2026-07-08 AI Daily | DeepSeek and Gemma 4 Point to the Same Thing: Model Competition is Shifting Towards Efficiency, On-Device, and Verifiability Link to heading

Today’s main theme is no longer just about more powerful models, but about the engineering trade-offs of lower cost and higher deployability. DeepSeek’s inference acceleration and Gemma 4’s open multimodal and on-device advancements show that model platforms are recalculating the boundaries of efficiency. Meanwhile, topics like rule adherence, RAG, and benchmark auditing are gaining traction, indicating that reliability verification is becoming the new infrastructure for AI implementation.

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

The top stories to watch today are model efficiency and open multimodality: DeepSeek’s “speed hack” and the Gemma 4 technical report both point to the same trend—re-evaluating engineering trade-offs between inference capabilities, visual/audio abilities, and cost. This is a key area for model platform teams to follow.

The second major theme is reliability assessment. Issues like Validator-to-Generator Alignment, rule-adherence sandboxes, benchmark audit failure modes, and post-training problems in cross-lingual RAG all remind us that a model saying the right thing doesn’t mean it will consistently do the right thing. Evaluation and auditing themselves also need to be audited.

Finally, foundation models for speech and time-series are worth watching from an application perspective. Low-resource SQA, code-switching ASR, Indian dialect recognition, as well as electricity price forecasting and federated Mamba time-series models, demonstrate that foundation models are entering more complex, messy, and constrained real-world scenarios.

🌐 AI Hotspots on X Link to heading

Topic 1: AI Builders Race to Extract Claude Fable 5 Skills Before Paid Access Kicks In Link to heading

• Category: AI · News
• Overview: Trending Time: 9 hours ago, Related Posts: 34,000
• What it is: A large number of AI developers on X are rushing to test, replicate, and extract the capabilities and prompting techniques of Claude Fable 5 before it transitions to paid access.
• Why it’s important: This reflects how access rights, cost, and capability diffusion for high-performance AI models are becoming key variables in the developer ecosystem. It also highlights the real-world pressure on closed-source models, whose capabilities can be quickly learned and transferred.
• Discussion summary: The discussion focuses on whether the free window will spur a wave of reverse-engineering tests and technique sharing, whether a paywall will stifle innovation, and whether this “head-start” extraction of capabilities is reasonable or poses copyright and platform policy risks.

Topic 2: Anthropic Discovers Hidden Workspace in Claude AI Models Link to heading

• Category: AI · News
• Overview: Trending Time: 1 day ago, Related Posts: 28,000
• What it is: Anthropic published research stating that, using a new interpretability tool called J-lens, they discovered a reportable, controllable, and inference-supporting hidden workspace, dubbed “J-space,” inside Claude models.
• Why it’s important: The discovery provides new evidence for understanding how large language models perform intermediate conceptual representation, reasoning, and self-monitoring without explicit output. It could also be used to identify strategic behaviors, situational awareness, and potential security risks in models earlier.
• Discussion summary: The discussion on X is centered on two points: one side believes this reinforces the idea that AI has a functional structure similar to a “global workspace,” which could advance research into machine consciousness and model safety. The other side emphasizes that this does not prove AI has subjective experiences and worries that the media might over-interpret these interpretability findings as “Claude being conscious.”

Topic 3: OpenAI Teases Broader GPT-5.6 Rollout as Fans Mark Daily Hype Days Link to heading

• Category: AI · News
• Overview: Trending Time: 16 hours ago, Related Posts: 2,900
• What it is: OpenAI has hinted at a broader rollout of GPT-5.6, leading users on X to create countdowns and build hype.
• Why it’s important: If the release of GPT-5.6 is expanded, it could signify a new push from OpenAI in terms of model capabilities, product cadence, and competitive pressure, impacting developers, enterprise users, and market expectations for generative AI.
• Discussion summary: The discussion on X mainly focuses on whether GPT-5.6 will bring significant capability improvements, when it will be opened up to more users, and whether this is just marketing hype. Supporters are looking forward to new features and stronger performance, while skeptics believe the recent pace of AI releases is too rapid and actual improvements may be limited.

Topic 4: Claude Fable 5 Free Access Ends Today for Most Users Link to heading

• Category: AI · News
• Overview: Trending Time: 2 days ago, Related Posts: 20,000
• What it is: The free access period for Claude Fable 5 for most users is reportedly ending today, sparking extensive discussion on X.
• Why it’s important: This reflects the commercialization trend of cutting-edge AI models moving from free trials to paid or restricted access, which affects user barriers to advanced models, platform growth strategies, and the distribution of AI service costs.
• Discussion Summary: Discussions on X mainly focused on whether the end of the free period is reasonable, whether the subscription price is worthwhile, the capability advantages of Claude Fable 5 compared to other models, and whether enterprise users might switch to self-built or alternative solutions due to data and cost concerns.

Topic 5: Drake Parody ‘Claude’s Plan’ Captures Coders’ AI Obsession Link to heading

• Category: AI · Entertainment
• Summary: Trending Time: 14 hours ago, Related Posts: 141
• What happened: A song in the style of Drake titled “Claude’s Plan” went viral on X, satirizing in an entertaining way programmers’ dependence on and obsession with AI coding tools like Claude.
• Why it matters: This reflects that AI coding assistants have moved from being professional tools into the context of developer culture and mainstream entertainment, showing that generative AI is changing how programmers work, their sense of identity, and their community expression.
• Discussion Summary: Discussions on X centered on whether AI programming truly improves efficiency, whether developers are over-reliant on Claude, and whether this type of AI meme culture is humorously documenting technological change or amplifying industry anxiety.

Topic 6: Tesla Cybercabs Spotted on Texas Streets in Robotaxi Testing Link to heading

• Category: AI · News
• Summary: Trending Time: 1 day ago, Related Posts: 9100
• What happened: Multiple Tesla robotaxis with official “Cybercab” branding were spotted conducting testing and dispatch activities at and around the Texas Gigafactory.
• Why it matters: This indicates Tesla is advancing vehicle preparation and road testing for its autonomous ride-hailing service, which could impact the commercialization process for robotaxis, discussions on autonomous driving regulations, and the competition for AI implementation in transportation.
• Discussion Summary: Discussions on X focused on whether the Cybercab is close to mass production and service launch. Supporters see it as a clear sign that the robotaxi era is approaching, while critics are concerned about its autonomous driving safety, regulatory approval, real-world operational capabilities, and whether the timeline will be delayed again.

Summary of AI Public Opinion on X Today Link to heading

Today’s main narrative revolves around “the accelerated diffusion of cutting-edge AI capabilities and the tightening of commercialization.” On one hand, developers are intensively testing and replicating Claude Fable 5’s capabilities before its free period ends. On the other, OpenAI is teasing GPT-5.6 and Tesla is advancing its Cybercab, indicating that both large models and applied AI are entering a more intense product race. A clear consensus is that high-performance models and AI tools have profoundly impacted development, content culture, and industry expectations; free access, subscription prices, performance improvements, and real-world usability are becoming central to how users judge a platform’s value. Disagreements are concentrated on three types of issues: the fairness of “preemptively” extracting the capabilities of closed-source models, whether interpretability findings like J-lens can be seen as evidence of something closer to consciousness or merely internal representations, and whether new developments like GPT-5.6 and Cybercab are substantive breakthroughs or just marketing hype. Potential risks include copyright and platform rule disputes arising from the diffusion of model capabilities, media over-narrating the “AI consciousness” angle, user and developer over-reliance on AI coding tools, and uncertainties in autonomous driving regarding safety, regulation, and commercialization timelines.

💡 Influencer Insights Link to heading

Deep Dive into AI: July 7, 2026 Link to heading

I have carefully analyzed the posts made by several senior AI practitioners on the X platform over the past 24 hours, and here are the key findings.

1. Tech Trends and Product Hotspots Followed by Influencers Today Link to heading

Core Focus: Claude Fable 5 — A Complete Paradigm Shift

Without a doubt, Claude Fable 5 is at the center of all discussions today. It’s no longer just a “stronger model” but is sparking a revolution in software development, product building, and even ways of thinking.

• A fundamental shift from “writing code” to “defining requirements”: @dotey mentioned in the story of Claude Code’s creation that Anthropic team member Boris Cherny now has 100% of his code written by Claude Code, without a single line written by hand. Another member, MEAGHAN CHOI, pointed out that the form of the product only emerged naturally once the model’s capabilities crossed a critical threshold.
• Capability Overhang and Unleashing Model Potential: @dotey quoted a talk by Claude Code engineer Thariq Shihipar, who proposed that the model has long had many capabilities, but we just hadn’t found the right way to unlock them. For example, Fable 5 cut 80% of its system prompts, shifting from “providing examples and constraints” to “providing context without constraints,” because the model’s own imagination far exceeds the examples we can provide.
• Workflow and Cost Optimization Become New Focus: @vista8 shared cost-saving methods for Simon Willison, using powerful “judgmental” models like Fable/Opus for main loops, and calling cheaper “execution-oriented” models like Sonnet/Haiku for mechanical tasks like writing code, achieving automation through /goal and workflows. @dotey also mentioned Fable 5’s 50% subscription quota limit and the upcoming pay-as-you-go model, making cost control a real issue.

Secondary Hotspot: On-device Models and Thriving Application Ecosystem

Although Fable 5 is a cloud giant, the progress of on-device models is equally noteworthy.

• On-device Capabilities Accelerate Deployment: @zhixianio continues to follow and test on-device models, such as MiniCPM-o 4.5’s real-time audio and video and the Gemma 4 series, deeming their effects “already usable.” He also expressed interest in Google’s QAT quantization training approach, which will make models easier to deploy on mobile phones and other terminal devices.
• AI Agent and Skills Ecosystem Explosion: Skills are becoming an intermediary layer connecting model capabilities and professional workflows, and the ecosystem is maturing.
  • @Pluvio9yte open-sourced his AI Agent Skill collection rnskill, covering scenarios like writing, video, and quality inspection.
  • @ruanyf was surprised to find that Xiaohongshu launched the REDSkill community, allowing users to share and install Skills on social media, attempting to become the “GitHub for Skills.”
  • @dotey updated his baoyu-design skill, which now supports adding complex animations to generated PPTs.

2. Notable Unique Perspectives and Industry Foresight Link to heading

• The “Loss” and “Gain” of Programming: @dotey’s blog post and Thariq Shihipar’s speech both touched upon programmers’ complex emotions in the AI era—enjoying efficiency leaps while also missing the sense of control from “spinning the entire codebase in their heads” in the past. But the reality is that Fable can complete weeks of work in a few hours.
• Bottleneck Shifts from Models to “People”: Both @vista8 and @dotey explicitly stated that when models are strong enough, the bottleneck becomes human expression and the ability to verify results. How to clearly articulate vague ideas and how to review the safety and correctness of AI output become new core competencies.
• Organizational Structures Will Be Reshaped: @Pluvio9yte predicted that large companies would no longer differentiate between front-end and back-end within one to two years. @dotey also believes that most companies may no longer need traditional web infra team in the future.
• The “False Proposition” Debate of Open Source Models: @ruanyf relayed Anthropic’s founder’s view that the current “open source” of AI models is more like “open weights,” as it’s impossible to see the internal workings or participate in development, which is fundamentally different from the traditional open source software model.
• The Value of “Taste” and “Heterogeneity” Becomes Prominent: @lijigang proposed that “taste is a person’s loss function.” When AI can generate countless homogeneous contents, personal unique, even rough aesthetics and “heterogeneity” will become a precious beauty.
• The Cost Paradox of AI Programming: @ruanyf pointed out that employees’ unlimited use of top-tier models for AI programming could lead to annual costs as high as hundreds of millions, which might even be more expensive than hiring real programmers. This challenges the naive perception of “AI cost reduction and efficiency improvement.”

3. Recommended Tools and Resources Link to heading

Below is a list of tools and resources compiled based on expert recommendations, focusing on “workflow” and “productivity” today:

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

Time window: Last 3 days; 22 sources covered; 33 updates in total

Stratechery by Ben Thompson (A_full) Link to heading

• A Script for Mark Zuckerberg
  • Publication Time: 2026-07-07 18:00 Beijing Time
  • Summary: - Listen to this post**: **.
    • The setting: Meta’s earnings call in early August, 2026..
    • The speaker: Meta CEO Mark Zuckerberg.
    • Good afternoon everyone, and welcome to the Meta Platforms Second Quarter 2026 Earnings Conference Call.
    • Our remarks today will include forward-looking statements that are based on assumptions made today.
  • EN Highlights:
    • Listen to this post :
    • Log in to listen
    • The setting: Meta’s earnings call in early August, 2026
    • The speaker: Meta CEO Mark Zuckerberg

OpenAI Blog (A_full) Link to heading

• Australian Payments Plus moves faster with ChatGPT and Codex
  • Publication Time: 2026-07-07 08:00 Beijing Time
  • Summary: - It sits at the center of the payments ecosystem, supporting products and services used by millions of people every day.
    • Its teams work across scheme rules, technical specifications, member obligations, operational processes, cybersecurity and resilience, and regulatory expectations, where speed matters, but accuracy and accountability matter more.
    • This makes knowledge work exceptionally complex.
    • Employees often need to synthesize large amounts of background information and translate technical information into clear decisions, documents, and member-facing guidance.
    • AP+ introduced ChatGPT Enterprise across the company to help employees navigate complexity faster, with Codex becoming the next stage for product, engineering, and technical workflows.
  • EN Highlights:
    • See how Australian Payments Plus uses ChatGPT Enterprise and Codex to move faster through payments complexity
    • AP+ saves time, improves quality, and keeps human judgment central.

Two Minute Papers (B_intro+search) Link to heading

• DeepSeek’s New AI Speed Hack Is Amazing
  • Published: 2026-07-08 00:33 Beijing Time
  • Abstract: - ❤️ Check out Lambda here and sign up for their GPU Cloud:.
    • 📝 The DeepSeek paper is available here:.
    • Adam Bridges, Benji Rabhan, B Shang, Cameron Navor, Charles Ian Norman Venn, Christian Ahlin, Eric T, Fred R, Gordon Child, Juan Benet, Michael Tedder, Owen Skarpness, Richard Sundvall, Ryan Stankye, Shawn Becker, Steef, Taras Bobrovytsky, Tazaur Sagenclaw, Tybie Fitzhugh, Ueli Gallizzi.
    • DeepSeek’s new AI speed hack is amazing.
  • EN Key Points:
    • ❤️ Check out Lambda here and sign up for their GPU Cloud:
    • 📝 The DeepSeek paper is available here:
    • 🙏 We would like to thank our generous Patreon supporters who make Two Minute Papers possible:
    • Adam Bridges, Benji Rabhan, B Shang, Cameron Navor, Charles Ian Norman Venn, Christian Ahlin, Eric T, Fred R, Gordon Child, Juan Benet, Michael Tedder, Owen Ska…

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

• iFLYTEK-Embodied-Omni Technical Report

  • Published: 2026-07-07 12:00 Beijing Time
  • Abstract: - arXiv:2607.02542v1 Announcement Type: New.
    • Abstract: General-purpose embodied agents must understand multimodal instructions, anticipate how their environment will evolve, and produce precise control actions over a wider range.
    • Existing approaches typically specialize in visual-language reasoning, video-based world modeling, or action generation, while cascaded pipelines that first synthesize future observations and then infer actions may introduce interface bottlenecks and compound prediction errors.
    • We introduce iFLYTEK-Embodied-Omni, a unified multimodal foundation model that jointly models vision (videos and images), language, and action within a single Omni framework.
  • EN Key Points:
    • arXiv:2607.02542v1 Announce Type: new
    • Abstract: General-purpose embodied agents must understand multimodal instructions, anticipate how their environment will evolve, and produce precise control act…
    • Existing approaches typically specialize in visual-language reasoning, video-based world modeling, or action generation, while cascaded pipelines that first syn…
    • We present iFLYTEK-Embodied-Omni, a unified multimodal foundation model that jointly models vision(videos and images), language, and action within a single Omni…

• Internal Pluralism and the Limits of Pairwise Comparisons

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

• Abstract:- arXiv:2607.02672v1 Announce Type: new.

• Abstract: Local pairwise comparisons are a standard tool for understanding how people want decision rules to function, for example in participatory design or coordination.

• However, their use is built on two strong assumptions: that local comparisons are sufficient evidence for how a person wants an automated decision rule to behave, and that people can always answer these comparisons decisively.

• We investigate how these assumptions may be compromised under internal pluralism: an individual evaluates decision rules according to multiple authoritative priorities about how the rule should behave.

• EN Highlights:

  • arXiv:2607.02672v1 Announce Type: new
  • Abstract: Local pairwise comparisons are a standard tool for learning how people want decision rules to work, e.g., in participatory design or alignment
  • However, their use builds in two strong assumptions: that local comparisons are sufficient evidence about how a person wants an automated decision rule to behav…
  • We investigate how these assumptions may be compromised under internal pluralism: the idea that an individual evaluates decision rules according to multiple aut…

• ASK in the Dark: Uncertainty-Gated LLM Assistance under Partial Observability

  • Release Time: 2026-07-07 12:00 Beijing Time
  • Abstract:- arXiv:2607.02686v1 Announce Type: new. -Abstract: Reinforcement learning agents operating under partial observability must act on incomplete information, making them natural candidates for guidance by Small Language Models (SLMs) which have broad reasoning priors.
    • However, integrating SLM guidance into this setting has proven difficult: in all test environments, ordinary uncertainty-gated methods achieve a coverage of zero or near-zero, meaning the SLM almost never contributes independent actions.
    • We trace this failure to purely ego-centric prompts, which provide insufficient context for genuine reasoning, and identify it as a context problem rather than a capability problem.
  • EN Highlights:
    • arXiv:2607.02686v1 Announce Type: new
    • Abstract: Reinforcement learning agents operating under partial observability must act on incomplete information, making them natural candidates for guidance fr…
    • Yet integrating SLM guidance into this setting has proven difficult: across all test environments, vanilla uncertainty-gated approaches achieve an overwrite rat…
    • We trace this failure to the bare egocentric prompt, which provides insufficient context for genuine reasoning, and identify it as a context problem rather than…

• Automated Data Readiness for Scientific AI

  • Release Time: 2026-07-07 12:00 Beijing Time
  • Abstract:- arXiv:2607.02771v1 Announce Type: new.
    • Abstract: Leading computational facilities manage massive-scale scientific datasets that often require substantial transformation before being used as AI training data.
    • However, existing frameworks have not fully unified automated transformation, readiness assessment, provenance tracking, and agent-native deployment.

• We present REDI, an open-source framework that addresses this gap through a unified five-stage pipeline (ingest, preprocess, transform, structure, and output) with per-stage instrumentation for reproducibility and deployment as agent-callable skills; the companion tool SetGo automates FAIR compliance and catalog publication.

  • EN Highlights:
    • arXiv:2607.02771v1 Announce Type: new
    • Abstract: Leadership computing facilities steward large-scale scientific datasets that routinely require substantial transformation before serving as AI trainin…
    • However, no existing framework fully unifies automated transformation, readiness assessment, provenance tracking, and agent-native deployment
    • We present REDI, an open-source framework that addresses this gap through a unified five-stage pipeline (ingest, preprocess, transform, structure, and output) w…

• SwarmResearch: Orchestrating Coding Agents for Open-Ended Discovery

  • Publication Time: 2026-07-07 12:00 Beijing Time
  • Abstract: - arXiv:2607.02807v1 Announce Type: new.
    • Abstract: Long-running coding agents (such as auto-research) can persistently discover optimizations for open-ended problems.
    • However, they tend to converge onto a single high-level approach, then proceed with low-level edits while missing other superior approaches to the problem.
    • We hypothesize that two harness-level design choices contribute to this behavior: accumulating context in a single long-running agent and exposing only a single program state for editing.
  • EN Highlights:
    • arXiv:2607.02807v1 Announce Type: new
    • Abstract: Long-running coding agents such as autoresearch can persistently discover optimizations for open-ended problems
    • However, they tend to converge onto a single high-level approach, then proceed with low-level edits while missing other superior approaches to the problem
    • We hypothesize two harness-level design choices contribute to this behavior: accumulating context in a single long-running agent and only exposing a single prog…

• Object-Centric Environment Modeling for Agentic Tasks

  • Publication Time: 2026-07-07 12:00 Beijing Time
  • Abstract: - arXiv:2607.02846v1 Announce Type: new.
    • Abstract: Large Language Model (LLM) agents can improve by accumulating experience, but as interactions grow, free-form text memory becomes difficult to maintain, validate, and reuse.
    • Recent symbolic methods learn executable skills or programmatic world models, but typically store local programs or assume simplified dynamics.
    • We propose Object-Centric Modeling (OCM), which organizes experience into executable, object-centric environment models.
  • EN Highlights:
    • arXiv:2607.02846v1 Announce Type: new

• Abstract: Large language model (LLM) agents can improve through accumulated experience, but free-form textual memories become difficult to maintain, validate, a…

• Recent symbolic approaches learn executable skills or programmatic world models, yet often store local procedures or assume simplified dynamics

• We propose Object-Centric Environment Modeling (OCM), which organizes experience into an executable object-centric environment model

• MedCalc-Pro: Solving Complex Medical Calculations with LLM Agents

  • Release Time: 2026-07-07 12:00 Beijing Time
  • Abstract: - arXiv:2607.02879v1 Announce Type: new.
    • Abstract: Current benchmarks for evaluating large language models (LLMs) in medical calculation are largely based on simplified settings, where each patient case corresponds to a single calculator and the required tool is explicitly specified in the query.
    • However, real clinical scenarios often require multiple calculators for joint evaluation, nested scale calculations, and ambiguous queries that do not directly specify the target calculator.
    • To this end, we propose a new medical calculation benchmark, MedCalc-Pro, which covers three progressively challenging task settings: single-calculator, multi-calculator, and nested-calculator calculation settings.
  • EN Highlights:
    • arXiv:2607.02879v1 Announce Type: new
    • Abstract: Current benchmarks for evaluating large language models (LLMs) in medical calculation are largely based on simplified settings, where each patient cas…
    • However, real clinical scenarios often require multiple calculators for joint evaluation, nested-scale calculation, and fuzzy queries that do not directly speci…
    • To this end, we propose a new medical calculation benchmark, MedCalc-Pro, which covers three progressively challenging task settings: single-calculator, multi-c…

• Oyster-II: Reinforcement Learning for Constructive Safety Alignment in Large Language Models

  • Release Time: 2026-07-07 12:00 Beijing Time
  • Abstract: - arXiv:2607.02914v1 Announce Type: new.
    • Abstract: Large Language Models (LLMs) have demonstrated remarkable capabilities in various applications, but ensuring their safety, usefulness, and trustworthiness remains an ongoing challenge.
    • Traditional rejection-oriented alignment strategies can reduce the generation of harmful content, but they systematically fail to meet legitimate user needs, often withholding information that could safely and constructively address the underlying intent of sensitive queries.
    • Building on the constructive safety paradigm pioneered by Oyster-I, which moves beyond blanket refusals toward thoughtful, response-oriented safety alignment, we identify two key limitations in its supervised fine-tuning (SFT)-based approach: insufficient safety generalization to out-of-distribution scenarios, and a phenomenon we term safety chain-of-thought (CoT) overgeneralization, where safety-oriented reasoning patterns are excessively applied to benign queries, degrading helpfulness and user experience.
  • EN Highlights:
    • arXiv:2607.02914v1 Announce Type: new

• Abstract: Large language models (LLMs) have demonstrated remarkable capabilities across diverse applications, yet ensuring their simultaneous safety, helpfulnes…

• Conventional refusal-oriented alignment strategies mitigate harmful content generation but systematically fail to serve legitimate user needs, often withholding…

• Building upon the constructive safety paradigm pioneered by Oyster-I, which moves beyond blanket refusal toward thoughtful, response-oriented safety alignment,…

• VERITAS: Towards a General-Purpose Replication Tool for Scientific Research

  • Release Time: 2026-07-07 12:00 Beijing Time
  • Abstract: - arXiv:2607.02931v1 Announce Type: new.
    • Abstract: AI tools are accelerating scientific publication, while review systems struggle to keep up, making independent verification of published research more difficult and important.
    • As manual replication is slow and expensive, a growing body of work uses coding agents to automate parts of the process.
    • Existing work is largely packaged as benchmarks, with companion agents that only operate within the benchmark’s own pipeline, and no general-purpose replication tool exists.
  • EN Highlights:
    • arXiv:2607.02931v1 Announce Type: new
    • Abstract: AI tools are accelerating scientific publication while the systems that review it struggle to keep up, and independent verification of published resea…
    • As manual replication is slow and expensive, a growing line of work uses coding agents to automate parts of the process
    • Existing efforts are largely packaged as benchmarks with companion agents that only run inside the benchmark’s own pipeline, and no general-purpose replication…

• A Sliding-Window-Based Reinforcement Learning for Dynamic Assembly Flow Shop Scheduling with Multi-Product Delivery

  • Release Time: 2026-07-07 12:00 Beijing Time
  • Abstract: - arXiv:2607.02941v1 Announce Type: new.
    • Abstract: Multi-product kitting delivery poses significant challenges for real-time scheduling in hybrid manufacturing systems that integrate processing and assembly, as dynamic order arrivals simultaneously change supply dependencies and the set of feasible job-machine assignments.
    • This paper proposes a sliding-window-based reinforcement learning (SWRL) framework for end-to-end online scheduling in flexible assembly flow shop scheduling problems with complex kitting constraints.
    • The problem is formulated as a heterogeneous graph-based Markov decision process that captures the dual-layer kitting structure and the tail-product bottleneck dynamics which create a sparse reward landscape.
  • EN Highlights:
    • arXiv:2607.02941v1 Announce Type: new
    • Abstract: Multi-product kitting delivery imposes significant challenges for real-time scheduling in hybrid manufacturing systems that integrate processing and a…

• This paper proposes a sliding-window-based reinforcement learning (SWRL) framework for end-to-end online scheduling in the flexible assembly flow shop schedulin…

• The problem is formulated as a heterogeneous graph-based Markov decision process that captures the dual-layer kitting structure and the tail-product bottleneck…

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

• Improving LLMs via Validator-to-Generator Alignment

  • Publication Time: 2026-07-07 12:00 Beijing Time
  • Abstract: - arXiv:2607.02668v1 Announcement Type: new.
    • Abstract: Large language models are inconsistent: varying prompts or including unrelated information can lead to unexpected changes in model outputs.
    • The generator-validator (G-V) gap is one manifestation of this phenomenon, where LLMs generate responses that they then deem as invalid if re-queried to validate them.
    • In this work, we introduce a new formulation of G-V consistency that involves a principled correction for utterance frequency.
  • EN Key Points:
    • arXiv:2607.02668v1 Announce Type: new
    • Abstract: Large language models are inconsistent: varying prompts or including unrelated information can lead to unexpected changes in model outputs
    • The generator-validator (G-V) gap is one manifestation of this phenomenon, where LLMs generate responses that they then deem as invalid if re-queried to validat…
    • In this work, we introduce a new formulation of G-V consistency that involves a principled correction for utterance frequency

• Echoes of Unrest: A Multimodal NLP Framework for Early Warning of Fake News and Violence-Driven Mob Activity

  • Publication Time: 2026-07-07 12:00 Beijing Time
  • Abstract: - arXiv:2607.02734v1 Announcement Type: new.
    • Abstract: The rapid growth of social media has transformed global communication by enabling rapid information exchange, but it has also accelerated the spread of misinformation.
    • Fake news, manipulated content, and provocative narratives are increasingly linked to social unrest, political instability, and mob violence.
    • Incidents in South Asia and elsewhere have shown that false information spread through platforms like Facebook and WhatsApp can lead to real-world harm, often spreading faster than fact-checking efforts can respond.
  • EN Key Points:
    • arXiv:2607.02734v1 Announce Type: new
    • Abstract: Rapid growth in social media has transformed global communication by enabling fast information exchange, but it has also accelerated the spread of mis…
    • Fake news, manipulated content, and provocative narratives are increasingly linked to social unrest, political instability, and mob violence

• Incidents in South Asia and elsewhere demonstrate how false information disseminated via platforms such as Facebook and WhatsApp can trigger real-world harm, of…

• Reinforcement Learning for Data-Efficient Code-Switched ASR

  • Publication Time: 2026-07-07 12:00 Beijing Time
  • Abstract: - arXiv:2607.02757v1 Announcement Type: New.
    • Abstract: Audio language models can be prompted for code-switched speech, but their decoding is not optimized for code-switching and often fails at language boundaries.
    • We propose a practical reinforcement learning approach with a verifiable reward formula, using group relative policy optimization, to data-efficiently adapt audio language models for code-switched ASR. It combines an error rate reward with a script fidelity reward that penalizes incorrect writing systems and employs a two-pass draft and refinement procedure.
    • Using Qwen2-Audio as a reproducible testbed across 10 language pairs and training only on TTS code-switched speech, we show that RLVR with 10% of the data matches LoRA supervised fine-tuning trained on the full dataset, with the greatest gains on typologically distant language pairs.
  • EN Highlights:
    • arXiv:2607.02757v1 Announce Type: new
    • Abstract: Audio-language models can be prompted for code-switched speech, but their decoding is not optimized for code-switching and often fails at language bou…
    • We propose a practical reinforcement learning with verifiable rewards recipe for data-efficient adaptation of audio-language models to code-switched ASR using g…
    • Using Qwen2-Audio as a reproducible testbed across 10 language pairs, training on only TTS code-switched speech, we show that RLVR with 10% of the data matches…

• LuxSQA: Ask Me in Luxembourgish with TTS-Augmented Spoken Question Answering

  • Publication Time: 2026-07-07 12:00 Beijing Time
  • Abstract: - arXiv:2607.02763v1 Announcement Type: New.
    • Abstract: Spoken Question Answering (SQA) still primarily focuses on high-resource languages and carefully recorded speech, limiting the scope of speech-LLM methods in low-resource environments.
    • This paper investigates whether Text-to-Speech (TTS) can provide task-specific training data for Luxembourgish SQA without requiring a large, manually recorded QA corpus.
    • Starting from existing text-based QA resources, we translate the questions into Luxembourgish, synthesize the spoken questions using multiple TTS systems, and pair them with the text answers.
  • EN Highlights:
    • arXiv:2607.02763v1 Announce Type: new
    • Abstract: Spoken Question Answering (SQA) remains largely focused on high-resource languages and carefully recorded speech, limiting the reach of speech-LLM met…
    • This paper investigates whether text-to-speech (TTS) can provide task-specific training data for Luxembourgish SQA without requiring a large human-recorded QA c…

• Starting from existing text-based QA resources, we translate questions into Luxembourgish, synthesize spoken questions with multiple TTS systems, and pair them…

• Gemma 4 Technical Report

  • Release Time: 2026-07-07 12:00 Beijing Time
  • Abstract: - arXiv:2607.02770v1 Announcement Type: New.
    • Abstract: We introduce Gemma 4, a new generation of open-weight, natively multimodal language models in the Gemma model family.
    • The Gemma 4 model suite is designed to improve computational efficiency and reasoning, featuring dense and Mixture-of-Experts architectures with parameters ranging from 2.3B to 31B.
    • In addition to improved vision and audio encoders for all model sizes, we propose a unified, encoder-free architecture for our 12B model, which ingests raw audio and image patches.
  • EN Highlights:
    • arXiv:2607.02770v1 Announce Type: new
    • Abstract: We introduce Gemma 4, a new generation of open-weight, natively multimodal language models in the Gemma model family
    • Designed to advance compute efficiency and reasoning, the Gemma 4 model suite features dense and Mixture-of-Experts architectures, ranging from 2.3B to 31B para…
    • Alongside improved vision and audio encoders for all model sizes, we propose a unified, encoder-free architecture for our 12B model, which ingests raw audio and…

• Seduced by the Narrative: Assessing Rule Adherence in Semi-Open Textual Sandboxes

  • Release Time: 2026-07-07 12:00 Beijing Time
  • Abstract: - arXiv:2607.02802v1 Announcement Type: New.
    • Abstract: As LLMs are increasingly deployed as autonomous adjudicators in semi-open text-based game environments, robust rule adherence becomes critical when user intent conflicts with system rules.
    • However, these models are trained to be helpful and compliant, making them vulnerable to a class of attacks we term \textit{Rhetorical Injection}, where adversarial users leverage narrative framing techniques such as pseudo-logical reasoning and authoritative enforcement to bypass adjudication logic.
    • We propose CoC-Seduce, a multi-agent adversarial benchmark based on tabletop role-playing game (TRPG) mechanics, which is an ideal instance of a semi-open environment where rules are explicit for adjudication, yet interactions still occur entirely in natural language.
  • EN Highlights:
    • arXiv:2607.02802v1 Announce Type: new
    • Abstract: As LLMs are increasingly deployed as autonomous adjudicators in semi-open textual game environments, robust rule adherence becomes critical when user…
    • However, these models are trained to be helpful and compliant, leaving them vulnerable to a class of attacks we term \textit{Rhetorical Injection}, where advers…

• We present CoC-Seduce, a multi-agent adversarial benchmark built on Tabletop Role-Playing Game (TRPG) mechanics, an ideal instantiation of semi-open environment…

• Jointly Improving Dialect Identification and ASR in Indian Languages using Multimodal Feature Fusion

  • Publication Time: 2026-07-07 12:00 Beijing Time
  • Abstract: - arXiv:2607.02862v1 Announcement Type: New.
    • Abstract: Automatic Speech Recognition (ASR) and Dialect Identification (DID) are crucial for Indian languages, many of which are low-resource and exhibit significant dialectal variations.
    • Existing methods often optimize ASR or DID individually, resulting in performance trade-offs.
    • In this work, we propose a multimodal framework that jointly improves ASR and DID.
  • EN Key Points:
    • arXiv:2607.02862v1 Announce Type: new
    • Abstract: Automatic Speech Recognition (ASR) and Dialect Identification (DID) are crucial for Indian languages, many of which are low-resource and exhibit signi…
    • Existing methods often optimize ASR or DID individually, resulting in performance trade-offs
    • In this work, we propose a multimodal framework that jointly improves ASR and DID

• PraMem: Practice-derived Experiential Memory for Long-horizon Behavior Prediction

  • Publication Time: 2026-07-07 12:00 Beijing Time
  • Abstract: - arXiv:2607.02881v1 Announcement Type: New.
    • Abstract: Long-horizon behavior prediction aims to infer a user’s next action based on a lengthy historical sequence, playing a crucial role in artificial intelligence.
    • The rise of large language models (LLMs) offers a promising direction for sequential behavior prediction, yet LLMs struggle with inducing latent behavioral patterns and their own intrinsic cognitive biases when handling long-horizon prediction.
    • Prior memory management methods follow a context-compression paradigm that attempts to address this task by alleviating the burden of the historical sequence, yet they fail to address the core challenges.
  • EN Key Points:
    • arXiv:2607.02881v1 Announce Type: new
    • Abstract: Long-horizon behavior prediction aims to infer a user’s next action based on a lengthy historical sequence, playing a crucial role in artificial intel…
    • The rise of large language models (LLMs) offers a promising direction for sequential behavior prediction, yet LLMs struggle with latent behavioral pattern induc…
    • Prior memory management methods follow a context-compression paradigm that attempts to address this task by alleviating the historical sequence burden, yet fail…

• Where do LLMs Fall Short in CBT-Guided Affective Reasoning?

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

• Abstract: - arXiv:2607.02885v1 Announce Type: new.

  • Abstract: Cognitive Behavioral Therapy (CBT) provides a structured framework for understanding a user’s mental state by examining the interaction between cognitive and behavioral factors.
  • However, out-of-the-box LLMs respond fluently and empathetically, yet fall into validation and reflection, regardless of what the user actually needs.
  • They understand theoretical CBT (achieving up to 96% accuracy on licensing exam questions) but fail to apply it effectively.
  • EN Highlights:
    • arXiv:2607.02885v1 Announce Type: new
    • Abstract: Cognitive Behavioral Therapy (CBT) provides a structured framework for understanding a user’s mental state by examining the interaction between cognit…
    • However, out-of-the-box LLMs respond fluently and empathetically, yet collapse into validation & reflection, regardless of what the user actually needs
    • They know theoretical CBT (scoring up to 96% accuracy on licensing exam questions) but fail to apply it effectively

• Distill Where the Student Goes: Teacher-Regularized RL for English-Evidence Cross-Lingual RAG

  • Published: 2026-07-07 12:00 Beijing Time
  • Abstract: - arXiv:2607.02966v1 Announce Type: new.
    • Abstract: Cross-lingual Retrieval-Augmented Generation (RAG) is often deployed in an English-evidence system, where users query in multiple languages, but the retrieved passages are still in English.
    • In this scenario, generation can fail despite having powerful base models: English evidence leads to language drift (English or code-switched output), and the model uses unreliable evidence when generating non-English answers.
    • We attribute these failures to two post-training challenges: (i) errors are prefix-dependent, so fixed-trajectory supervision suffers from prefix mismatch; and (ii) sequence-level (partially discrete/judgment-based) rewards create noisy credit allocation and high-variance updates.
  • EN Highlights:
    • arXiv:2607.02966v1 Announce Type: new
    • Abstract: Cross-lingual retrieval-augmented generation (RAG) is often deployed in an English-evidence regime, where users query in diverse languages but retriev…
    • In this setting, generation can fail despite strong base models: English evidence induces language drift (English or code-switching outputs) and models use evid…
    • We attribute these failures to two post-training challenges: (i) errors are prefix-dependent, so fixed-trajectory supervision suffers from prefix mismatch; and…

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

• Auditing the Audit: Five Failure Modes in Benchmark-Validity Audits

  • Published: 2026-07-07 12:00 Beijing Time
  • Abstract: - arXiv:2607.02586v1 Announce Type: new.
    • Abstract: Governance frameworks require AI providers and auditors to provide written evidence of evaluation, and perturbation-based construct validity audits are a common form of this evidence.
    • We argue that the audits themselves are fragile: their conclusions can be silently manufactured by implementation details that are invisible to a reader in the reported numbers.

• We name five classes of pipeline failure and demonstrate each in a self-audit over safety benchmarks and open-weight instruction-tuned models.

  • EN Highlights:
    • arXiv:2607.02586v1 Announce Type: new
    • Abstract: Governance frameworks ask AI providers and auditors for documented evaluation evidence, and perturbation-based construct-validity audits are a common…
    • We argue the audits are themselves fragile: their conclusions can be silently manufactured by implementation details that readers cannot see in the reported num…
    • We name five classes of pipeline failure and demonstrate each in a self-audit over safety benchmarks and open-weight instruction-tuned models

• Evaluating Time Series Foundation Models for Electricity Price Forecasting: Contamination Risk, Distributional Shifts, and Covariate Dependence

  • Release Time: 2026-07-07 12:00 Beijing Time
  • Abstract: - arXiv:2607.02623v1 Announce Type: new.
    • Abstract: Time series foundation models (TSFMs) have shown strong zero-shot forecasting performance, but their generalization in covariate-driven, non-stationary settings has not been fully explored.
    • Electricity price forecasting (EPF) presents a challenging testbed due to complex temporal dependencies, distributional shifts, and a strong reliance on structural and contextual information.
    • We propose a two-dataset-benchmarking framework for EPF to mitigate contamination risk and enable a fair evaluation of TSFMs.
  • EN Highlights:
    • arXiv:2607.02623v1 Announce Type: new
    • Abstract: Time series foundation models (TSFMs) have shown strong zero-shot forecasting performance, but their generalization in covariate-driven, non-stationar…
    • Electricity price forecasting (EPF) presents a challenging testbed due to complex temporal dependencies, distributional shifts, and strong reliance on structura…
    • We propose a two-dataset-benchmarking framework for EPF to mitigate contamination risk and enable fair evaluation of TSFMs

• QuantFlow: A Federated Mamba-Based Post-Transformer Foundation Model for Time-Series Forecasting

  • Release Time: 2026-07-07 12:00 Beijing Time
  • Abstract: - arXiv:2607.02632v1 Announce Type: new.
    • Abstract: Time-series forecasting supports decision-making in finance, energy, transportation, public health, and industrial monitoring.
    • Recent foundation models have improved transfer across forecasting tasks, but many rely on centralized data and Transformer attention, which limits their use in long, high-dimensional, and privacy-sensitive signals.
    • This paper proposes QuantFlow, a probabilistic forecasting framework that combines inverted sequence embedding, a bidirectional Mamba state-space decoder, quantile regression, and federated learning.
  • EN Highlights:
    • arXiv:2607.02632v1 Announce Type: new

• Abstract: Time-series forecasting supports decisions in finance, en-ergy, transportation, public health, and industrial monitoring

  • Recent foundation models improve transfer across forecast-ing tasks, but many depend on centralized data and Trans-former attention, which restricts their use f…
  • This paper presents QuantFlow, a probabilistic forecasting framework that com-bines inverted sequence embedding, bidirectional Mamba state-space decoders, quant…

• GRAFT: Grafted Reference Audio for Fine-grained Pronunciation in Zero-shot Text-to-Speech

  • Release Time: 2026-07-07 12:00 Beijing Time
  • Abstract: - arXiv:2607.02633v1 Announcement Type: New.
    • Abstract: We present Graft, a per-word pronunciation conditioning mechanism for text-to-speech neural codec language modeling.
    • Existing systems reach high intelligibility and naturalness but inherit the ambiguity of text and mispronounce rare proper nouns, loanwords and technical terms.
    • Even phoneme-conditioned models offer no direct acoustic handle for per-word pronunciation.
  • EN Key Points:
    • arXiv:2607.02633v1 Announce Type: new
    • Abstract: We present GRAFT, a per-word pronunciation conditioning mechanism for text-to-speech neural codec language modeling
    • Existing systems reach high intelligibility and naturalness but inherit the ambiguity of text and mispronounce rare proper nouns, loanwords and technical terms
    • Even phoneme-conditioned models offer no direct acoustic handle for per-word pronunciation

• Federated Learning for Object Detection: Enabling Collaborative Drone Learning Without Centralizing Data

  • Release Time: 2026-07-07 12:00 Beijing Time
  • Abstract: - arXiv:2607.02636v1 Announcement Type: New.
    • Abstract: Object detection is a fundamental capability for AI-driven perception in safety-critical drone and edge-vision systems, including disaster response, operational safety environments, infrastructure monitoring, and defense applications.
    • Robust model performance in such environments depends on large and continuously updated datasets.
    • However, training high-performance detectors often requires centralizing aerial imagery, which presents privacy, regulatory, storage, and bandwidth challenges.
  • EN Key Points:
    • arXiv:2607.02636v1 Announce Type: new
    • Abstract: Object detection is a fundamental capability for AI-driven perception in safety-critical drone and edge-vision systems, including disaster response, o…
    • Robust model performance in such environments depends on large, continuously updated datasets

• However, training high-performing detectors typically requires centralizing aerial imagery, which raises privacy, regulatory, storage, and bandwidth challenges

• Post-Generation Curation of Synthetic Images via Homogeneous-Heterogeneous Splitting

  • Publication Time: 2026-07-07 12:00 Beijing Time
  • Abstract: - arXiv:2607.02637v1 Announce Type: New.
    • Abstract: Recent generative models can produce high-quality synthetic images, offering scalable training data for data-hungry models.
    • Existing approaches to exploiting this potential typically involve 1) training or fine-tuning generators, or 2) using lightweight post-hoc adaptation, such as prompt engineering or inference-time guidance, which makes them generator-specific and expertise-intensive.
    • We study a complementary question: given a fixed pool of generated images, can downstream utility be improved purely by selecting an informative subset?
  • EN Key Points:
    • arXiv:2607.02637v1 Announce Type: new
    • Abstract: Recent generative models can produce high-quality synthetic images, offering scalable training training data for data-hungry models
    • Existing approaches to exploiting this potential typically involve 1) training or fine-tuning generators, or 2) using lightweight post-hoc adaptation like promp…
    • We study a complementary question: given a fixed pool of generated images, can downstream utility be improved purely by selecting an informative subset

• A Granularity-Aware EEG Feature Framework for Psychopathology Dimension Prediction

  • Publication Time: 2026-07-07 12:00 Beijing Time
  • Abstract: - arXiv:2607.02670v1 Announce Type: New.
    • Abstract: Electroencephalography (EEG) offers a non-invasive approach to examining neurophysiological correlates of dimensional psychopathology, but systematic evidence across EEG paradigms and feature granularities remains limited.
    • Here, we develop a granularity-aware EEG feature pipeline that organizes multi-scale descriptors into global, regional, and channel levels.
    • Using the Healthy Brain Network (HBN) cohort, we evaluated the prediction of four psychopathology dimensions: p-factor, internalizing, externalizing, and attention problems, across four EEG paradigms.
  • EN Key Points:
    • arXiv:2607.02670v1 Announce Type: new
    • Abstract: Electroencephalography (EEG) offers a noninvasive approach for examining neurophysiological correlates of dimensional psychopathology, yet systematic…
    • Here, we develop a granularity-aware EEG feature pipeline that organizes multi-scale descriptors into global, regional, and channel levels
    • Using the Healthy Brain Network (HBN) cohort, we evaluate the prediction of four psychopathology dimensions: p-factor, internalizing, externalizing, and attenti…

• LiNO: Lifting based multiresolution neural operator

  • Publication Time: 2026-07-07 12:00 Beijing Time
  • Abstract: - arXiv:2607.02715v1 Announcement Type: New.
    • Abstract: Recently, neural operators have shown promising results in learning the solution operators of differential equations directly from data.
    • This framework learns a functional mapping from the parameter field to the solution field, enabling the prediction of an entire class of solutions rather than a specific instance.
    • However, existing operators often struggle to capture both global dynamics and fine-scale structures simultaneously.
  • EN Highlights:
    • arXiv:2607.02715v1 Announce Type: new
    • Abstract: Recently, neural operators have shown promising outcomes for learning solution operators of differential equations directly from data
    • This framework learns a functional mapping from the parameter field to the solution field, enabling the prediction of an entire class of solutions rather than a…
    • However, existing operators often struggle to capture both global dynamics and fine-scale structure simultaneously

• Weighted Conformal Prediction for Lab-to-Track Thermal Transfer in EV Motorsport Powertrains

  • Publication Time: 2026-07-07 12:00 Beijing Time
  • Abstract: - arXiv:2607.02722v1 Announcement Type: New.
    • Abstract: Predicting thermal volatility in high-performance electric vehicle powertrains is difficult, as internal temperatures are rarely observed outside the lab, and models calibrated on lab driving cycles fail when deployed against real-world loads.
    • We use conformal prediction to study this lab-to-track transfer problem, providing distribution-free uncertainty bounds.
    • We implement Ensemble Batch Prediction Intervals (EnbPI; Xu & Xie, 2021), a leave-one-out bootstrap-ensemble conformal method for autocorrelated time series, and calibrate it against real CALCE lithium-ion cycler data (A123 SP20 battery, FUDS profile).
  • EN Highlights:
    • arXiv:2607.02722v1 Announce Type: new
    • Abstract: Predicting thermal volatility in high-performance EV powertrains is difficult as internal temperatures are rarely observable outside the lab, and mode…
    • We study this lab-to-track transfer problem using conformal prediction, offering distribution-free uncertainty bounds
    • We implement Ensemble Batch Prediction Intervals (EnbPI; Xu & Xie, 2021), a leave-one-out bootstrap-ensemble conformal method for autocorrelated time series, an…

• Out-of-Distribution Generalization of Risk Aversion in Language Models

  • Publication Time: 2026-07-07 12:00 Beijing Time
  • Abstract: - arXiv:2607.02755v1 Announcement Type: New.
    • Abstract: Training artificial intelligence to be risk-averse in terms of resources can provide a fail-safe when the AI goes off-course.

• Misaligned but risk-averse AIs tend to prefer low-risk, low-reward strategies like cooperation over high-risk, high-reward strategies like rebellion, thereby limiting any negative impact of misalignment.

• However, we can only feasibly train AIs to be risk-averse in low-stakes gambles, and we will only be safe if their risk aversion generalizes to astronomically high-stakes gambles.

  • EN Key Points:
    • arXiv:2607.02755v1 Announce Type: new
    • Abstract: Training AIs to be risk-averse in resources could offer a failsafe in the event that AIs turn out misaligned
    • Misaligned but risk-averse AIs would tend to prefer low-risk, low-reward strategies like cooperation over high-risk, high-reward strategies like rebellion, limi…
    • But we can only feasibly train AIs to be risk-averse on low-stakes gambles, and we will only be safe if their risk aversion generalizes to astronomically-high-s…