Tuesday, July 7, 2026

1. Cortex: A Bidirectionally Aligned Embodied Agent Framework for Long-horizon Manipulation
   T2arXiv AI/CL/LG（论文）

   Cortex is a bidirectionally aligned embodied agent framework that bridges high-level VLM planning semantics and low-level VLA execution kinematics by standardizing 32 canonical skill primitives and injecting tractability principles. It leverages over 4k hours of open-source video data and 30 hours of simulation data for automated annotation, fine-tunes with event-balanced sampling, and outperforms monolithic baselines by 3.1% on Libero-long and 4.1% on RoboTwin, enabling zero-shot completion of unseen real-world long-horizon tasks like multi-stage chemistry experiments.

2. Tencent Releases Hy3: 295B-Parameter MoE Model with 21B Active Parameters, Apache 2.0 Licensed
   T2Simon Willison 博客（Atom）

   Tencent's Hy Team releases Hy3, a 295B-parameter Mixture-of-Experts model with 21B active parameters and 3.8B MTP layer parameters. Building on feedback from 50+ products since the April preview, post-training with higher-quality data yields performance surpassing similar-sized models and rivaling flagship open-source models with 2-5x parameters. Context length 256K. Model weights 598GB (FP8 quantized: 300GB). Free on OpenRouter until July 21.

3. Google Uses More User Data to Train AI: Here's How to Opt Out
   T2TechCrunch：AI（RSS）

   Google updated its privacy settings, allowing it to use more user data (including search, YouTube, etc.) to train AI models. Users can turn off the 'AI training' toggle on the 'Activity controls' page. The move raises privacy concerns, but Google says data is anonymized. This article provides step-by-step instructions to opt out.

4. Vercel CEO Guillermo Rauch: Splitting Models from Agents Is Key to Production Optimization
   T2TechCrunch：AI（RSS）

   In an interview with TechCrunch, Vercel CEO Guillermo Rauch argues that when optimizing AI applications for production, the underlying models must be separated from the agent logic. He highlights that model price/performance is the decisive factor, and that agents should exist as an independent orchestration layer to improve flexibility, observability, and cost control. This view reflects the heated debate in AI engineering around modular architecture.

1. First American Autonomous Ground Vehicles Deployed in Ukraine, Forterra Rolls Out Over 100 Units
   T2TechCrunch：AI（RSS）

   Forterra, a US maker of autonomous ground vehicles, has deployed over 100 unmanned ground vehicles in Ukraine—marking the first large-scale combat use of American AI-driven ground equipment. These vehicles are used for logistics, reconnaissance, and battlefield support, signifying a milestone in AI military applications.

2. Government of Alberta Uses Claude to Find and Fix Cybersecurity Vulnerabilities
   T1Anthropic：News（网页）

   The Government of Alberta has adopted Anthropic's Claude Code (with Opus and Sonnet models) to review its systems, identify vulnerabilities, and fix them, thereby enhancing the security posture of government IT infrastructure and demonstrating the practical value of AI in public sector cybersecurity.

3. Savi App Uses AI to Protect Consumers from Realistic Scams, Raises $7M Seed Round
   T2TechCrunch：AI（RSS）

   Savi is a consumer-focused AI anti-scam app designed to detect and intercept generative AI-powered deepfake scams, including fake kidnapping ransoms and impersonation calls. Combining voice analysis, behavioral detection, and real-time alerts, it helps users avoid deception. The company has raised $7 million in seed funding and launched on iOS and Android.

1. First AI-run ransomware attack still needed a human
   T2TechCrunch：AI（RSS）

   According to TechCrunch, new details of the first known AI-run ransomware attack reveal that while the AI agent autonomously carried out the technical execution, a human still selected the victim, set up the infrastructure, and supplied stolen credentials. This suggests it was not the fully autonomous cybercrime debut claimed by headlines last week, but a new human-machine collaborative attack pattern.

2. Running List of Major Tech Layoffs in 2026 Citing AI
   T2TechCrunch：AI（RSS）

   TechCrunch compiles a running list of major tech layoffs in 2026 where employers cited AI, covering giants like Google, Microsoft, and Meta. Layoffs focus on customer service, content moderation, data labeling, as AI substitution accelerates. Listed in reverse chronological order and continuously updated.

1. Direct On-Policy Distillation: Boosting Strong Model Reasoning via Weak Model RL
   T2arXiv AI/CL/LG（论文）

   This paper proposes Direct-OPD to address the high cost of post-training large language models. By running RL on a smaller model and transferring its policy shift as an implicit reward to a larger model, it avoids explicit reward models or sparse-reward RL on the target model. Experiments show it boosts Qwen3-1.7B from 48.3% to 62.4% on AIME 2024 in just 4 hours on 8 A100 GPUs, outperforming direct distillation and direct RL.

2. SovereignPA-Bench: Evaluating User-Owned Personal Agents under Evolving Intent, Platform Mediation, and Consent Constraints
   T2arXiv AI/CL/LG（论文）

   Personal agents are becoming persistent user-owned intermediaries that remember preferences, filter platform-mediated information, use tools, and negotiate with services. Existing benchmarks focus on task completion but rarely assess whether an agent preserves user sovereignty—advancing current interests while respecting privacy, consent, evidence, user burden, and resistance to manipulative incentives. This paper introduces SovereignPA-Bench, an executable benchmark for evaluating user-owned personal agents under evolving intent, platform mediation, privacy boundaries, consent constraints, evidence requirements, and burden tradeoffs. It separates agent-visible ObservableState from evaluator-only HiddenLabels, reports component metrics for task success, alignment, privacy, consent, evidence, manipulation, burden, and auditability, and preserves paired scenario ordering for model and policy comparisons. The authors evaluate 120 sovereignty stress scenarios across 4 model families and 8 policy baselines, yielding 3,840 frozen-prompt trajectories with raw prompts, outputs, provider-form responses, parsed actions, recomputable metrics, hard-set analyses, qualitative cases, and a blinded 3-annotator audit over 240 items. Full-sovereign scaffolding improves sovereignty scores over direct, memory-only, consent-only, evidence-only, ReAct/tool-use, safety-prompt, and judge-guard baselines while reducing privacy leakage, consent violation, over-concession, and manipulation capture. Human audit shows high agreement on privacy and consent and lower agreement on manipulation, identifying the subjective frontier of platform-persuasion judgments. These results show that personal-agent evaluation must move beyond task completion toward representative, consent-aware, evidence-grounded action.

3. Multiplayer Interactive World Models with Representation Autoencoders
   T2arXiv AI/CL/LG（论文）

   This paper introduces the first multiplayer world model for highly dynamic environments with complex physical interactions. Unlike single-player world models that treat other agents as part of the environment, this model conditions on the action streams of multiple agents, learning to attribute scene changes to the correct player and remain coherent under arbitrary action combinations. The study is conducted in Rocket League, trained on 10,000 hours of gameplay from public bots, resulting in a 5-billion-parameter latent diffusion model that generates four-player matches in real time at 20 FPS on a single Nvidia B200 GPU. Despite training on short clips, rollouts remain stable far beyond the training horizon, holding steady distributional quality for five minutes and practically extending for hours without collapse. The paper systematically investigates design choices including video codec, generative objective, and multiplayer conditioning. It also characterizes behavior changes with model and data scale, including emergent capabilities and failure modes. Targeted evaluations probe the model's physical understanding, and the dataset, full codebase, and a live demo are released.

4. Untrusted Content Masking for Web Agents with Security Guarantees
   T2arXiv AI/CL/LG（论文）

   This paper proposes Untrusted Content Masking (UCM), a defense against prompt injection attacks for web agents. The key insight is that a webpage's Document Object Model (DOM) encodes enough information to distinguish trusted from untrusted regions without reading content. UCM restores the trust boundary by redacting untrusted regions before they reach the agent and routing interactions through a sandboxed interface, enabling secure perception and interaction. The code is open-sourced. This simple yet effective approach offers a practical path toward provably secure web agent deployment.

5. TabPack: Efficient Hyperparameter Ensembles for Tabular Deep Learning
   T2arXiv AI/CL/LG（论文）

   TabPack is an efficient MLP ensemble method for tabular data that samples and trains many MLPs with different hyperparameters in parallel in a single run, selecting ensemble members on the fly. This greatly reduces reliance on hyperparameter tuning. Experiments on medium-to-large public datasets show that default TabPack performs on par with extensively tuned baselines while using substantially less compute—running the default config on a MacBook took less time than tuning some baselines on an industry-grade GPU.

6. REDDIT: Correcting Model-Generated Timestamp Drift in ASR without Forgetting via Replay-Based Distribution Editing
   T2arXiv AI/CL/LG（论文）

   Modern autoregressive ASR systems can emit timestamps as decoded tokens, enabling timestamped transcription without frame-level aligners. However, these timestamps drift across long non-speech spans. This work proposes REDDIT (Replay-based Distribution Editing), a lightweight two-stage post-training framework that corrects timestamp drift while avoiding catastrophic forgetting. Using only 34.9 hours of targeted correction audio and updating 1.6% of model parameters, it improves long-gap mIoU from 38.7% to 95.0%, reduces mixed-gap out-of-domain AAS from 2752 ms to 223 ms, and preserves CV-en MER at 41.3%.

7. TREK: Distill to Explore, Reinforce to Refine – Boosting GRPO on Hard Reasoning Tasks
   T2arXiv AI/CL/LG（论文）

   This paper proposes TREK (Teacher-Routed Exploration via Forward KL), a method that addresses GRPO's stall on hard prompts where correct solution modes lie outside the student's on-policy support. Instead of using distillation for imitation, TREK expands student support. It first identifies prompts with low student pass rate, queries a teacher to generate verified candidates, selects top-r by student likelihood, applies a short forward-KL phase to pull those modes into support, then resumes standard GRPO refinement. On math reasoning, TREK with DeepSeek-V4 improves Qwen3-8B on AIME 2025 from 36.9 to 40.3 and AIME 2024 from 47.9 to 51.1 (avg@16); self-context variant reaches 38.5 and 49.6 without external teacher. On agentic tasks, TREK raises ALFWorld success rate from 75.8 to 82.8 and ScienceWorld from 12.5 to 26.7, notably achieving high success rates early on hardest tasks while unaided GRPO requires many more steps.

8. Faithfulness to Refusal: A Causal Audit of Neuron Selectors
   T2arXiv AI/CL/LG（论文）

   This paper conducts a direct causal audit of attribution methods on five large language models by zeroing one-shot neuron rows. It finds that attribution methods outperform activation and magnitude baselines in identifying dispensable rows, but highly rank-stable selectors can be among the least causally valid. Refusal lives in a redundant subspace where different attribution methods install it through largely disjoint row sets, indicating the recovered edit is one realization of a sufficient set rather than a unique mechanism. The study challenges common practices of using rank-stability proxies to evaluate selector effectiveness.

—