feat(daily): 2026-04-22 digest

[yayajjiang]

Papers added (2026-04-22)

• KnowRL (2604.12627) — 将提示分解为最小充分原子知识点，约束子集搜索缓解RLVR奖励稀疏，8项推理基准准确率平均提升9.63点。⭐ Editor's Pick
• LLM Reasoning Is Latent, Not the Chain of Thought (2604.15726) — 形式化三种推理假设，现有证据最支持潜在状态轨迹为LLM推理真实载体，而非表面思维链，重塑基准设计方向。
• Demystifying the Unreasonable Effectiveness of Online Alignment Methods (2604.17207) — 证明在线RLHF和DPO的零温度遗憾界优于KL分析所示，从理论上解释在线对齐方法的强实证表现。

Also fixed

Two pre-existing malformed entries in daily.ts where two object literals were merged into one (duplicate property names caused a TypeScript build error). Removed the duplicate second halves, as those papers already appear as separate valid entries elsewhere in the array.

https://claude.ai/code/session_015EPed7aML4r9hjyCwtBieW