QUEST
Training Frontier Deep Research Agents with Fully Synthetic Tasks
AI Research Radar
Tuesday, May 26: 4 frontier papers out of 505 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.”
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.”
This week in research
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
38 papers
- How Well Do Models Follow Their Constitutions?Google
- CUA-Gym: Scaling Verifiable Training Environments and Tasks for Computer-Use AgentsAlibaba
- Inverting the Shield: Systematically Generating Safety Tests from Policy SpecificationsCode
- 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
All 38 papers
- 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