DAS 加速 RL 滾動高達 50%
RL 後訓練中的滾動是主要瓶頸。Together AI 的分布感知推測解碼 (DAS) 使用自適應推測解碼，將滾動加速高達 50%。獎勵品質零衰減。
Search
Tag: #rl8 results
vLLM V0 到 V1：在 RL 中正確性優先於修正
Hugging Face 部落格探討 vLLM 從 V0 到 V1 的演進，在強化學習工作流程中優先考慮正確性而非修正。文章強調建構可靠 LLM 伺服系統的關鍵原則。提供平衡準確性和效率在 RL 應用中的洞見。
代理逐一繪製單一部件
研究人員開發多模態 LLM 代理，使用監督微調後的多輪過程獎勵 RL，逐部件生成向量草圖。他們透過自動分割與標註管線，建立具豐富部件級註解的 ControlSketch-Part 資料集。此方法透過視覺回饋實現可解釋、可控制且可局部編輯的文字轉向量草圖生成。
🔬
PyTorch RL 自訂演算法最佳實作
Reddit 貼文尋求 PyTorch 自訂 RL 演算法資源、在 Gym 基準測試對比基線。詢問程式碼優化、目錄結構、Docker、Mac/Linux 相容。
MAPLE Boosts Multimodal RL Post-Training
MAPLE is a modality-aware ecosystem for post-training multimodal LLMs, including MAPLE-bench, MAPO optimization, and adaptive curricula. It stratifies training by modality needs to cut variance and speed convergence. It closes uni/multi-modal gaps by 30% and converges 3x faster.
VESPO Stabilizes Off-Policy LLM Training
VESPO introduces variational sequence-level soft policy optimization to tackle training instability in RL for LLMs caused by policy staleness and async execution. It derives a closed-form reshaping kernel for importance weights without length normalization. Experiments demonstrate stable training up to 64x staleness on math benchmarks.
V-STAR: Value-Guided RecSys Sampling
V-STAR addresses probability-reward mismatch in generative recsys via value-guided decoding and sibling-relative RL. VED efficiently explores high-potential prefixes; Sibling-GRPO focuses on decisive branches. Outperforms baselines in accuracy and diversity.
OmniSapiens: HARPO for Social Behaviors
OmniSapiens-7B 2.0 uses HARPO RL to train a unified model across heterogeneous social tasks. HARPO balances learning via modulated advantages. It outperforms baselines by up to 16.85% with robust reasoning.