MIMO paper

9 atoms · 跨 2 天 · 首见 2026-06-30 · 最近 2026-07-01

三色: 🟦 fact 0 · 🟥 take 9 · stance ▲3/▼1/◆5

来源: X 9

时态: fresh:9

标签: 好观点:7 · 好问题:2 · 好信源:1

叙事 (narrative) (8)

• 2026-06-30 · top-k OPD not useful result contrasts DeepSeek findings
  Also they show top-k OPD not being useful going against deepseek results 👀
  tags: 好观点·好信源 · → daily
• 2026-07-01 · mix-RL not bitter lesson pilled in terms of large scale generalisation
  But also would say its not very bitter lesson pilled in terms of large scale generalisation.
  tags: 好观点 · → daily
• 2026-07-01 · RL on SWE with length penalty generalizes to other domains
  Also best example of generalisation is that if u ran RL on SWE with a length penalty then eval'd on other domains, it would 100% also have gotten more efficient there.
  tags: 好观点 · → daily
• 2026-07-01 · zero generalization across domains not believed
  But like do u fully believe there is 0 generalisation across these domains
  tags: 好问题 · → daily
• 2026-07-01 · extra MOPD compute needed in comparison with mix-RL
  Bcs u need to count the extra MOPD compute. In SWE even without that mix RL is the same
  tags: 好观点 · → daily
• 2026-06-30 · SWE teacher equivalent to mix with MOPD
  The SWE teacher seems equivalent to the mix with MOPD then spending more compute for the same performance.
  tags: 好观点 · → daily
• 2026-06-30 · mix-RL better than MOPD for SWE
  For SWE u would argue here mixRL was better here.
  tags: 好观点 · → daily
• 2026-06-30 · generalization between domains is non-zero
  Like do we fully think theres like 100 opus and codex experts, and that the generalisation between these domains is nothing?
  tags: 好问题 · → daily

预测 (forecast) (1)

• 2026-07-01 · possible optimum is MixRL for scale-up, MOPD for scale-out
  Perhaps the optimum is MixRL for scale-up, MOPD for scale-out
  tags: 好观点 · → daily

时间轴 (近 20)

• 2026-07-01 · forecast · possible optimum is MixRL for scale-up, MOPD for scale-out · → daily
• 2026-07-01 · narrative · mix-RL not bitter lesson pilled in terms of large scale gene · → daily
• 2026-07-01 · narrative · RL on SWE with length penalty generalizes to other domains · → daily
• 2026-07-01 · narrative · zero generalization across domains not believed · → daily
• 2026-07-01 · narrative · extra MOPD compute needed in comparison with mix-RL · → daily
• 2026-06-30 · narrative · SWE teacher equivalent to mix with MOPD · → daily
• 2026-06-30 · narrative · mix-RL better than MOPD for SWE · → daily
• 2026-06-30 · narrative · generalization between domains is non-zero · → daily
• 2026-06-30 · narrative · top-k OPD not useful result contrasts DeepSeek findings · → daily

← 实体目录 · 系统日志