One Token to Fool LLM-as-a-Judge
“Our systematic evaluation demonstrates this is a widespread failure affecting a diverse range of models, including leading proprietary systems such as GPT-o1 and Claude-4.”
The authors, in the abstract
AI Research Radar
June 26 – July 2: 3 frontier papers out of 436 read.
Frontier
Papers rated as moving the field
How we chooseSWE-MERA
“Evaluation using the Aider coding agent demonstrates strong discriminative power in state-of-the-art models.”
Trojan Horse Prompting
The week by field
Jun 26 – Jul 2: 12 cleared, 3 frontier
Also cleared
9 papers
- InvisibleInk: High-Utility and Low-Cost Text Generation with Differential PrivacyGoogleCode
- RoboEval: Where Robotic Manipulation Meets Structured and Scalable EvaluationAllenAI
- Efficient Unlearning with Privacy GuaranteesCode
- Reconstructing Template-Memorized Images from Natural PromptsCode
- Exposing the Illusion of Fairness: Auditing Vulnerabilities to Distributional Manipulation AttacksCode
- Intrinsic Fingerprint of LLMs: Continue Training is NOT All You Need to Steal A Model!
- LoRAShield: Data-Free Editing Alignment for Secure Personalized LoRA Sharing
- Policy-Grounded Safety Evaluation of 20 Large Language Models
- OASIS: Outlier-Aware LUT-Based GEMM with Dual-Side Quantization for LLM Inference Acceleration