feat(daily): 2026-04-17 digest

[yayajjiang]

Papers added (2026-04-17)

• From P(y|x) to P(y): Investigating Reinforcement Learning in Pre-train Space (2604.14142) — PreRL在预训练空间扩展推理边界，NSR快速剪枝错误路径，DSRL全面超越强基线。⭐ Editor's pick
• LongCoT: Benchmarking Long-Horizon Chain-of-Thought Reasoning (2604.14140) — 2500道专家题需推理10万token，最优模型不足10%，长链推理缺口巨大。
• Calibration-Aware Policy Optimization for Reasoning LLMs (2604.12632) — CAPO通过AUC替代损失修正GRPO过度自信，同时优化准确率与不确定性校准，ACL 2026录用。

Fixes

• Resolved pre-existing TypeScript errors: 4 object literals had duplicate property keys from accidentally merged entries — split into proper standalone entries.
• Fixed tsconfig.json: ignoreDeprecations changed from "6.0" to "5.0" (TypeScript 5.9.3 does not recognise "6.0").

https://claude.ai/code/session_01S9HrHYJFWbhKJqq7jCNFDA