AI Unlearning Benchmarks Are Broken. Here Is What We Should Measure Instead
中文摘要
AI遗忘基准有缺陷。我们应建立新指标来衡量模型遗忘信息的能力，而不仅是测试其知识。
English Summary
AI unlearning benchmarks are flawed. We need new metrics to measure how models forget, rather than just testing what they know.
原文节选
AI is very good at proving what it knows. Accuracy scores, reasoning tests, coding benchmarks, math exams, latency charts, and leaderboard… Continue reading on Medium »