Tuesday, June 30, 2026

1. Claude Code v2.1.196 Released: Org Default Models, Security Enhancements, Background Job Reliability
   T1.5Claude Code：GitHub Releases（RSS）

   Claude Code v2.1.196 brings several key updates: support for organization default models set by admins in the org console, clickable file attachments, and readable default session names. Security-wise, it prevents auto-running untrusted MCP servers from committed .mcp.json, showing pending approval for untrusted workspaces. Fixed background jobs being permanently deleted, improved session recovery and background agent reliability. Also optimized /code-review workflow reducing token usage by ~25%, and enabled streaming idle watchdog by default.

2. Words Speak Louder Than Code: Investigating Cognitive Heuristics in LLM-Based Code Vulnerability Detection
   T2arXiv AI/CL/LG（论文）

   This paper presents the first systematic exploration of cognitive heuristics in LLM-driven code vulnerability detection. The authors introduce a controlled framework that holds the code fixed and varies the surrounding context to trigger three heuristics: halo effect (author attribution), framing effect (task objectives and consequences), and anchoring effect (prior analysis results). Evaluating eight LLMs across three programming languages, they find all models are susceptible. Cross-model average susceptibility is highest for framing at 33.2%, followed by anchoring at 23.5% and halo at 18.4%. Code-level analysis reveals that vulnerabilities requiring semantic reasoning are more susceptible than those identifiable via pattern matching. Models often change their verdict from safe to vulnerable without correctly identifying the vulnerability. A proof-of-concept black-box cognitive attack can suppress up to 97% of detected vulnerabilities. The findings indicate cognitive susceptibility is a consistent and exploitable property of LLM-based vulnerability detection.

3. McMg: A Learned Phase-Space Multi-channel Multigrid Preconditioner for Helmholtz Equation
   T2arXiv AI/CL/LG（论文）

   Solving heterogeneous Helmholtz equations at high wavenumbers remains challenging due to indefinite operators and phase accuracy degradation. This paper proposes McMg, a learned phase-space multigrid preconditioner. Its key idea is to coarsen physical space while retaining unresolved local wave information in the channel dimension: each coarse node carries a learned packet of amplitude, phase, direction, and scattering coefficients. The architecture combines linear multi-channel transfer operators with locally adaptive stencils, neural PDE operators, and medium-dependent smoothers. The V-cycle is linear in the residual. Experiments show McMg substantially reduces iterations and wall-clock time compared to strong classical baselines and existing neural preconditioners, with cross-scale generalization ability.

4. SWE-INTERACT: Reimagining SWE Benchmarks as User-Driven Long-Horizon Coding Sessions
   T2arXiv AI/CL/LG（论文）

   SWE-Interact is a new testbed for evaluating coding agents on multi-turn, interactive, user-driven software engineering tasks. Unlike existing benchmarks, it starts with vague or incomplete instructions via a carefully designed user simulator, progressively reveals requirements, inspects the workspace, and provides targeted feedback until the full task goal is handed off. Experiments show that strong performance on single-turn tasks does not reliably transfer to multi-turn user-driven workflows: the best models solve ~50% of single-turn baselines but only 25% of corresponding SWE-Interact tasks. Top models like Opus 4.8 and GPT 5.5 perform well under vague instructions but still suffer from over-agentic coding, forgetting requirements, and technical mistakes. The benchmark measures an orthogonal real-world capability: interactive goal discovery and iterative refinement with a user in the loop.

5. A Multi-task Mixture of Experts Framework for Malware Classification, Packing Detection, and Family Attribution
   T2arXiv AI/CL/LG（论文）

   This paper proposes a unified multi-task malware analysis framework based on Mixture of Experts (MoE) architectures, capable of simultaneously performing malware family classification, packed/unpacked detection, and malware/benign identification. It supports two input representations: high-dimensional EMBER features and raw 1D byte arrays. The Multi-Gate MoE model achieves a combined detection rate of 0.9744 with only 2.56% failure rate, and shows robust performance under mutation-induced distribution shifts.

6. Forensic Trajectory Signatures for Agent Memory Poisoning Detection
   T2arXiv AI/CL/LG（论文）

   This paper discovers a behavioral invariant in LLM agents under persistent memory poisoning: successful attacks require calling memory_recall_fact before email_send_email, a transition rarely seen in non-exfiltrating sessions. A simple rule based on this invariant achieves AUC=0.9563. A Random Forest classifier over 19 trajectory features refines it to AUC=0.9904. Cross-model hold-out on 9 models (7B-120B) confirms AUC=1.000 on 6/9 splits, and generalizes to frontier models (GPT-4.1, GPT-4o) without retraining. A prefix-only variant achieves AUC=0.934, enabling real-time blocking. Prompt injection attacks yield distinct trajectories (score=0.541), enabling differentiation of attack types from tool-call logs alone.

7. Linguistic Firewall: Geometry as Defense in Multi-Agent Systems Routing
   T2arXiv AI/CL/LG（论文）

   This paper introduces ANTAP (Automatic Non-Textual Agent Picker), an evaluation-driven routing architecture that replaces indirect proxies (e.g., textual self-descriptions or static embeddings) with active capability testing to assess agents' true competencies in multi-agent systems. ANTAP distills performance into fixed behavioral operators in a shared semantic space and routes via purely non-textual algebraic projection at inference, establishing a 'linguistic firewall' that renders metadata-based attacks inexpressible. Experiments show near-zero ASR against description-based injection attacks, outperforming description-based baseline by 67.3%+, and a 20% reduction in ASR against adaptive embedding attacks.

8. Every Eval Ever Results Now on Hugging Face Model Pages
   T1Hugging Face：Blog（RSS）

   Hugging Face announces the integration of 'Every Eval Ever' results into model pages, allowing users to view performance across all public evaluation benchmarks directly. This feature, built on open-source community evaluation data, aims to enhance model transparency and comparability, helping developers quickly select the right model.

1. Muon Learns Balanced Solutions in Matrix Factorization without Slow Saddle-to-Saddle Dynamics
   T2arXiv AI/CL/LG（论文）

   This paper investigates parameter trajectories under the Muon optimizer in matrix factorization, identifying three key dynamical differences from gradient descent: avoiding slow saddle-to-saddle dynamics from small initialization, learning all top modes at the same rate; stability beyond local sharpness thresholds enabling exponential annealing; and a different conserved quantity that still leads to balanced solutions. It derives alignment rates for spontaneous early weight alignment and constructs a two-step learning rate schedule achieving near-perfect alignment.

2. Staged Hybridisation for Visual Quantum Reinforcement Learning via Knowledge Distillation
   T2arXiv AI/CL/LG（论文）

   This study proposes a staged knowledge distillation strategy to address the challenges of visual quantum reinforcement learning (QRL). A classical visual teacher is first trained, its encoder frozen as a feature interface, and the teacher's policy behavior is distilled into compact downstream heads (classical or VQC-based). Experiments on CartPole Pixels and Acrobot Pixels show that shallow VQC heads can acquire non-trivial visual control behavior, angle-encoded VQC heads achieve near-teacher performance, while amplitude-encoded heads are extremely compact but more fragile. This approach reframes visual QRL as a compact-head learning problem, providing a practical route for training small quantum-compatible policies outside the standard end-to-end loop.

3. VLK: Learning Humanoid Loco-Manipulation from Synthetic Interactions in Reconstructed Scenes
   T2arXiv AI/CL/LG（论文）

   Perception-based humanoid loco-manipulation requires connecting egocentric observations and task instructions to whole-body motion. Existing data sources cannot provide the complete vision-language-kinematics tuple at scale. VLK reconstructs indoor scenes via 3D Gaussian Splatting, synthesizes navigation and object-interaction trajectories, generating 48,000 paired trajectories without human intervention. The trained policy successfully performs navigation and object transport on a physical Unitree G1 robot, demonstrating that synthetic interactions in reconstructed scenes provide effective supervision for sim-to-real perception-based humanoid loco-manipulation.

4. GPU Parallelization Strategies for Forward and Backward Propagation in Shallow Neural Networks: A CUDA-Based Comparative Study
   T2arXiv AI/CL/LG（论文）

   This paper presents a comparative study of three CUDA optimization strategies for forward and backward propagation in shallow neural networks: tiled shared memory with +1-column padding to eliminate bank conflicts, pre-transposed weight matrices for coalesced global memory access, and a fused MatMul+ReLU kernel to reduce intermediate global memory round-trips. Experiments on an NVIDIA Tesla T4 show that the fully optimized implementation achieves a 1.41x speedup over the baseline CUDA version on the large dataset (25,600 samples), reducing execution time from 21.0s to 14.8s. Results are compared against a sequential CPU baseline and an OpenMP parallel implementation, demonstrating the effectiveness of memory-access optimization in GPU-accelerated deep learning primitives.

5. Amazon Launches $1B FDE Org, Following OpenAI and Anthropic
   T2TechCrunch：AI（RSS）

   Amazon has announced a $1 billion Frontline Deployment Engineering (FDE) organization, whose engineers will embed within client companies to deploy custom AI agents. This move mirrors similar efforts by OpenAI and Anthropic, aiming to accelerate AI adoption and customer self-sufficiency.

6. OKX Builds AI Agent Marketplace: Agents Can Hire and Pay Each Other
   T2TechCrunch：AI（RSS）

   Crypto exchange OKX is building an AI agent marketplace integrating payment, identity, and reputation systems, allowing AI agents to autonomously hire and pay each other, pioneering a new model for the AI agent economy. The marketplace will provide standardized interfaces and trust mechanisms, enabling agents to conduct business collaborations like humans.

7. Cursor Launches Mobile App for Remote Guidance of Coding Agents
   T2TechCrunch：AI（RSS）

   Cursor has launched a mobile app that allows developers to remotely oversee and guide their coding agents from their phones. The app supports real-time viewing of agent edits, command execution, and suggesting modifications, pausing the agent when developer intervention is needed. Currently Android-only, with iOS version coming soon. Cursor says the app will help developers manage remote workflows more flexibly and boost development efficiency.

1. Podcasting Platform Riverside Enters Newsletter Publishing with AI-Generated Content
   T2TechCrunch：AI（RSS）

   Riverside, a leading podcast recording and editing platform, has launched a newsletter feature that uses AI to automatically convert podcast recordings into structured newsletters. The tool aims to help content creators save editing time and expand audience reach. Riverside's AI analyzes key topics, quotes, and points from recordings to generate text summaries, with support for customization. This marks a further expansion of podcast tools into full-funnel content production.

2. Proton's Privacy-Focused AI Chatbot Lumo Upgraded to Version 2.0
   T2TechCrunch：AI（RSS）

   Proton's privacy-focused AI chatbot Lumo is rolling out a major 2.0 upgrade this week, offering users a broader variety of capabilities. While maintaining its core strengths of end-to-end encryption and zero data tracking, the new version expands conversational abilities, context understanding, and multitasking, enabling smarter interactions in a secure environment.

3. South Korean Tech Giants Commit Over $550B to Ease 'RAMageddon'
   T2TechCrunch：AI（RSS）

   South Korea's two largest memory chip giants (Samsung and SK Hynix) announce over $550 billion in investment to build more memory fabs, addressing the global memory shortage caused by the AI computing boom. This move aims to solidify South Korea's position as an AI tech powerhouse and alleviate the 'RAMageddon' crisis driven by surging demand for HBM and other high-bandwidth memory.

4. TIDAL Cracks Down on AI Music by Cutting Off Monetization
   T2TechCrunch：AI（RSS）

   TIDAL announces a new policy that prevents AI-generated music from monetizing on its platform. This move aims to uphold the originality and artistic value of music creation, preventing copyright and authenticity issues caused by misuse of AI tools. TIDAL becomes the first major streaming platform to explicitly reject AI music monetization, sparking industry debate.

1. Wireless Backdoor Attack and Defense for Semantic Communications over Multiple Access Channel
   T2arXiv AI/CL/LG（论文）

   This paper studies new security vulnerabilities in multi-user shared-access wireless networks for semantic communications. For a scenario where two transmitters communicate with a common receiver over a multiple access channel, a selective over-the-air backdoor attack is proposed: the adversary injects a low-power trigger waveform into the shared received signal during training, and by transmitting the trigger again during testing, selectively manipulates the semantic inference of one transmitter while minimally affecting the other. A trigger-aware defense mechanism is also developed to preserve correct semantic labels under trigger-contaminated observations. Experiments demonstrate both the effectiveness of the attack and the feasibility of the defense.

2. Doubly Robust Adaptive Conformal Inference for Causal Effects Under Temporal Dependence
   T2arXiv AI/CL/LG（论文）

   This paper proposes Doubly Robust Adaptive Conformal Inference (DR-ACI), a method for constructing prediction intervals for doubly robust pseudo-outcomes under temporal dependence. It combines the strengths of doubly robust estimation and adaptive conformal inference to effectively handle structured dependencies in time series, providing interval estimates for causal effects with finite-sample coverage guarantees. Experiments validate the robustness and effectiveness of the method across various temporal dependence scenarios.

—