QUEST
Training Frontier Deep Research Agents with Fully Synthetic Tasks
AI Research Radar
May 20–26: 5 frontier papers out of 1,612 read.
Frontier
Papers rated as moving the field
How we chooseDirectional Alignment Mitigates Reward Hacking in Reinforcement Learning for Language Models
JT-SAFE-V2
“Extensive evaluations demonstrate that JT-Safe-V2 achieves state-of-the-art performance across both general intelligence and safety benchmarks.”
Hallucination as Exploit
“A HACR audit on 500 stratified task keys shows that unsupported action-critical claims reach unsafe execution for naive agents (100.0%) and prompt-only defenses (49.6%), but not for ECA.”
DiscoverPhysics
“Across eleven frontier models, we find that the strongest agents pass only half of the worlds and consistently fail on those where latent structure must be uncovered.”
The week by field
May 20–26: 88 cleared, 5 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
83 papers
- How Well Do Models Follow Their Constitutions?Google
- CUA-Gym: Scaling Verifiable Training Environments and Tasks for Computer-Use AgentsAlibaba
- Robust LLM Watermarking with Minimal Semantic Distortion for IP ProtectionMicrosoft
- XWind: A Cross-site Router for Large Language Model Inference Serving at Renewable Energy FarmsMicrosoft
- SCI-Defense: Defending Manipulation Attacks from Generative Engine OptimizationAmazon
- Causal Unlearning in Collaborative Optimization: Exact and Approximate Influence Reversal under Adversarial ContributionsMeta
- The Illusion of Intervention: Your LLM-Simulated Experiment is an Observational StudyGoogle
- DECOR: Auditing LLM Deception via Information Manipulation TheoryAmazon
- Library Drift: Diagnosing and Fixing a Silent Failure Mode in Self-Evolving LLM Skill LibrariesAmazon
- Inverting the Shield: Systematically Generating Safety Tests from Policy SpecificationsCode
- LLM Agents Already Know When to Call Tools -- Even Without ReasoningCode
- Behavior Cue Reasoning: Monitorable Reasoning Improves Efficiency and Safety through OversightCode
All 83 papers
- Conditional Equivalence of DPO and RLHF: Implicit Assumption, Failure Modes, and Provable AlignmentCode
- Auditing Stealth Sycophancy in Mental-Health Dialogue: Structured Clinical-State Diagnostics and Clean Matched Benchmarks
- PASA: A Principled Embedding-Space Watermarking Approach for LLM-Generated Text under Semantic-Invariant Attacks
- Tokenizer Fertility and Zero-Shot Performance of Foundation Models on Ukrainian Legal Text: A Comparative Study
- Is VLA Reasoning Faithful? Probing Safety of Chain-of-Causation in Autonomous Driving Models
- LivePI: More Realistic Benchmarking of Agents Against Indirect Prompt Injection
- Language-Switching Triggers Take a Latent Detour Through Language Models
- Lying Is Just a Phase: The Hidden Alignment Transition in Language Model Scaling
- Agent-Facing Information Design in LLM Tool Registries
- Artificial Effort
- When the Manual Lies: A Realistic Benchmark to Evaluate MCP Poisoning Attacks for LLM Agents
- WideDepth: Millimeter-Accurate Benchmark for Fisheye Depth Estimation
- ContextEcho: A Benchmark for Persona Drift in Long Agentic-Coding Sessions
- A governance horizon for ethical-use constraints in open-weight AI models
- Poisoning the Watchtower: Prompt Injection Attacks Against LLM-Augmented Security Operations Through Adversarial Log Content
- IterInject: Indirect Prompt Injection Against LLM Agents via Feedback-Guided Iterative Optimization
- The Tokenizer Tax Across 25 European Languages: Domain Invariance, Cross-Lingual Few-Shot Effects, and the Ukrainian Penalty
- Who judges the judges? Governance from metrics: a runtime framework for continuous LLM compliance monitoring
- CONF-KV: Confidence-Aware KV Cache Eviction with Mixed-Precision Storage for Long-Horizon LLM
- RouteScan: A Non-Intrusive Approach to Auditing MoE LLMs Safety via Expert Routing Telemetry
- Reflect-Guard: Enhancing LLM Safeguards against Adversarial Prompts via Logical Self-Reflection
- Memory-Induced Tool-Drift in LLM Agents
- APT-Agent: Automated Penetration Testing using Large Language Models
- Latent Q-Barrier Shielding for Safe In-Context Reinforcement Learning
- READER: Reasoning-Enhanced AI-Generated Text Detection
- When Interpretability Becomes a Liability: Adversarial Attacks on CBM Concept Layers
- Evo-Attacker: Memory-Augmented Reinforcement Learning for Long-Horizon Tool Attacks on LLM-MAS
- LLM-as-a-Reviewer: Benchmarking Their Ability, Divergence, and Prompt Injection Resistance as Paper Reviewers
- Security of OpenClaw Agents: Fundamentals, Attacks, and Countermeasures
- Llamion Technical Report
- Capability and Robustness Cannot Both Be Free: An Information-Theoretic Bound for Vision-Language-Action Models
- PolyGnosis 2.0: Enhancing LLM Reasoning via Agentic Harness Engineering for Polymarket and OSINT Insight Extraction
- Hidden in Plain Tokens: Simply Robust, Gradient-Free Watermark for Synthetic Audio
- What Makes a Medical Checker Trainable? Diagnosing Signal Collapse and Reward Hacking in Checker-Guided RAG for Biomedical QA
- Paris 2.0: A Decentralized Diffusion Model for Video Generation
- Claw-Anything: Benchmarking Always-On Personal Assistants with Broader Access to User's Digital World
- Codec-Robust Attacks on Audio LLMs
- Is Capability a Liability? More Capable Language Models Make Worse Forecasts When It Matters Most
- ObjectCache: Layerwise Object-Storage Retrieval for KV Cache Reuse
- How Far Will They Go? Red-Teaming Online Influence with Large Language Models
- When AI Takes Sides on Questions of Faith: Persistent Asymmetries in AI-Mediated Faith Guidance
- Same Model, Different Weakness: How Language and Modality Reshape the Jailbreak Attack Surface in Frontier MLLMs
- Are Frontier LLMs Ready for Cybersecurity? Evidence for Vertical Foundation Models from Dual-Mode Vulnerability Benchmarks
- Cultural Adaptation in Large Language Models for Political Discourse
- MemAudit: Post-hoc Auditing of Poisoned Agent Memory via Causal Attribution and Structural Anomaly Detection
- Frequency-Domain Regularized Adversarial Alignment for Transferable Attacks against Closed-Source MLLMs
- What Counts as AI Sycophancy? A Taxonomy and Expert Survey of a Fragmented Construct
- On-Policy Consistency Training Improves LLM Safety with Minimal Capability Degradation
- Evaluating Large Language Models as Live Strategic Agents: Provider Performance, Hybrid Decomposition, and Operational Gaps in Timed Risk Play
- Benchmarking Autonomous Agents against Temporal, Spatial, and Semantic Evasions
- Compiling Agentic Workflows into LLM Weights: Near-Frontier Quality at Two Orders of Magnitude Less Cost
- Can AI Make Conflicts Worse? An Alignment Failure in LLM Deployment Across Conflict Contexts
- HealthCraft: A Reinforcement Learning Safety Environment for Emergency Medicine
- CR4T: Rewrite-Based Guardrails for Adolescent LLM Safety
- Blind Spots in the Guard: How Domain-Camouflaged Injection Attacks Evade Detection in Multi-Agent LLM Systems
- Boundary-targeted Membership Inference Attacks on Safety Classifiers
- Evaluating Commercial AI Chatbots as News Intermediaries
- The Impact of AI Search on the Online Content Ecosystem: Evidence from Google and Reddit
- Conformal Selective Acting: Anytime-Valid Risk Control for RLVR-Trained LLMs
- Mechanistic Interpretability for Learning Assurance of a Vision-Based Landing System
- Trusted Weights, Treacherous Optimizations? Optimization-Triggered Backdoor Attacks on LLMs
- REFLECTOR: Internalizing Step-wise Reflection against Indirect Jailbreak
- Decoupling Communication from Policy: Robust MARL under Bandwidth Constraints
- roto 2.0: The Robot Tactile Olympiad
- PrivScope: Task-scoped Disclosure Control for Hybrid Agentic Systems
- Prompt2Fingerprint: Plug-and-Play LLM Fingerprinting via Text-to-Weight Generation
- UCCI: Calibrated Uncertainty for Cost-Optimal LLM Cascade Routing
- Benchmarking Commercial ASR Systems on Code-Switching Speech: Arabic, Persian, and German
- Agent Meltdowns: The Road to Hell Is Paved with Helpful Agents
- GoLongRL: Capability-Oriented Long Context Reinforcement Learning with Multitask Alignment
- Auditing Privacy in Multi-Tenant RAG under Account Collusion