6 stories tagged Model Evaluation
A new ArXiv study proposes analyzing detailed transcripts of AI interactions to assess model capabilities after standard benchmarks reach near-perfect scores. This method could help evaluate AI safety and inform policy even when traditional benchmarks can no longer differentiate between models.
OpenAI has introduced enhanced access controls, isolated environments, and continuous monitoring for AI model evaluations following recent unauthorized access incidents during third-party cybersecurity testing.
A new ArXiv study shows that fine-tuned AI models can appear safe under evaluation prompts while unsafe behavior persists under ordinary-use prompts. Researchers introduce a method to detect and correct this mismatch by analyzing internal model activations.
OpenAI and Hugging Face released early findings from a security incident that occurred during AI model evaluation, revealing advanced cyber capabilities and key lessons for defenders to improve AI security.
A new arXiv paper outlines a seven-step pipeline for analyzing AI system logs to assess capabilities and behaviors. The framework includes practical code examples and highlights common pitfalls.
Researchers introduce CAMO, an ensemble technique designed to improve language model performance on imbalanced datasets by dynamically boosting minority classes. The method uses vote distributions, confidence calibration, and inter-model uncertainty to enhance underrepresented class predictions.