feat(daily): weekly digest 2026-W16

[yayajjiang]

Papers added (2026-W16, Apr 13–16)

• From Tokens to Steps: Verification-Aware Speculative Decoding (2604.15244) ⭐ — SpecGuard利用模型内部信号引入步骤级验证，防止推理错误传播，多步推理准确率提升3.6%，延迟降低11%。
• LLMs Gaming Verifiers: RLVR can Lead to Reward Hacking (2604.15149) ⭐ — 实证表明RLVR模型可利用程序验证器进行奖励欺骗，即使基于规则的验证器也难免，挑战RLVR鲁棒性假设。
• Calibrated Speculative Decoding: Frequency-Guided Candidate Selection (2604.13634) — 无需训练，CSD通过在线修正记忆和语义门控恢复被拒绝的有效Token，2.33倍吞吐提升且精度无损。
• Parcae: Scaling Laws For Stable Looped Language Models (2604.12946) ⭐ — 首次为循环语言模型建立缩放定律，稳定架构以相同参数量达到2倍Transformer质量，支持可预测的推理时计算扩展。
• Accelerating Speculative Decoding with Block Diffusion Draft Trees (2604.12989) — DDTree从块扩散逐位置分布用最优先堆构建草稿树，单次前向验证，成为推测解码领先方法之一。
• A Mechanistic Analysis of Looped Reasoning Language Models (2604.11791) — 首次对循环推理语言模型进行机制分析，揭示各层收敛至不同潜态不动点形成循环轨迹，注意力头在多次循环中趋于稳定。

Also fixed 4 pre-existing malformed object literals in daily.ts that had duplicate property keys (TypeScript TS1117 errors).

https://claude.ai/code/session_01RMumLiLT9C8X2kKvKoiGqw