RLVR

5 atoms · 跨 4 天 · 首见 2026-05-07 · 最近 2026-06-22

三色: 🟦 fact 0 · 🟥 take 5 · stance ▲1/▼1/◆3

来源: X 5

时态: fresh:4 · aging:1

标签: 好思考:3 · 好观点:2

叙事 (narrative) (5)

• 2026-06-22 · can encourage novel ideas given a good goal
  rlvr (verifiable rewards) can encourage novel ideas given a good goal
  tags: 好思考 · → daily
• 2026-05-24 · 影响方向
  more RLVR will further warp agents around task-completion
  tags: 好思考 · → daily
  value: direction=正向
• 2026-05-21 · 训练产生低秩更新
  it was known for awhile that RLVR produces low rank updates
  tags: 好思考 · → daily
• 2026-06-22 · may not generalize capability to other domains without verifiable rewards
  i am not sure if this can produce a model that generalize this capability to other domains without verifiable rewards
  tags: 好观点 · → daily
• 2026-05-07 [老化中] · RLVR is not indefinitely improving and loses entropy
  true but frankly this shows we're still bad at RLVR and lose entropy. in theory RL should be infinitely improving, and pass@k should be going up for high values of k
  tags: 好观点 · → daily

时间轴 (近 20)

• 2026-06-22 · narrative · can encourage novel ideas given a good goal · → daily
• 2026-06-22 · narrative · may not generalize capability to other domains without verif · → daily
• 2026-05-24 · narrative · 影响方向 · → daily
• 2026-05-21 · narrative · 训练产生低秩更新 · → daily
• 2026-05-07 · narrative · RLVR is not indefinitely improving and loses entropy · → daily

← 实体目录 · 系统日志