Выпуск
頂尖模型只能靠砸資源硬爬嗎？Ricursive 自稱少試跑約 70 倍！靠的是什麼
脈報 · 5 ноября 2026 г. · 6 мин
Ricursive Intelligence 自述只用 1.4M 次 rollouts 就在 TerminalBench 2.1 拿到 80.5 分，用量比 Reflection Beam 少大約 70 倍。對照 Reflection 公告的 over 100 million rollouts，倍數規模兜得起來，80.5 分仍是團隊自報成績。 ⭐ 文章深度讀：頂尖模型只能靠砸資源嗎？Ricursive 自稱少試跑約 70 倍 → https://heymaibao.com/ricursive-fewer-rollouts-model-training-418d43/ ⚡ 章節重點 砸資源以外的路 00:00 少大約 70 倍的自述與對帳 00:30 團隊自述的省運算配方 01:54 訓練偏差的病根與解法 03:09 論文勝績與配方的界線 04:20 帶得走的類比與結論 05:49 📝 懶人包 ∙ 團隊自述只用 1.4M 次 rollouts 拿到 80.5 分 ∙ 團隊自述比 Reflection Beam 少大約 70 倍 ∙ 論文推導出 score centering，靠抵銷 drift 穩定 RL ∙ 我的觀點：1.4M 比 80.5 分更值得看，站得住的話砸資源不是唯一的路 📚 參考資料 Recursive Self-Improvement via AI for Chip Design & Chip Design for AI - Ricursive Intelligence → https://www.ricursive.com/ripple Score Centering Stabilizes Off-policy Reinforcement Learning → https://arxiv.org/abs/2609.20807 Introducing Ripple-65B Preview → https://www.ricursive.com/ Terminal-Bench 2.1 → https://www.tbench.ai/news/terminal-bench-2-1 Introducing Beam: Reflection’s 501B open-weight model → https://reflection.ai/blog/introducing-beam How an LLM RL System Fits Together → https://ai-systems-engineering.com/en/posts/rl-system-architecture/ Rethinking Training–Inference Mismatch inLLM Reinforcement Learning:Where It Arises and How to Correct It → https://arxiv.org/abs/2609.32444 On the Origin and Evolution of Training-Inference Mismatch in LLM Reinforcement Learning → https://arxiv.org/abs/2610.40945