
    • As modern systems scale, both ceilings tighten.
    • We collapse the class-count axis to $\lceil \log_2 C \rceil$ bits per probe by transmitting only each peer’s $\arg\max$ class index, where $C$ is the number of output classes.
  • EN Key Points:
    • arXiv:2607.00173v1 Announce Type: new
    • Abstract: Federated learning is bandwidth-bound on two orthogonal axes: model size, which limits how often parameter-averaging methods can afford to merge, and…
    • Both ceilings tighten as modern systems scale
    • We collapse the class-count axis to $\lceil \log_2 C \rceil$ bits per probe by transmitting only each peer’s $\arg\max$ class index, where $C$ is the number of…