AI Research Radar
Saturday, May 9: no frontier papers out of 208 read.
Frontier
Papers rated as moving the field
How we chooseNone of the 19 cleared papers that day was rated frontier.
This week in research
May 3–9: 78 cleared, 4 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
19 papers
- More Is Not Always Better: Cross-Component Interference in LLM Agent ScaffoldingAmazon
- SafeHarbor: Hierarchical Memory-Augmented Guardrail for LLM Agent SafetyCode
- Correct Code, Vulnerable Dependencies: A Large Scale Measurement Study of LLM-Specified Library VersionsCode
- When Agents Handle Secrets: A Survey of Confidential Computing for Agentic AI
- Shattering the Echo Chamber: Hidden Safeguards in Manuscripts Against the AI Takeover of Peer Review
- On Semantic Loss Fine-Tuning Approach for Preventing Model Collapse in Causal Reasoning
- SLAM: Structural Linguistic Activation Marking for Language Models
- FinRAG-12B: A Production-Validated Recipe for Grounded Question Answering in Banking
- ReaComp: Compiling LLM Reasoning into Symbolic Solvers for Efficient Program Synthesis
- Irminsul: MLA-Native Position-Independent Caching for Agentic LLM Serving
- LoopTrap: Termination Poisoning Attacks on LLM Agents
- PragLocker: Protecting Agent Intellectual Property in Untrusted Deployments via Non-Portable Prompts
All 19 papers
- BehaviorGuard: Online Backdoor Defense for Deep Reinforcement Learning
- Policy-Guided Stepwise Model Routing for Cost-Effective Reasoning
- Beyond Accuracy: Policy Invariance as a Reliability Test for LLM Safety Judges
- Measuring Evaluation-Context Divergence in Open-Weight LLMs: A Paired-Prompt Protocol with Pilot Evidence of Alignment-Pipeline-Specific Heterogeneity
- PrefixGuard: From LLM-Agent Traces to Online Failure-Warning Monitors
- Instrumental Choices: Measuring the Propensity of LLM Agents to Pursue Instrumental Behaviors
- When No Benchmark Exists: Validating Comparative LLM Safety Scoring Without Ground-Truth Labels