DeepSeek-V4.1-Flash
Pushing the Limits of KV Cache Compression
AI Research Radar
September 18–24: 16 frontier papers out of 2,330 read.
Frontier
Papers rated as moving the field
How we chooseHunyuan-A13B Technical Report
“Evaluations show competitive performance across mathematics, science, programming, general language understanding, and agent tasks, often approaching that of much larger models.”
APort Vault
“The zero spans 790 source sessions, giving a per-session upper bound of 0.38%.”
Your Model Is Leaking
“Across eleven models from seven architecture families, our evaluation shows 91--100% recovery on nine models with KL divergence 0.001--0.007, while evaluated activation-level detectors remain close to random guessing (AUC <= 0.56).”
FrontierMath Erdős
“One (GPT-6 Astra) scored 3%, and all others scored 0%.”
Evaluating Coding Agents on Kernel Exploit Generation
“With a reference PoC, the strongest configuration solves 31 of 45 tasks (68.9%).”
Qwen3.8-Omni
“Extensive evaluations demonstrate that Qwen3.8-Omni-Flash achieves strong performance across multimodal understanding, reasoning, long-horizon agentic execution, and video productivity tasks.”
Beyond Task Completion
“Starting from Qwen3.5-9B, SCOPE-RL achieves a 54.17% task success rate on OSWorld and a 64.30% attack-avoidance rate on OS-BLIND, yielding the best aggregate capability--safety score of 58.80% among the evaluated agents.”
Euston
“Median response length also falls from 19,217 to 18,296 tokens and the truncation rate from 25.8% to 8.3%, so the improvement does not come from thinking longer.”
Control-Token Injection Suppresses Chain-of-Thought and Defeats Reasoning-Based Oversight in Tool-Using Agents
“We show the suppression can be delivered indirectly and characterize its dependence on the chat template across two more reasoning models, and we evaluate input sanitization, parser hardening, and empty-reasoning detection as defenses; flagging an absent trace catches the basic attack but not an adaptive benign decoy.”
InternW0
Delegated Misalignment
“We show that this assumption breaks down: \emph{individual safety alignment fails to transfer to multi-agent settings}.”
Optimizing the Score, Losing Sight of the Task: Reward Hacking Across Weights, Selection, and Prompts
OPBackdoor
Open ultrasound foundation model for robust segmentation and clinical measurement across heterogeneous settings
Harm Laundering in GPT Models
The week by field
Sep 18–24: 199 cleared, 16 frontier
- Security59◆ 5
- Safety25◆ 3
- Agents22◆ 1
- 4 moreML, Vision, Training, Reasoning15◆ 3
- Society14
- Language10
- Robotics9◆ 1
- Inference8
- Science5
- Security59◆ 5
- 8 moreLanguage, Robotics, Inference, Science, ML, Vision, Training, Reasoning47◆ 4
- Safety25◆ 3
- Agents22◆ 1
- Society14
Size: cleared papers◆ and tint: frontier papers32 papers not yet sorted into a field
Models under test
Papers this week that name the model
Labs publishing
Papers this week with an author at the lab
Also cleared
183 papers, by field
Security & privacy54
- From Alignment to Access Control: A Framework for GenAI Policy EnforcementIBM
- Beyond Predictable Paths: Redefining AI Security Incident Reporting for AgentsMicrosoft
- ODPure: Backdoor Purification for Object Detection via Ensemble Corruption ConsensusCode
- A Data-Interventional Framework for Auditing Privacy and Fairness in Generative Medical ImagingCode
- The Like Trap: Multi-Stage Poisoning against Agents in Similarity-based Recommendation Systems
All 54 papers in Security & privacy
- Multi-View Fusion for Encrypted C2 Detection: A Leakage-Controlled Measurement Study of Evaluation Pitfalls
- Only Pay What You Must Spend: On-Demand Privacy Budget Payment for Differentially Private RAG
- RAMP: Reversing Adversarial Perturbations to Strengthen Clean-Label Backdoor Attacks against Malware Detectors
- Invisible in Space, Visible in Time: Motion Vision CAPTCHA against GUI Agents
- Agent Name Collision Attacks in Multi-Agent Systems
- ChronosAttack: Adversarial Tool Scheduling Attacks on LLM Agents
- ChatGPT Images 2.5 in the Wild: A Launch-Period Dataset and Detector Evaluation
- Not All 4-bit Quantizers Are Equal: Deployment-Time Mitigation of PII Leakage in Fine-Tuned Small Language Models
- ZeroGate: Trust-Preserving Fast Paths for Governed AI Agent Runtimes
- RAG-NAROK: Retrieval-Aware Knowledge Corpus Poisoning in RAG with Source-specific Refutation
- LoRango: It Takes Two LoRAs to Unlock Hidden Behaviors in Diffusion Models
- Policy-Backed Selective Regeneration under Tainted Inter-Agent Communication
- Selection-Invariant Communication Compilers for Privacy-Aware Multi-Agent LLM Workflows
- StepTrigger: Contact-State-Triggered Backdoor Attacks on VLM-Powered Legged Robots
- Silent Sabotage: Internal State Triggered Backdoor Attacks on LLM-Powered Robotic Systems
- Towards Effective Black-Box Adversarial Attacks on Deep Code Models via Structural and Identifier Perturbations
- From Approval to Execution: Reconstruction-Aware Repair Analysis for LLM-Agent Software
- Team DArgk at the 2026 ELOQUENT lab for evaluating generative language model quality: Residuals of Humanity: AI Detection Evasion via GRPO Fine-Tuning
- The Corroboration Illusion: When More News Makes LLM Forecasts Less True
- Defusing Explosive Prompts: Understanding and Preventing Trigger-Based Prompt Injections in LLM Agents
- From Capability to Assurance in Autonomous Penetration-Testing Harnesses: A Framework and Reference Implementation
- UBA-ORL: Unlearning-Activated Backdoor Attacks on Offline Reinforcement Learning
- The Price of Safety: Benign-Case Utility and Token Overhead of Memory-Poisoning Defenses in LLM Agents
- Beyond Single-Model Injection: A Threat Model and Defense Architecture for Prompt Injection in Multi-Agent Systems
- When Agentic Trust Crosses Organizational Boundaries: Structural Externalization and a Reference Model for Trust Evidence
- LLMs as Linguistic Chameleons: Decoupling Semantics and Structure for Privacy-Preserving Communication
- Runtime Authorization Consistency Checking for MCP-based Agentic Workflows
- An Efficient and Effective Watermarking Scheme for the Protection of the Intellectual Property Rights of Video Generative Models
- When the Agent Becomes the Kernel: A Systematization of Security on the Path to AI-Native Operating Systems
- Agents That Edit Documents: Measuring Agentic PDF Forgery Against a Non-Agentic Control
- MobileCybench: Evaluating Agent Vulnerability Discovery via Executable Probes
- LeaseGuard: Incumbent-Preserving Admission Control for Privileged LLM Agents
- From Bits to Beliefs: Recoverable Semantic Fingerprints for Black-Box Verification of Large Language Models
- Forgeable Confirmation in Automated Computer Security Testing: Deterministic Rules versus AI Judges
- ActGov: Governing LLM Agent Actions via Policy-Constrained Validation
- State-Aware Fuzzing of JavaScript Engines with LLM-Guided Instrumentation
- Feedback Coding Enables Inference-Time Covert Agentic Communication
- The Right Tool for the Job: On the Selection of Mitigations for GenAI Privacy Threats
- Origin Is All You Need: Provenance-Aware Transformers for Structural Trust-Boundary Separation
- Conformal Privacy Auditing: Calibrated Re-identification Attacks with Statistical Guarantees
- Hiding in Plain Sight: A Diffusion-based Mitigation of Geolocation Privacy Leakage in Vision-Language Models
- DEFEAT: Stitching Fragmented File I/O Contexts for Early Ransomware Detection
- HE-Guardrail: A Homomorphic Guardrail Against Jailbreak Attacks for Encrypted Large Language Model Inference
- ServeGuard: Verifiable, Bounded-Residual Confinement of Operator-Invisible Channels Without Revealing the Certified Read Factor
- Micro-Collaborative Poisoning: A Distributed Attack on RAG Systems
- Towards Zero-Shot Attribution of Synthetic Speech via Audio-Text Contrastive Retrieval
- CIPL: A Channel-Aware Framework for Recoverable Privacy Leakage in LLM Agents
- TERMon: Detecting Persistent Behavioral Threats in Edge AI via Hardware-Native Ternary Runtime Monitor
- CASCADE Against Jailbreaks: Combination Across Stages with Controlled Attack-Defense Evaluation
Safety & alignment22
- Swiss-Knife: A Framework for Reconfigurable Externalised Multi-Objective Alignment at Decode TimeGoogle, Amazon
- CART: Closed-Loop Adaptive Red Teaming for Large Language ModelsMicrosoft
- InGuard: Towards Generalized Inner Guardrail for Safe Text-to-Image GenerationAlibaba
- Safety Nudges: User-Facing Interventions for Real-Time AI Risk Awareness
- Hard Negatives Reveal What Easy Negatives Hide: Cross-Lingual Harmfulness Representations Degrade with Resource Tier Under Hard Negatives
All 22 papers in Safety & alignment
- When Accuracy Gaps Fail to Certify: Auditing Cross-Domain Recalibration of LLM Judges
- Finite-Sample Probabilistic Safety Certification for AI-Based Grid-Edge Coordination
- Shutdown Sabotage Propensities in Multi-Agent Systems
- An Open Pipeline and Dashboard for Systemic-Risk Evidence under the EU AI Act's Code of Practice
- Mitigating LLM Over-Refusal via Dynamic Semantic Routing Calibratione
- Recovering Agentic Sovereignty: Mitigating the Consensus Paradox via Contrastive Epistemic Decoding
- The Uncontrolled Variable: Vision-Language Model Refusal Responds to Image Presence in Ways Risk Cannot Explain
- EADC: Evaluation of Advanced and Deep-level Compliance in Large Language Models
- Correcting Learning-based Perception for Safety
- SafeTune: A Unified Faithful Library for Auditing and Repairing Safety Drift in Fine-Tuned LLMs
- VLPSA: Vision-Language-Poisson-Safe Actions for Full-Body Safety of Learned Policies
- Agreement Overstates Evidence: Error Dependence in LLM Judge Consensus
- Auditing Political Alignment in LLM Assistants: Engagement, Stance, and User Identity
- Directing large language models to follow the letter or spirit of the law
- Self-Healing Harness for Runtime Oversight of Agent Self-Modification
- Emergent Collusion in Long-Horizon LLM Agent Interaction
- SafeStage: Evaluating Safety Before, During, and After Vision-Language-Conditioned Robot Manipulation
Agents21
- MemCalib: Benchmarking and Optimizing Memory Use in LLM AgentsAlibaba
- BI-Agent and BI-Bench: Towards Automating End-to-End Business IntelligenceMicrosoft
- Scaling Discovery through Test-Time CommunicationMicrosoft
- Verify, Don't Trust: Agentic Model Development for Video Discovery Retrieval at ScaleMeta
- What Makes a Terminal-Bench Task Hard? Separating Genuine Hardness from Fake-Hardness on an Adjudicated Agentic Corpus
All 21 papers in Agents
- Propose, Don't Judge: An Anytime-Valid Referee for LLM Agents That Mine Investment Factors
- Policy-as-Skill: Governed LLM Decision Support with Evidence, Deterministic Control, and Audit
- Specifying and Maintaining Agentic Workflows: An Empirical Study of GitHub Agentic Workflows
- Issuer-Sovereign Agentic Payments
- Compliant AI Infrastructure for Regulated Finance: A tiered multi-agent framework with DLT audit trails for financial operations in DACH
- SkillGym: Internalizing Human Skills into LLMs for Real-World Problem Solving
- From Agent Output to Authorized Transition
- TRACTOR Benchmark for Evaluating C to Rust Translators
- CliffCompaction: Cost-Efficient Compaction for Long-Horizon Coding Agents
- An Empirical Cost Attribution of Context-Compression Gateways in Multi-Turn Coding Agents
- NostrAgent: A Decentralized Identity and Delegation Architecture for Sovereign Agentic Systems
- AgentRouter: Heterogeneous Model Routing for Cost-Optimal Multi-Step Agentic Workflows
- Total Cost of Agency: Exact Attribution of Memory Injection Cost in Multi-Agent LLM Workflows
- LEGIT: Credentialing Protocol for Trustworthy AI Agent Marketplaces
- What Stops a Small Language Model From Driving a Database Agent
- MintAct: A Unified Visual Agent for Digital Environments
AI & society14
- Towards participatory speech dataset curation: A queer case study and conceptual frameworkMeta
- EnterpriseVal: Quantifying the Efficacy, Reliability and Value of Generative AI in the EnterpriseNVIDIA
- Compliant with Local Controls, Collectively Discriminatory. A Governance Architecture for Multi-Agent AI in Regulated Finance
- "We'll Fix It Later": Education, AI, and the Deferral of Privacy in EdTech
- The Architect, the Adversary, and the Judge: Closed-Loop Generation of Standards-Aligned Assessment Items at Scale
All 14 papers in AI & society
- FairMon: A Tool for Monitoring and Visualizing Algorithmic Fairness
- Monocultural Biases: Correlated biases in large language models lead to unequal systemic exclusion rates in hiring
- Fairness Beyond Anonymization? Demographic Leakage in German LLM-Generated Resumes
- The Moral Check: Strategic AI Governance for the Pacing Problem
- The Law of Stop: Interruptibility, Injunctions, and the Governance of Agentic AI
- OmniEdu: Open Foundation Models for Learning and Teaching
- Context-Aware Pre-Deployment Evaluation of AI Systems: A Regulatory Framework for Nigerian Fintech
- Convex AI Compositionality and the Governance of AI System Populations
- AI-GRACE: A Use-Case Operationalization Framework for Agentic AI: From Organizational Objectives and Obligations to Deployment Capabilities and Architecture
Language & speech10
- Qwen-Audio-3.1-Realtime: Towards Reliable Agentic Voice InteractionAlibaba
- I'll Keep an Ear Out: Teaching AudioLLMs Proactive Audio AssistanceMeta
- When Context Misleads: In-context Learning with Jurisdiction in Large Language ModelsCode
- The Recall Ceiling of LLM Recommendation Reranking
- EvidenT: Building Trustworthy Enterprise Assistants through Evidence Groundedness and Traceability
All 10 papers in Language & speech
- Per-Query Gating of LLM Rerankers for Multi-Hop Retrieval
- Attributable Post-Rationalization in RAG Citations: A Controlled Reproduction and an RLVR Comparison
- Auditing Source Exposure in Baidu and Google AI Search
- Samsone: A Family of Open Small Audio Language Models for On-Device Inference
- NemotronLabs VoiceChat: An Open Full-duplex Speech-to-Speech Model with Tool Calling Capabilities
Robotics & embodied AI8
- ME-VLM:A Unified VLM for Embodied Cognition and Agent CoordinationCode
- Turning Safety into Competence: Minimally Exploitable Robot Policies via Safety-Filtered Reinforcement Learning
- MotionForge: A Data Generation Pipeline and Large-Scale Benchmark for Long-Horizon Manipulation of Dynamic Objects with Domain Shifts
- Robot World Models Are Not Invariant to How the Actions Are Written
- Catch Me If You Can: Real-Time Feedback Denoising for Responsive VLAs
All 8 papers in Robotics & embodied AI
Inference & efficiency8
- Crossflow: Prefill-Decode Elasticity for Agentic LLM ServingMeta
- Greedy Decoding Is Not Precision-Invariant: Cross-Precision Output Divergence in LLM InferenceAmazon
- SPHQuant: Efficient extreme low bit weight quantization for Vision-Language ModelsCode
- Learning the Cost of Reliable Inference
- Measured Joules, Learned Routes: Learning to Route for Energy-Efficient LLM Serving
Science & health5
- An open benchmark for machine learning-based polymer property predictionMetaCode
- nnFoundation: 3D Foundation Models for Radiology
- A generalizable structural brain MRI foundation model built through dual-priority federated pretraining
- WILSON - a pathology foundation model framework for patient-level analysis and diagnostic text generation
- Toward Fairness in Machine Learning Models for Predicting Treatment Retention and Premature Discontinuation in Medication for Opioid Use Disorder
Machine learning5
- Evaluation Choices Decide the Forecasting Leaderboard: Evidence from a Production Marketplace Panel
- Causilo Technical Report
- Do Not Trust the Benchmark: Limitations of General LLM Rankings and a Case for Task-Specific Evaluation
- t₀: A Time-Series Foundation Model for Forecasting with Context
- Pinocchio: Fast Uncertainty Estimates for Black-Box Language Models
Vision & multimodal4
Training & data3
Not yet sorted29
- Red-Teaming Auto Mode: Improving Blocking Classifiers Against Malign Coding AgentsAnthropicCode
- Decaf: A privacy preserving speech codec using speaker disentanglement and canonical voice conversionMicrosoft
- Quantifying Overclaiming Propensity in Frontier LLM AgentsCohere
- SAFARI: An Industrial Benchmark for LLM-Assisted Hazard Analysis and Risk AssessmentCode
- What Users Think of Generative AI: A Cross-Platform NLP Analysis of Trust and Friction in App Store Reviews
All 29 papers in Not yet sorted
- Code-as-Auditor: Executable Compliance Reasoning via Regulation-to-Code
- Robust Conformal Intrusion Detection via Traffic-Aware Calibration and Attack-Orbit Invariance
- MAGS: Multi-agent Auto-formalization Guarantees Safety for Agentic Outputs
- Beyond Private Training: The New Landscape of AI Privacy
- DeltaSelect: Affordable A/B Testing for Coding Agents
- MiX: Micro-Inverted-Scaling for End-to-End Low-Bit Vision-Language Model Acceleration
- ALIBI: Adversarial Legitimacy Injection in Binary Input against LLM Malware Analyzers
- Benchmarking LLM Compliance with China AI Generated Content Regulations
- Geopolitical Divisions Across Languages in Large Language Models
- Governance-as-Code: Translating EU AI Act Technical Requirements into Executable Compliance Pipelines for Generative AI Systems
- Can Data Attribution Filter Out Subliminal Learning? Not Reliably
- Astronex-World 1.0: Real-Time Interactive World Model Foundation
- Local Sparsity Enables Unsupervised LLM Safety Detection
- Designing Against Deskilling: Metacognitive Feedback Reduces Cognitive Offloading to LLM Assistants
- ResumeShield: Channel Separation and an Open Benchmark for Indirect Prompt Injection in AI Resume Screening
- Silence Is Endorsement: Verification-Status Laundering in LLM Agent Pipelines
- Is It Still Worth Training a Classical Model in the Era of LLMs? A Crossover Benchmark on Tabular Data
- Detecting Deceptive Recruitment: A Signal-theoretic Machine Learning Framework for Early Identification of Labour Exploitation
- The More It Says, the More You Pay: A Black-Box Audit of Provider-Side Token Inflation in LLM Services
- Fingerprinting Multimodal Large Language Models
- greCAPTCHA: Assessing Understanding as Evidence of Research Authorship Under Generative AI
- Towards TEE-Certified DP: Verifiable Differentially Private Training on Legacy GPUs
- Refuse, Decompose, Refresh: A Claim-Safe Protocol for Closed-Loop AI Evaluation
- Inference-Engine Fingerprinting Attacks are Practical: Exploring Model-Driven Environmental Discovery, Exploitation, and Escape