Projection sampling: SFT learns like RL without catastrophic forgetting

A NeurIPS paper introduces projection sampling, an MCMC-based data transformation that lets SFT acquire new capabilities with less forgetting, matching or beating RL-based post-training in generalization.

2026-10-02 ~ 2026-10-03 · 4 related posts

• Pope says AI lacks the spark of humanity; Dan Shipper calls it a category error — danshipper · 2026-10-02
• Pope comments on AI art, Schmidhuber fires back a response — Merzmensch · 2026-10-03
• Pope Leo XIV: AI-generated images and human art differ ontologically, not just aesthetically — bennash · 2026-10-03
• Pope says algorithms lack the human spark: AI art and human art differ ontologically — sorenrood · 2026-10-03