on-policy distillation (OPD)

5 atoms · 跨 1 天 · 首见 2026-06-09 · 最近 2026-06-09

三色: 🟦 fact 2 · 🟥 take 3 · stance ▲1/▼0/◆4

来源: X 5

时态: fresh:5

标签: 好观点:4 · 好问题:2 · 好数字:2

叙事 (narrative) (3)

• 2026-06-09 · is not just SFT-lite
  On-policy distillation is not just SFT-lite
  tags: 好观点·好问题 · → daily
• 2026-06-09 · challenges intuition that OPD sits between SFT and RLVR
  This challenges the intuition that OPD sits between SFT and RLVR.
  tags: 好观点·好问题 · → daily
• 2026-06-09 · defines its own geometric regime
  It defines its own geometric regime.
  tags: 好观点 · → daily

事实 (fact) (2)

• 2026-06-09 · rapidly locks into a narrow subspace that alone sustains strong reasoning
  it rapidly 'locks' into a narrow subspace that alone sustains strong reasoning.
  tags: 好数字·好观点 · → daily
• 2026-06-09 · avoids principal directions and sparsely updates fewer weights
  It avoids principal directions and sparsely updates fewer weights.
  tags: 好数字 · → daily

时间轴 (近 20)

• 2026-06-09 · narrative · is not just SFT-lite · → daily
• 2026-06-09 · fact · avoids principal directions and sparsely updates fewer weigh · → daily
• 2026-06-09 · fact · rapidly locks into a narrow subspace that alone sustains str · → daily
• 2026-06-09 · narrative · challenges intuition that OPD sits between SFT and RLVR · → daily
• 2026-06-09 · narrative · defines its own geometric regime · → daily

← 实体目录 · 系统日志