feat(daily): 2026-04-23 digest

src/lib/daily.ts

@@ -16,6 +16,37 @@ export interface DailyPaper {
 }
 
 export const dailyPapers: DailyPaper[] = [
+  {
+    date: "2026-04-23",
+    title: "LLaDA2.0-Uni: Unifying Multimodal Understanding and Generation with Diffusion Large Language Model",
+    titleZh: "LLaDA2.0-Uni：基于扩散大语言模型的多模态理解与生成统一框架",
+    authors: "Inclusion AI et al.",
+    arxivId: "2604.20796",
+    tags: ["Diffusion LM", "Multimodal", "MoE"],
+    why: "LLaDA 2.0 goes multimodal — MoE dLLM unifies image understanding and generation via mask token prediction, matching dedicated VLMs with 8-step distilled inference.",
+    whyZh: "LLaDA 2.0进化为多模态：MoE扩散LLM通过掩码token预测统一图像理解与生成，8步蒸馏推理媲美专用VLM。",
+    pick: true,
+  },
+  {
+    date: "2026-04-23",
+    title: "Self-Consistency from Only Two Samples: CoT-PoT Ensembling for Efficient LLM Reasoning",
+    titleZh: "仅需两次采样的自一致性：面向高效LLM推理的CoT-PoT集成方法",
+    authors: "Raman Saparkhan et al.",
+    arxivId: "2604.17433",
+    tags: ["Reasoning", "Efficient Inference"],
+    why: "Combining chain-of-thought and program-of-thought cuts self-consistency sampling 9.3x — strong reasoning accuracy with just two samples.",
+    whyZh: "结合思维链与程序化思维，将自一致性采样需求降低9.3倍，仅需两次采样即可获得强推理准确率。",
+  },
+  {
+    date: "2026-04-23",
+    title: "KnowRL: Boosting LLM Reasoning via Reinforcement Learning with Minimal-Sufficient Knowledge Guidance",
+    titleZh: "KnowRL：基于最小充分知识引导的强化学习提升LLM推理",
+    authors: "Linhao Yu et al.",
+    arxivId: "2604.12627",
+    tags: ["Reasoning", "RLHF"],
+    why: "Decomposes hints into atomic knowledge points and finds minimal sufficient subsets — fixes reward sparsity in RLVR without adding token overhead.",
+    whyZh: "将提示分解为原子知识点并搜索最小充分子集，解决RLVR中奖励稀疏问题，不增加token开销。",
+  },
   {
     date: "2026-04-16",
     title: "Introspective Diffusion Language Models",
@@ -36,15 +67,6 @@ export const dailyPapers: DailyPaper[] = [
     tags: ["Diffusion LM", "Efficient Inference"],
     why: "DDTree builds a best-first draft tree from block diffusion per-position distributions — SOTA speculative decoding verified in one target model forward pass.",
     whyZh: "DDTree从块扩散逐位置分布构建最优优先草稿树，单次目标模型前向传播完成验证，达到推测解码SOTA。",
-    date: "2026-04-15",
-    title: "Introspective Diffusion Language Models",
-    titleZh: "内省扩散语言模型",
-    authors: "Yifan Yu et al.",
-    arxivId: "2604.11035",
-    tags: ["Diffusion LM", "Reasoning", "Efficient Inference"],
-    why: "Introspective strided decoding fixes diffusion LM's consistency gap — I-DLM-8B beats LLaDA-2.1-mini (16B) at 2.9–4.1x higher throughput.",
-    whyZh: "内省跨步解码修复扩散LM一致性缺陷，I-DLM-8B以2.9-4.1倍吞吐超越更大规模LLaDA。",
-    pick: true,
   },
   {
     date: "2026-04-15",
@@ -96,14 +118,6 @@ export const dailyPapers: DailyPaper[] = [
     tags: ["Diffusion LM", "Efficient Inference"],
     why: "Block diffusion drafter achieves 6x lossless speedup over base LLM — 2.5x faster than EAGLE-3 with no quality loss.",
     whyZh: "用块扩散模型作为推测解码草稿器，实现6倍无损加速，比EAGLE-3快2.5倍。",
-    date: "2026-04-12",
-    title: "SUPERNOVA: Eliciting General Reasoning in LLMs with Reinforcement Learning on Natural Instructions",
-    titleZh: "SUPERNOVA：利用自然指令强化学习激发LLM通用推理能力",
-    authors: "Ashima Suvarna et al.",
-    arxivId: "2604.08477",
-    tags: ["Reasoning", "RLHF"],
-    why: "Mines natural instruction datasets for verifiable rewards — extends RLVR beyond math/code to causal, temporal, and abductive reasoning without hand-crafted reward functions.",
-    whyZh: "从自然指令数据集中挖掘可验证奖励，将RLVR扩展至因果、时序和溯因推理，无需手工设计奖励函数。",
   },
   {
     date: "2026-04-12",