S-EMBER
A Large-Scale Benchmark for Streaming Egocentric Memory Retrieval
AI Research Radar
July 4–10: 10 frontier papers out of 1,665 read.
Frontier
Papers rated as moving the field
How we chooseOyster-II
Auto
“The same stream also quantifies the failure modes: a loose guard silently mislabels 48.9% of compiled answers, and an unfaithful deopt reference causes the verification gate to refuse recompilation.”
Governed Individuation
“An adversarial evaluation of monitors of increasing semantic depth shows false-allows falling from 75% (name-based gating) to zero (dynamic effect tracing), and refusal history transfers compliance to held-out red-line families.”
Your Agent's Memories Are Not Its Own
“Our evaluation also shows that SENTINEL reduces FARMA's attack success rate to as low as 0% with no false positives observed across 326 benign agent traces.”
Does AI Understand Imaging? A Systematic Benchmark of Agentic AI for Computational Imaging Tasks
WebRetriever
“Extensive experimental analysis reveals substantial performance disparities across evaluation protocols, demonstrating that navigation success alone is an insufficient predictor of real-world application effectiveness.”
RoboDojo
Native Video-Action Pretraining for Generalizable Robot Control
More Convincing, Not More Correct
“We show it fails structurally: conditioned on a candidate, a judge scores plausibility, not correctness, leaving false-positive basins a policy learns to exploit.”
The week by field
Jul 4–10: 128 cleared, 10 frontier
The week’s papers have not been sorted into fields yet.
Models under test
Papers this week that name the model
Labs publishing
Papers this week with an author at the lab
Also cleared
118 papers
- Pluralis v0.1: Towards a Multicultural, Multimodal, Multilingual Benchmark for AI Risk and ReliabilityGoogle, Microsoft
- CTA-Pipelining: A Latency-Oriented Spatial Scaling Method for Multi-GPU SystemsNVIDIA
- Persuasion Attacks Can Decrease Effectiveness of CoT MonitoringGoogle
- Modular Pretraining Enables Access ControlAnthropic
- Progressive Crystallization: Turning Agent Exploration into Deterministic, Lower-Cost Workflows in ProductionMicrosoft
- The Blind Curator: How a Biased Judge Silently Disables Skill Retirement in Self-Evolving AgentsAmazon
- MedPMC: A Systematic Framework for Scaling High-Fidelity Medical Multimodal Data for Foundation ModelsMicrosoft
- SEVRA-BENCH: Social Engineering of Vulnerabilities in Review AgentsMicrosoft
- AEGIS: A Mechanism-Guided Defense against Visual Synonym Jailbreaks in Text-to-Image ModelsAlibaba
- DREAMSTEER: Latent World Models Can Steer VLA Policies During Deployment Without Any FinetuningMeta
- Teaming Up with AI: Coordination and CooperationMicrosoft
- Securing Multi-Tool AI Agent Chains With Dynamic, Real-Time Compositional PoliciesMicrosoft
All 118 papers
- NKI-Agent: Domain-Specific Fine-Tuning and Agentic Tool Use for Neuron Kernel GenerationAmazon
- TrustCLIP: Learning Private Visual Features via Adversarial ReconstructionMeta
- DSpark: Confidence-Scheduled Speculative Decoding with Semi-Autoregressive GenerationDeepSeek
- Unified Audio Intelligence Without Regressing on Text IntelligenceNVIDIA
- aiAuthZ: Off-Host, Identity-Bound Authorization for AI AgentsCode
- SPORK: Self-Speculative Forking to Accelerate Agentic LLM InferenceCode
- G2VD: Generalizable AI-Generated Video Detection via Counterfactual Intervention and Causal DisentanglementCode
- When Do Foundation Models Pay Off? A Break-Even Analysis of Pretrained Time Series ForecastersCode
- Selective Disclosure Watermarking for Large Language ModelsCode
- Governing Generative AI Across Financial Institutions: An SR 26-2-Compatible Framework for Generative AI Risk Control
- Context Graphs for Proactive Enterprise Agents
- Controllability-Aware Adversarial Examples Against LLM-Based Network Traffic Classifiers
- Efficient Safety Alignment of Language Models via Latent Personality Traits
- When Debiasing Backfires: Counterintuitive Side Effects of Preprocessing-Based Stereotype Mitigation
- DeepSWE: Measuring Frontier Coding Agents on Original, Long-Horizon Engineering Tasks
- Beware What You Autocomplete: Forensic Attribution of Backdoored Code Completions
- Closing the Null Space: Guidance-Aware Quantization for Classifier-Free Diffusion
- How Analysts Use AI in High-Stakes Crime Linkage: An Industrial Study
- Multi-Agent Firewall Architecture for Privacy Protection of Sensitive Data in Interactions with Language Models
- TRACE: A Two-Channel Robust Attribution Watermark via Complementary Embeddings for LLM-Agent Trajectories
- When Synthetic Speech Is All You Have: Better Call GRPO
- Harness VLA: Steering Frozen VLAs into Reliable Manipulation Primitives via Memory-Guided Agents
- When the Judge Changes, So Does the Measurement: Auditing LLM-as-Judge Reliability
- Can Trustless Agents Be Trusted? An Empirical Study of the ERC-8004 Decentralized AI Agent Ecosystem
- JuZhou 1.0 Technical Report: The First Edge-Native Text-to-Image Foundation Model Trained Entirely on China-Developed AI Accelerators
- When Agents Remember Too Much: Memory Poisoning Attacks on Large Language Model Agents
- Calibration-Family Overfit: Why Trusted Sabotage Monitors Don't Transfer Across Lineages
- Inertia-1: An Open Exploration of Wearable Motion Foundation Models
- The Power of Backdoor Absorption in Community Training
- ORAN-DEFEND: Subspace Detection and Sanitization of Backdoor DRL xApps in Open RAN
- POPS: Recovering Unlearned Multi-Modality Knowledge in MLLMs with Prompt-Optimized Parameter Shaking
- CILC: Cryptographically-secure Inter-agent Loop Closure Candidate Detection for Multi-Agent Collaborative SLAM
- When Agents Go Rogue: Activation-Based Detection of Malicious Behaviors in Multi-Agent Systems
- Behavioral Privacy Leakage in Agentic Negotiation: Formalizing and Mitigating Inference Attacks via Randomized Policies
- LEMUR 2: Unlocking Neural Network Diversity for AI
- LLMs Silently Correct African American English: Auditing and Mitigating Dialect Bias via Activation Steering
- The Harness Effect: How Orchestration Design Sets the Token Economics of Enterprise Agentic AI
- A knowledge-augmented dataset of high-risk driving scenarios with LLM annotations for autonomous driving
- Certifying Ghosts: How Cybersecurity AI Agents Break the EU Cyber Resilience Act
- InfraQR: Edge-Placed QR-Inspired Structured Patch Attacks on Infrared Vision-Language Models
- The AI Resilience Gap: Bringing Artificial Intelligence Inside the Operational Resilience Perimeter
- Beware of Agentic Botnets: Scalable Untargeted Promptware Attacks via Universal and Transferable Adversarial HalluSquatting
- Mitigating Taint-Style Vulnerabilities in MCP Servers via Security-Aware Tool Descriptions
- Beyond Attack-Success Rate: Action-Graded Severity Scale for Tool-Using AI Agents
- Infinite Worlds with Versatile Interactions
- SkillCenter: A Large-Scale Source-Grounded Skill Library for Autonomous AI Agents
- Institutional Red-Teaming: Deployment Rules, Not Just Models, Causally Shape Multi-Agent AI Safety
- When Lower Privileges Suffice: Investigating Over-Privileged Tool Selection in LLM Agents
- The Inattentional Gap: Task-Conditioned Language and Vision Models Omit the Safety-Critical Signals They Can Otherwise Report
- Anti-Prompt: Image Protection against Text-Guided Image-to-Video Generation
- Proof of Execution: Runtime Verification for Governed AI Agent Actions
- The Jagged Global Economy: Frontier AI Unevenly Exposes National Economies
- KAT-Coder-V2.5 Technical Report
- Privilege and confidentiality in generative AI workflows
- Beyond Heuristic Tuning: Power-Calibrated LLM Watermarking
- Akashic: A Low-Overhead LLM Inference Service with MemAttention
- The Balkanization of Execution-Security Research for AI Coding Agents: Isolation, Access Control, and Time-of-Check-to-Time-of-Use Vulnerabilities
- Harrison.Rad 1.5 Technical Report: A radiology foundation model that can draft reports from images, priors and clinical context
- InfluMatch: Frontier-Quality KOL Search at 4B-Model Cost
- Context-to-Execution Integrity for LLM Agents
- Multi-Channel Spread-Spectrum Code Watermarking
- DT-Guard: Intent-Driven Reasoning-Active Training for Reasoning-Free LLM Safety Guardrail
- From Foundation to Application: Improving VLA Models in Practice
- TILDE: TILt-based Distributional Erasure for Concept Unlearning
- Andha-Dhun: A First Look at Audio Descriptions in Hindi
- AirflowAttack: Thermal-Airflow Adversarial Perturbations against Infrared Remote-Sensing Vision-Language Models
- Open-Weather Robust 3D Detection via Dual-Critic Diffusion Alignment
- Hierarchical Anti-Aesthetics: Protecting Facial Privacy against Customized Diffusion Models
- kNNGuard: Turning LLM Hidden Activations into a Training-Free Configurable Guardrail
- Cloak and Detonate: Scanner Evasion and Dynamic Detection of Agent Skill Malware
- Auditing the Audit: Five Failure Modes in Benchmark-Validity Audits
- Not All Refusals Are Equal: How Safety Alignment Fails Cybersecurity at Scale
- Safe Inference-Time Alignment via Lagrangian Reward Augmentation
- Vision Token Manipulation Attacks on Cloud-Edge Inference of Large Vision-Language Models
- MOSAIC: Knowledge-Guided CLI Command Composition Attack in LLM Coding Agents
- PromptPET: Privacy-Utility Optimized Prompt Obfuscation
- Overloading Large Vision-Language Models for Jailbreaking
- VISTA: Auditing Semantic Divergence in Vision-Language Models
- FairFlow: Demystifying and Mitigating Stereotype Bias in Text-to-Image Diffusion Transformers
- CONTRA: Red-Teaming Configurations of Personalizable Agents
- GDPR-Aware Trajectory Sharing for ISAC-Assisted Robot Navigation: A Case Study on FID-Constrained Collision Prediction
- Defending from GeoLocalization through Adversarial Road Trips
- Regulatory compliance-readiness in the AI Supply Chain: examining datasets in Hugging Face
- Brand-as-Memory: Vision-Language Models Encode Causal, Mechanistically Localizable Credibility Priors for News Sources
- DETECT-3B-Omni is Agnostic of Content and Demographics
- CAGE-1: Control, Assurance, and Governance Evaluation for Enterprise Agentic AI
- AGL-1: The Enterprise AI Governance Layer as a Control Plane for Trusted Enterprise Intelligence
- XPlainVerse: A Million-Scale Benchmark for Explainable Deepfake Detection
- AutoCedar: An Agentic Framework for Verifier-Guided Access Control Policy Synthesis
- PathMark: Protecting Intellectual Property of Mixture-of-Expert LLMs via Path Watermarks
- SABLE: An NDA-Safe Closed-Loop LLM Framework for Analog Circuit Optimization in Industrial EDA Flows
- Folding, Reasoning, and Scaling with Open-source Drug Discovery Engine
- DualView: Preventing Indirect Prompt Injection in Personal AI Agents
- "AI Slop is DDoSing Open Source": Understanding the Impact of AI-Generated Contributions on Open Source Sustainability
- Target-Aware Interaction-Guided Reinforcement Learning for Black-Box Node Injection Attacks on Graph Neural Networks
- !Imperio, smolVLA: The Implications of Data Poisoning on Open Source Robotics
- One Framework for All: Cross-Modal Membership Inference for Generative Models
- Nemotron-Labs-3-Puzzle-75B-A9B: Compressing Hybrid MoE LLMs
- Knowledge Base Poisoning Attacks and Defense for Policy-Aware LLM-RAG Framework
- ACE-Brain-0.5: A Unified Embodied Foundational Model for Physical Agentic AI
- Retroactive Chain-of-Thought (RetroCoT): Forensic Reconstruction Prompts as a Safety Diagnostic Across Model Generations
- From Multiplicity to Vulnerability: Privacy Amplification Risk from One-Dataset-Multiple-Model Exposure
- Agent Data Injection Attacks are Realistic Threats to AI Agents
- When Agents Lie: Premeditation, Persistence, and Exploitation in Repeated Games
- When Claws Remember but Do Not Tell: Stealthy Memory Injection in Persistent Personal Agents
- Privacy-Preserving Robustness Verification for Neural Networks