feat(daily): weekly digest 2026-W19

[yayajjiang]

Weekly Digest — 2026-W19 (May 4–10)

6 new papers added to src/lib/daily.ts (3 editor's picks ⭐):

• ⭐ Rethinking RL for LLM Reasoning: It's Sparse Policy Selection, Not Capability Learning (2605.06241) — RL不教推理新技巧，仅在1-3%高熵决策点稀疏调整概率，从根本上重构了对RL后训练的认知。
• Can RL Teach Long-Horizon Reasoning to LLMs? Expressiveness Is Key (2605.06638) — ScaleLogic分离推理深度与逻辑表达力，指数由表达力决定（1.04→2.60），揭示长程推理泛化关键。
• RVPO: Risk-Sensitive Alignment via Variance Regularization (2605.05750) — RVPO惩罚优势聚合中的跨奖励方差，修复多目标RLHF的均值聚合缺陷，17个并发奖励信号验证。
• ⭐ A Unified Pair-GRPO Family (2605.06375) — Soft/Hard Pair-GRPO统一隐式与显式偏好约束，解决GRPO类方法训练不稳定问题。
• Federation of Experts: Communication Efficient Distributed Inference (2605.06206) — FoE将MoE模块重组为专家集群，切断跨节点token嵌入通信瓶颈。
• ⭐ Making Every Verified Token Count: Adaptive Verification for MoE Speculative Decoding (2605.00342) — EVICT无需训练实现专家感知草稿树剪枝，比AR快2.35倍、比EAGLE-3快1.21倍。

https://claude.ai/code/session_015ootf94u65YKT9eXtiMwVz

---

Generated by Claude Code