Infrathon
Claude Opus 5.5 raises the mean KV-cache compression ratio from 6.82% with the human-designed Quant-VideoGen codec to 19.86%.
Benchmarking Language Agents for Real-World, Long-Horizon AI Infra Optimization
AI Research Radar
Tuesday, November 10: 7 frontier papers out of 843 read.
Frontier
Papers rated as moving the field
How we chooseWhen Experience Becomes a Backdoor
BadEvolve demonstrates effective task-level backdoor attacks on self-evolving LLM agents across diverse frameworks, benchmarks, and backbone LLMs.
Mitigating Agentic Deception via Evaluation-Side Partially Observable Games
e-PoG outperforms the strongest baselines by 26.0% in deception mitigation and by up to 39.8% in task success.
Jais 3
Jais 3 matches or outperforms open-weight models of comparable size on Arabic and agentic benchmarks with fewer hallucinations.
Explaining and interrupting collective misalignment in language model systems
Collective Dependence Analysis reduces violations from 93% to 6% in controlled cases using O(N) model evaluations.
Compass 1
Compass 1 achieves a 79.4% overall success rate in open-world embodied navigation, versus 17.8% for the strongest baseline.
Shortcuts and Boundaries
A binary compliance penalty raises safe completion on 32 unseen changing-evidence tasks from 53.1% to 89.1% while both policies achieve full real work.
This week in research
Nov 4–10: 282 cleared, 17 frontier
- Security93◆ 4
- Agents42◆ 6
- Safety38◆ 4
- Society27
- Inference26
- Robotics22◆ 3
- Vision14
- 4 moreScience, Training, Reasoning, ML14
- Language6
- Security93◆ 4
- 8 moreInference, Robotics, Vision, Language, Science, Training, Reasoning, ML82◆ 3
- Agents42◆ 6
- Safety38◆ 4
- Society27
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
112 papers, by field
Security & privacy42
- MetaPI: Constructing Prompt Injection Benchmarks from LLM Agent BenchmarkNVIDIA
- zkARc: Succinctly Verifiable Proofs for Neurosymbolic Natural Language Formalization and VerificationAmazon
- Scaling Adaptive Prompt Injection Attacks with Agent SwarmsGoogle
- XposedMIA: Exploiting Explainable Artificial Intelligence (XAI) for One-Shot Membership Inference AttacksIBM
- DiffusedSplitFed: Latent Diffusion-based Obfuscation for Reconstruction-Resilient Split Federated Medical Image SegmentationCode
All 42 papers in Security & privacy
- BAPLe++: Cross-Model Transferability and Prompt-Level Defense of Backdoors in Medical Foundation ModelsCode
- Beyond Bit Matching: Orthogonal Watermarks for Collusion-Resistant Image FingerprintingCode
- MUTE: Multi-Level Alignment Uncoupling Against Talking-Head Exploitation for Voice ProtectionCode
- SecureGARD: Dynamic Ontology-Guided Governance for Secure Tool-Using LLM Agents
- Toward Safety-Aware AI Agents in Realistic Social Engineering Attack Scenarios
- Governed Self-Improvement for a Local-First Security Analyzer: Baseline Measurements and a Preregistered Evaluation Protocol for Tantular Cyber
- Bayesian Low-Rank Posteriors for Scalable Membership Inference
- PrivMark: Pixel-Space Steering against Privacy Leakage in Vision-Language Models
- EgoVeil: Personalized Privacy Reasoning for Egocentric Video
- MGE: A Membership Guided Data Extraction Attack against Diffusion Models
- CyberBinGym: Can AI Agents Turn Binary Patches into Real Exploits?
- Mirage: Provable Defense Against LLM KV Cache Timing Side Channels with Topology-Aware Isolation
- The Devil Wears Markdown: Detecting Malicious Intent Hidden in AI Agent Skill Instructions through Hybrid Static, Dynamic, and Agentic Analysis
- Agents of Chaos: A Systematization of Knowledge (SoK) on Malicious Attacks on Agentic AI Systems
- Validating AI-Discovered Cloud Threats on Live Infrastructure: A Multi-Model, Multi-Cloud Study
- Beyond AUROC: Deployment-Aware Evaluation of Model-Internal Jailbreak Detectors
- Tracing MCP Security Vulnerabilities: Constructing and Analyzing an MCP Security Vulnerability Dataset
- A Layer-wise Study of How Style Rewriting Breaks the Calibration of Zero-Shot LLM Fake-News Detectors
- Chatter Attack: Resource Consumption Attack for Large Language Models
- Least-Privilege Auditing of Agent Tool Calls with Intent Graphs and Reasoning-Guided Escalation
- SkillCODER: Towards Auditing and Attribution of Copyright Infringement in LLM Agent Skills
- RoCoMark: Robust Code Watermarking Against LLM-Based Functionality-Preserving Rewriting Attacks
- When Coding Agents Get Arms: Safety Risks Across the Digital-Physical Boundary
- Coding Agents Propagate Vulnerabilities in Code Bases
- On the Relative Security of Prompt Injection: Mechanised Definitions, Defences and Verdicts for LLM Agents
- Data Manipulation Attacks Against Machine Learning Integrated Cyber-Physical Systems
- PRISM: Causal Counterfactual Minimization of Agent Tool Privileges via Information Bottleneck Role Induction
- Treating the Model as an Untrusted Component: Deterministic Policy Enforcement, Human Approval and Taint Tracking Against Prompt Injection in Tool-Using Agents, with Validation on Open-Weights Models
- ShadowBench-MCP: Measuring Shadow Invocations in MCP-Style Tool Routing
- Ultrasound-Assisted Biometric Watermarking Against Face Video Deepfakes
- VeilTune: Secure Fine-Tuning of Private LoRA Adapters on Public Pretrained Language Models
- FALC: An Artifact-Centric Life Cycle Model and Research Agenda for AI Forensics
- StealthyPoison: Manipulating LLM Agent Behavior through Procedural Poisoning of Third-Party Skills
- SNAPRAG: Cost-Efficient Knowledge Extraction Attack on RAG Systems via Adversarial Perturbations
- Detecting Shadow AI: Unsanctioned LLM Agents on Enterprise Endpoints: A Taxonomy, an Open Detector and a Controlled Evaluation
- Say Shibboleth: Fingerprinting the LLMs and Harnesses Behind Agentic Attacks
- Suppression Is Not Erasure: Mechanistic Analysis and Reactivation Boundaries of Subspace Projection for LoRA Backdoor Defense
Safety & alignment16
- Double Blind Evals: Resolving the Dual Confidentiality Dilemma in AI Safety AuditingGoogle
- Enmo Safety: External, Non-invasive, Model-agnostic Oversight for Off-the-Shelf Language ModelsCode
- Personalized Alignment of Language Models with Human Value Profiles
- A Model Organism for Studying Unverbalized Evaluation Awareness
- META-IR: Meta-Introspective Reasoning for Adaptive Safety Alignment in LLM
All 16 papers in Safety & alignment
- Restoring General and Improving Domain-Specific Safety in Fine-Tuned LLMs Without Retraining
- Coding or Cheating? An Investigation of Frontier Coding LLMs
- What Do Collusion Probes Read? Three Pitfalls in Evaluating Multi-Agent Collusion Monitors
- Hallucination-Guided Unlearning: Using Hallucination Traces to Reveal Overfitted Memories
- R-BED: A Consequence-Aware Framework for System-Level AI Evaluation and Continuous Assurance
- When Safety Becomes An Outlier: Understanding the Retention of LLM Safety Behaviors
- System Prompts Affect Response Bias More Than Discriminability in LLM Safety
- ErosionBench: Probing Alignment Decay in Enterprise Multi-Agent Systems
- Safeguarding Agentic AI in Radiology Workflows: A Multimodel Evaluation of Rule-based and Large Language Model-based Verification
- FraudBench: A Legal Evaluation of AI Deception
- Can the Confidence and Decisions of System 1 Models Be Trusted? An Explainability and Perturbation Audit of Jev-Style Decision Models
Inference & efficiency14
- AutoEMatic: Runtime-Adaptive Mixed-Precision LLM Inference via Automatic Exponent-Mantissa RefinementNVIDIA
- Multi-GPU MoE Inference with Adaptive Expert OffloadingMicrosoft
- PACE: Keeping the Right Tokens in the KV Cache Is Not Enough for Long-Horizon LLM AgentsAmazon
- BridgeKV: Large-Model Prefill and Small-Model Decoding for LLM Inference via KV Cache TranslationMicrosoft
- Ensemble Routing for Prompt-Adaptive Cost-Aware LLM Inference
All 14 papers in Inference & efficiency
- Firepool: Brokering Warm GPU Capacity for Agentic and RL Workloads at LinkedIn
- Think or Remember: The Exchange Rate of Agentic Inference
- AutoQS: Automated Quantization Strategy Design with Large Language Models
- Translating KV Caches Between Hybrid-Attention Language Models: MLA Latents, Linear-Attention States and Sparse-Attention Indexers
- System Optimizations for Cost-Efficient Agent Harnesses
- Prune, Don't Merge: Expert Compression of Mixture-of-Experts LLMs on Multilingual Production Tasks
- MNN-Pocket-0.3B Technical Report
- Prediction with Foundation Models under Deployment Budgets
- Inference Validation for Regulated AI: A Governed Decision Pipeline for LLM Serving
AI & society10
- FAIDER: Fairness-Aware Item-Level Decoding for Large Language Model Recommenders
- Policy Over Purpose: Auditing LLM Content Moderation of Reclaimed Slur Usage
- Too Big to Inspect? Ethically Auditing Large Datasets Through Small Samples
- Behavioral Evaluation of Large Language Models Under Cognitive Biases in HAZOP Studies
- Measuring Generative AI Content in Political Campaigns: Evidence of Rising Adoption and Non-Disclosure
All 10 papers in AI & society
- State Media at Retrieval Time: A Large-Scale Audit of Russian State-Aligned Sources in AI Chatbot Responses
- CHO-Bench: How Do Context Hints Affect Ad Rankings in LLM Advertising?
- Detecting AI-Authored Text in Authentic Student Writing: Evidence from PISA 2025
- Fairness in Explainable AI Narratives: Decomposing Social Disparities in LLM-Generated Explanations
- Curbside Chatbots: Auditing the Mental Health Interventions Large Language Models Recommend to Clinicians, Caregivers, and Patients
Agents8
- VIVID: Bringing Tool-Augmented Agents into Enterprises
- Done Is Not Correct: Measuring Silent Failures and Self-Verification Calibration When LLM Agents Take CAD Actions
- Safe by Accident: Crash-Testing LLM-Written Code
- DeepFinance: Can AI Diligence a $100MM Deal?
- GARP: Real-Time Generative Agents on Local Hardware in a Game Engine
Robotics & embodied AI6
- HARP-Bench: Benchmarking Robot Manipulation Around Active HumansMetaCode
- DeltaForge: A Post-hoc Mixture-of-Experts VLA for Multi-Domain Autonomous DrivingNVIDIA
- From Game Generation to Game Playing: Fully Automated Ability Training for Vision-Language-Action Models
- Breaking survivorship bias in driving world models
- DriveFoundry: Co-Evolving Data and Model for Efficient Closed-Loop Iteration in Autonomous Driving
- Planners Exploit Deterministic Error: Dithered Low-Bit World Models
Vision & multimodal4
- Beyond Real or Fake: A Dual-Channel Authenticity and Reasoning Protocol for Photographic AssessmentGoogle
- ScreenProof: Where Is This Screen Wrong? Finding UI Faults Without a Reference
- From Weapons to Events: Evaluating Evidential Support in Multimodal AI for Policing
- Direct Trajectory Optimization for Inference-Time Alignment of Text-to-Image Diffusion Models
Language & speech4
- Yuanbao-Reranker: Document-Grounded Synthesis and Preference Distillation for Text RerankingTencent
- Retrofitting Unbounded Context onto Stock Language Models
- M5-TTS: A Mini-Scale Multilingual and Multi-Dialect Text-to-Speech Language Model with Multi-Token Prediction
- LAION-Audio-XL: A Web-Scale Corpus and Scaling Studies for Contrastive Language-Audio Pretraining