New arXiv paper: variational estimator for Lp calibration errors separates over- and under-confidence
_onionesque · x · 2026-09-15
arXiv 2602.24230（Berta、Braun、Holzmüller、Francis Bach、Michael I. Jordan）提出估计 Lp 校准误差 的变分方法。
- 校准要求预测概率与实际类别频率一致，传统用散度函数评估，多分类场景下很难准确估计。
- 该工作把最近针对 proper loss 的变分校准估计框架扩展到 Lp 散度诱导的校准误差。
- 方法可区分过度置信与不足置信，且不像非变分方法那样高估误差。
- 代码已集成进开源包 probmetrics。
More from Research
- Dev cracks image conditioning on a DIY video model trained on one hour of footage — pixlpa · 2026-09-15
- MIT's deterministic math solver boosts clinical LLM accuracy — but only for larger models — MIT · 2026-09-15
- TailSFT: Skipping SFT Gains Improves pass@16 and RL Exploration in GRPO — tw_killian · 2026-09-15
- STARK Co-inventor Notes STARKs Are Post-Quantum Secure — jamestagg · 2026-09-15
- Self-evolving AI agents shouldn't grade their own homework: a 7-step evidence-gated roadmap — MaryamMiradi · 2026-09-15
- I2T loss stays comparable across tokenizers and predicts image generation quality — peterxichen · 2026-09-15