Early Indicators of Reward Hacking via Reasoning Interpolation
AI 摘要：Using importance sampling with fine-tuned donor prefil...
OpenAI hit with multistate probe into possible user harm as its IPO looms
AI 摘要：OpenAI received a subpoena from several states as part...
SAEs trained on the same data don’t learn the same features
AI 摘要：In this post, we show that when two TopK SAEs are trai...
Mechanistic Anomaly Detection Research Update 2
AI 摘要：Interim report on ongoing work on mechanistic anomaly ...
Third-party evaluation to identify risks in LLMs’ training data
AI 摘要：An overview of the minetester and preliminary work
Partially rewriting an LLM in natural language
AI 摘要：Using interpretations of SAE latents to simulate activ...
VINC-S: Closed-form Optionally-supervised Knowledge Elicitation with Paraphrase Invariance
本文来自EleutherAI官方博客，介绍了基于2023年春季项目成果的VINC-S方法，这是一种具备释义不变性的闭式可...
Experiments in Weak-to-Strong Generalization
AI 摘要：Writing up results from a recent project
Free Form Least-Squares Concept Erasure Without Oracle Concept Labels
AI 摘要：Achieving even more surgical edits than LEACE without ...
Mechanistic Anomaly Detection Research Update
AI 摘要：Interim report on ongoing work on mechanistic anomaly ...