AI Research Radar

Tuesday, May 5: 1 frontier paper out of 417 read.

1 frontier15 cleared417 papers read

Frontier

Papers rated as moving the field

LLM Ghostbusters

“We demonstrate that AU reduces package hallucination rates by 81%, corresponding to a substantial reduction in slopsquatting attack surface, while maintaining performance on standard coding benchmarks.”

The authors, in the abstract

Surgical Hallucination Suppression via Adaptive Unlearning

arXiv PDF

This week in research

Apr 29 – May 5: 124 cleared, 12 frontier

The week’s papers have not been sorted into fields yet.

Models under test

Papers this week that name the model

GPT-54
GPT-4o mini3
GPT-4o3
GPT-5.42
Claude Opus 4.62
GPT-5.3-Codex1
Gemini 3 Pro1
Claude Sonnet 4.61

Labs publishing

Papers this week with an author at the lab

AllenAI2
NVIDIA2
Alibaba1
Anthropic1
Google1
IBM1
Meta1
Microsoft1

Also cleared

14 papers

ARGUS: Policy-Adaptive Ad Governance via Evolving Reinforcement with Adversarial UmpiringTencent
ContextualJailbreak: Evolutionary Red-Teaming via Simulated Conversational PrimingIBM
SRTJ: Self-Evolving Rule-Driven Training-Free LLM JailbreakingCode
SplitZip: Ultra Fast Lossless KV Compression for Disaggregated LLM ServingCode
The Provenance Gap in Clinical AI: Evidence-Traceable Temporal Knowledge Graphs for Rare Disease Reasoning
When Embedding-Based Defenses Fail: Rethinking Safety in LLM-Based Multi-Agent Systems
Led to Mislead: Adversarial Content Injection for Attacks on Neural Ranking Models
Probe-Geometry Alignment: Erasing the Cross-Sequence Memorization Signature Below Chance
NH-CROP: Robust Pricing for Governed Language Data Assets under Cost Uncertainty
The Compliance Gap: Why AI Systems Promise to Follow Process Instructions but Don't
Zero-Shot Confidence Estimation for Small LLMs: When Supervised Baselines Aren't Worth Training
Decoding-Time Debiasing via Process Reward Models: From Controlled Fill-in to Open-Ended Generation

All 14 papers

Efficient Preference Poisoning Attack on Offline RLHF
PubMed-Ophtha: An open resource for training ophthalmology vision-language models on scientific literature

SIGNIFICANT ON THE BOARDS

MODEL TRACKER2,648 new benchmark results across 174 models this week DEAL FLOW$158B disclosed AI funding this week across 61 rounds POLICY125 policy events tracked across 7 days