Wednesday, July 29, 2026

1. Schrödinger's Cat: Probabilistic Representation and Prediction of Potential Scene Kinematics
   T2arXiv AI/CL/LG（论文）

   The paper introduces GARFIELD, a probabilistic model that learns a structured spatio-temporal latent distribution over possible scene kinematics from an image and sparse constraints. It enables joint trajectory sampling and direct density estimation via a deterministic decoder, localizing uncertainty to specific scene elements and timesteps. Compared to large video generation models, it samples trajectories 97× faster and estimates motion densities two orders of magnitude faster, enabling interactive exploration and uncertainty-aware planning.

2. Discovering Cryptographic Weaknesses with Claude Mythos
   T2Simon Willison 博客（Atom）

   Anthropic researchers used the Claude Mythos model, guided by carefully crafted prompts (including spelling mistakes), to discover mathematical flaws in HAWK and a weakened version of AES. While these findings have no practical impact on current systems, the experiment demonstrates the potential of LLMs in advanced research. The model ran for 60 hours, costing approximately $100,000 in API usage, with the main human intervention being encouragement to "not give up" and "find something worth publishing." The paper and code are publicly available.

3. Sharpness-Aware Minimization and Muon: Robustness under the Spectral Norm
   T2arXiv AI/CL/LG（论文）

   This paper reexamines Sharpness-Aware Minimization (SAM) from a geometric perspective, introducing layerwise spectral inner perturbations for matrix-valued hidden-layer parameters and combining them with AdamW/SGDW or Muon for outer updates. Experiments on ImageNet-1K with ViT-Small/16 and ResNet-50 show that the combination of a spectral inner step with a Muon outer step consistently achieves the best validation accuracy, highlighting the importance of matrix-aware geometry in both stages of SAM. This work offers new insights into optimizer-generalization understanding and advances practical training methods.

1. MemLens: A Value-Aware Memory Management System with Interactive Analytics for LLM-based Agents
   T2arXiv AI/CL/LG（论文）

   MemLens introduces a value-aware memory management approach that treats heterogeneous interaction records as first-class data objects, employing Shapley-style evaluation for memory utility, hierarchical visualizations of storage structures, and comparative analysis of strategies (response quality, retrieval latency, token consumption). The system provides an end-to-end interactive dashboard, enabling researchers and developers to intuitively understand and optimize LLM-based agents' long-term memory, enhancing personalized reasoning and knowledge reuse. Experiments on a study-copilot application validate its effectiveness in reducing redundancy and improving response quality.

1. Scientific Computing in the Age of Agentic AI: OpenAI Releases New Field Report
   T1OpenAI：News（RSS）

   OpenAI has released a new field report on the application of agentic AI in scientific computing, showcasing how scientists use AI coding agents to modernize scientific computing, accelerating software development and discovery in genomics and beyond. The report reveals the potential of AI agents in code generation, experimental design, and data analysis, transforming scientific workflows.

2. Cyera Acquires Oasis Security for $1B to Secure Proliferating AI Agents
   T2TechCrunch：AI（RSS）

   Data security firm Cyera has agreed to acquire Oasis Security, an AI agent security startup, for $1 billion, addressing identity and access management challenges posed by the mass deployment of AI agents. This is Cyera's third acquisition this year, signaling accelerating consolidation in AI security. Oasis's technology automatically discovers, authenticates, and monitors AI agent behaviors to prevent unauthorized data access.

1. Pictura: Perspective-View Self-Play at Scale for Driving
   T2arXiv AI/CL/LG（论文）

   Pictura is a GPU-accelerated multi-agent driving simulator that sustains up to 500K agent-steps/s (2M images/s) on a single H100. Using this simulator, researchers trained Alberti, a driving policy, via plain PPO self-play directly from perspective images without any privileged observations (e.g., exact poses or velocities). The training spanned 50B agent steps (~35M km of driving). Alberti approaches the performance of privileged vectorized counterparts and zero-shot transfers to Waymo Open Motion Dataset layouts re-rendered in Pictura, even outperforming privileged vectorized agents. This work bridges the representation gap between simulation training and real-world deployment.

2. Face De-Identification: A Domain-Centric Survey from Capture to Processing
   T2arXiv AI/CL/LG（论文）

   This survey provides the first unified overview of the full data acquisition pipeline for face de-identification (De-ID), covering physical, sensor, and digital domains. Through a domain-centric lens, it systematically analyzes methodologies, technical progress, and distinct challenges at each stage, reviews existing evaluation protocols, and highlights the urgent need for standardized benchmarks. It identifies key open problems and emerging research directions. An accompanying project page curates relevant literature, datasets, and open-source code.

3. Toward Standardized Cross-Vendor Agent Tool Trust Management in Autonomous Networks
   T2arXiv AI/CL/LG（论文）

   To address the lack of trust visibility when AI agents invoke tools across vendors in L4-L5 autonomous networks, this paper proposes AgentToolMO, an information model based on 3GPP NRM. The model includes a formal trust state machine, damped cascade propagation, cross-vendor notification via MnS, and retroactive impact assessment via dependency graph traversal. Simulations show standardized notifications reduce blast radius from hours to near-real-time, with bounded cascade convergence and sub-linear notification scaling. The framework operates within existing 3GPP infrastructure, offering a standardization path for trustworthy multi-vendor autonomous network management.

4. Interactive Reward Agent: GUI Task Evaluation via Environment-State Verification
   T2arXiv AI/CL/LG（论文）

   This paper proposes an Interactive Reward Agent (IRA) based on a 'propose-then-verify' framework that acquires and verifies evidence from the post-execution environment by invoking system tools, application tools, and GUI tools to determine task completion. It also introduces GUI-RewardBench, a benchmark of 321 GUI task trajectories across 10 Ubuntu desktop applications. IRA achieves 86.9% accuracy on this benchmark and is applied to reinforcement learning for GUI agents, achieving a 34.0% OSWorld success rate.

5. Pass the Baton: Trajectory-Relayed On-Policy Distillation
   T2arXiv AI/CL/LG（论文）

   This paper proposes Relay-OPD, an improved on-policy distillation method that addresses the prefix failure problem where the student model commits to wrong reasoning directions. By detecting failed prefixes and letting the teacher model briefly take over to produce a 'teacher leg', then resuming student generation for optimization, it effectively corrects erroneous trajectories. On mathematical reasoning benchmarks, the 1.7B model outperforms standard OPD by 5.73% on average, with training trajectory length reduced by over 50%.

6. πR²: Reactive Real-time Flow Policies — Enabling High-Frequency Closed-Loop Replanning for Generalist Manipulation Policies
   T2arXiv AI/CL/LG（论文）

   Generalist manipulation policies often take the form of action-chunking flow policies built on large pretrained backbones, but they execute open-loop and cannot react to mid-execution sensory input, with replanning limited by latency. This paper proposes πR², leveraging the per-position noise schedule of diffusion forcing to split conditioning into a fast channel (proprioception) and an asynchronously updated slow channel (vision-language features), plus a latency-adaptive flow schedule. Applied to GR00T-N1.7 on an xArm6+XHand platform, πR² achieves roughly 4x faster closed-loop replanning (~25Hz), improving success rates by 23% in simulation and 30% in the real world over the strongest baseline.

7. Desktop-Delta Bench: Do Computer-Use Models Understand Desktop GUI Transitions?
   T2arXiv AI/CL/LG（论文）

   This paper introduces Desktop-Delta Bench (DDB), an offline step-level benchmark with 2,013 human-verified instances across ~15 applications and 50 task domains. DDB evaluates computer-use models' ability to understand GUI state transitions through two tasks: temporal ordering and before-after comparison, filling a diagnostic gap in existing benchmarks. Evaluating 8 model families reveals consistent gaps: best exact-match rates for ordering tasks are around 65%, and inferring action families is harder than localizing them. DDB enables targeted improvements for verification, reliability, and recovery of desktop CUAs.

8. Parallel Decoding Distillation for Fast Image and Video Generation
   T2arXiv AI/CL/LG（论文）

   This paper introduces Parallel Decoding Distillation (PDD), a trajectory-based distillation method for fast inference of diffusion and flow matching models. PDD predicts multiple denoising steps per network evaluation, achieving SOTA performance with 4-8 NFE while avoiding mode collapse from traditional VSD and adversarial losses, improving video diversity. Validated on LTX-2.3, Wan 14B, and Qwen-Image models.

—