SPPO

5 atoms · 跨 1 天 · 首见 2026-04-15 · 最近 2026-04-15

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

来源: X 5

时态: fresh:5

标签: 好数字:3 · 好思考:1 · 好信源:1

事实 (fact) (5)

• 2026-04-15 · speedup over GRPO
  SPPO: 5.9× speedup over GRPO
  tags: 好数字 · → daily
  value: qty=5.9× · direction=faster
• 2026-04-15 · reformulates reasoning as Sequence-Level Contextual Bandit
  Reformulates reasoning as a Sequence-Level Contextual Bandit, enabling stable single-sample updates with 5.9× speedup over GRPO.
  tags: 好思考 · → daily
• 2026-04-15 · matches GRPO performance on AIME24/25, AMC23, and MATH500 at 1.5B and 7B scales
  Matches GRPO performance on AIME24/25, AMC23, and MATH500 at 1.5B and 7B scales.
  tags: 好数字 · → daily
• 2026-04-15 · decoupled small critic architecture VRAM reduction
  Decoupled small critic architecture allows a 1.5B model to align a 7B policy, cutting VRAM by 12.8%.
  tags: 好数字 · → daily
  value: qty=12.8% · direction=reduced
• 2026-04-15 · paper accepted to ACL 2026
  Accepted to ACL 2026.
  tags: 好信源 · → daily

时间轴 (近 20)

• 2026-04-15 · fact · speedup over GRPO · → daily
• 2026-04-15 · fact · reformulates reasoning as Sequence-Level Contextual Bandit · → daily
• 2026-04-15 · fact · matches GRPO performance on AIME24/25, AMC23, and MATH500 at · → daily
• 2026-04-15 · fact · decoupled small critic architecture VRAM reduction · → daily
• 2026-04-15 · fact · paper accepted to ACL 2026 · → daily

← 实体目录 · 系统日志