feat(daily): 2026-04-19 digest

[yayajjiang]

$(cat <<'EOF'

Papers added (2026-04-19)

• From P(y|x) to P(y): Investigating Reinforcement Learning in Pre-train Space (2604.14142) — 在预训练空间对边缘分布P(y)施加RL，双空间RL结合负样本强化与标准RLVR，持续拓展模型推理边界。⭐ Editor's Pick
• Parcae: Scaling Laws For Stable Looped Language Models (2604.12946) — 稳定循环架构以固定参数达到2倍大Transformer的87.5%质量，推导出参数高效语言模型的幂律FLOP缩放定律。
• A Mechanistic Analysis of Looped Reasoning Language Models (2604.11791) — 揭示循环LLM各层收敛至不同不动点，注意力头行为逐轮稳定，从机制层面解释循环为何提升推理能力。

Also fixed

• Pre-existing TypeScript errors in daily.ts (duplicate object keys from malformed entries)
• tsconfig.json: changed ignoreDeprecations from "6.0" to "5.0" to match installed TypeScript 5.9.3

https://claude.ai/code/session_0158MM7Pqei5atabQHd7AELS
EOF
)