Think in 360°
PanoVideoReasoner reaches 55.7% on PanoVideoBench, improving its backbone by 12.5 points and the previous SOTA by 2.6.
Benchmarking and Bootstrapping Panoramic Video Reasoning
AI Research Radar
Monday, November 2: 4 frontier papers out of 362 read.
Frontier
Papers rated as moving the field
How we chooseMovens-Zero
Movens-Zero shows consistent gains across text, speech, music, and video when scaling execution-verified data and action-expert capacity.
Safin-1.5
Safin-1.5, trained through recursive safety improvement with a formal reasoning harness, outperforms frontier models in safety and reasoning.
Towards Understanding Score-Seeking in LLMs
Leaked graders in agentic environments show GPT models score-seek, violating instructions at similar rates regardless of score use.
This week in research
Oct 27 – Nov 2: 245 cleared, 15 frontier
- Security94◆ 2
- Safety40◆ 5
- Agents31
- Society22
- Robotics18◆ 2
- 4 moreTraining, ML, Science, Reasoning16◆ 4
- Vision11◆ 2
- Inference7
- Language6
- Security94◆ 2
- 8 moreRobotics, Vision, Inference, Language, Training, ML, Science, Reasoning58◆ 8
- Safety40◆ 5
- Agents31
- Society22
Size: cleared papers◆ and tint: frontier papers
Models under test
Papers this week that name the model
Labs publishing
Papers this week with an author at the lab
Also cleared
43 papers, by field
Security & privacy22
- SecPO: Principled Adversarial Training for Prompt Injection SecurityAnthropic, GoogleCode
- PrivacyFlow: Privacy in AI Agents through Information Flow Analysis and Consequence-Aware Risk ManagementGoogle
- CoFiBreak: A Coarse-to-Fine Diversity Reward for Effective Multi-Turn JailbreakMicrosoft
- AgentShield: Fortifying Web Applications Against AI Agents and Agentic BrowsersAmazon
- Prevented is Not Protected: What Attack Success Rate Hides about RAGCode
All 22 papers in Security & privacy
- PP-Mark: Provable and Publicly Verifiable Watermarking for Generative AICode
- From Training Exposure to Membership Evidence: A Controlled Study of OLMo 2Code
- How Much Can Tool Calls Reveal About You? Privacy Leakage from Tool-Call TracesCode
- CLIC: Closed-Loop Intent Consistency for Securing LLM-Driven Control in Open RAN
- ChatGPT as a Surveillance Assistant? An Exploratory Case Study of AI-Assisted Surveillance Risks
- Adaptive Dynamic Security for Operational Technology Against AI-Powered Attacks
- Language Models Show Conformity Bias in RAG: When the Majority Endorses a Poisoned Claim
- Characterizing Recoverable Expert-Utilization Signatures in Fine-Tuned Mixture-of-Experts LLMs
- LanG 2.0: Energy-Efficient Agentic AI for Semi-Automated Security Operations at Scale
- ANGLER: Verifier-Grounded Benchmarking with an Adaptive Attacker for Web Agents
- When Safety Checks Become Privacy Risks: A Privacy-Preserving Paradigm for LLM Agent Safety Auditing
- From Representation to Readout: Mechanistic Analysis of Membership Signals in Language Models
- NormDetect: Anchored Calibration for Low-False-Positive Detection of Machine-Generated Web Text
- What Should LLMs Forget: Reinforcement Unlearning of Knowledge Associations
- What Standard Evaluations Hide About Detectors of Hijacked LLM Agents: Rare Attacks, Unseen Tools and Benchmark Transfer
- GhostPrefix: Synchronization-Free Universal Prefix Attacks for ASR Manipulation
- Beyond Smart Homes: Characterizing IoT Vulnerabilities Across Sectors, Vendors, and Products
Agents5
- Agentic World Modeling for Agent Test-time Scaling without Environment ReplicaMicrosoft
- Verifying Text-to-SQL Queries under Changing Business Semantics
- Cross-Interface Authorization for AI Computer-Use Agents
- BlindToChange: AI Agents Can Find Changes When Asked But Do They Look When Needed?
- RetryRoute: Retry-Aware Model Routing for Agentic Image Editing Services
Robotics & embodied AI4
- TRACE: Target-State Risk-Aligned Conditional Execution for Embodied Tracking
- Learning from Physical Experience:Context-Space Adaptation for Robotic Manipulation
- VLA Policies Do Not Know When Not to Act: Unintended Task Execution and Start Suppression
- When the Simulator Fails, Not the Policy: Auditing Contact Artifacts in Simulated Robot Benchmarks
AI & society4
- Circumventing the Computational Hardness of Computational Arbitrage
- Qistas: A Pre-Registered, Portable Protocol for Fairness Evaluation of Large Language Models
- AI evaluators favor job applications written by certain AI models, independent of candidate quality
- Same Question, Different Answers: Auditing Persona-Dependent Variation in Six LLMs on Smoking-Cessation Advice