feat(daily): weekly digest 2026-W20

[yayajjiang]

Weekly Digest 2026-W20 (May 11–17)

Week theme: Diffusion LMs enter the RL post-training era — plus new theory on what RL actually does to LLMs, and Apple's information-theoretic limits for alignment.

Papers added (6 entries, 3 picks)

• Beyond Reasoning: RL Unlocks Parametric Knowledge in LLMs (2605.07153) ⭐ — RL在无CoT零样本问答中提升约27%事实召回率，机制是重分配已有知识概率质量而非习得新知，有力挑战「强化学习=推理」叙事。
• Theoretical Limits of Language Model Alignment (2605.07105) ⭐ — Apple推导对齐的信息论KL-奖励帕累托前沿，best-of-N接近理论最优，PPO/GRPO仍大幅低于极限，为对齐研究提供理论锚点。
• Block-R1: Block Size in Multi-domain RL for Diffusion LLMs (2605.11726) ⭐ — 块尺寸冲突是多领域dLLM RL后训练的关键瓶颈，Block-R1-41K为每个样本分配最优块尺寸，有效缓解GRPO跨领域训练张力。
• Break the Block: Dynamic-size Reasoning Blocks for Diffusion LLMs (2605.02263) — 固定块尺寸破坏扩散LLM推理连贯性，单调熵下降RL框架自适应学习块边界，在多类推理任务中显著提升生成质量。
• Continuous Latent Diffusion Language Model (Cola DLM) (2605.06548) — 分层Text VAE与块因果DiT在连续潜空间中建模文本，彻底绕开离散token瓶颈，为扩散语言模型开辟全新范式。
• LatentRAG: Latent Reasoning and Retrieval for Efficient Agentic RAG (2605.06285) — 在连续潜空间中完成智能体RAG的推理与检索，完全替代逐token生成，7项基准上匹配显式推理精度，延迟降低约90%。

https://claude.ai/code/session_01DUXCCYevfJ8ALPNqk7sdoZ

---

Generated by Claude Code