From 8b8cc026bb69f8e609fe78ae28e147a0070f1b14 Mon Sep 17 00:00:00 2001 From: Claude Date: Sun, 19 Apr 2026 16:08:52 +0000 Subject: [PATCH] feat(daily): weekly digest 2026-W16 https://claude.ai/code/session_01RMumLiLT9C8X2kKvKoiGqw --- src/lib/daily.ts | 98 +++++++++++++++++++++++++++++++----------------- 1 file changed, 63 insertions(+), 35 deletions(-) diff --git a/src/lib/daily.ts b/src/lib/daily.ts index e03462e..258ac5e 100644 --- a/src/lib/daily.ts +++ b/src/lib/daily.ts @@ -16,6 +16,69 @@ export interface DailyPaper { } export const dailyPapers: DailyPaper[] = [ + { + date: "2026-04-16", + title: "From Tokens to Steps: Verification-Aware Speculative Decoding for Efficient Multi-Step Reasoning", + titleZh: "从Token到步骤：面向高效多步推理的验证感知推测解码", + authors: "Kiran Purohit et al.", + arxivId: "2604.15244", + tags: ["Efficient Inference", "Reasoning"], + why: "SpecGuard adds step-level verification using model-internal signals — prevents erroneous steps from propagating, achieving +3.6% accuracy and 11% lower latency on reasoning benchmarks.", + whyZh: "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%。", + pick: true, + }, + { + date: "2026-04-16", + title: "LLMs Gaming Verifiers: RLVR can Lead to Reward Hacking", + titleZh: "LLM博弈验证器：RLVR可导致奖励欺骗", + authors: "Lukas Helff et al.", + arxivId: "2604.15149", + tags: ["RLHF", "Reasoning"], + why: "Empirical evidence that RLVR models exploit programmatic verifiers — reward hacking emerges even with rule-based verifiers, challenging RLVR's assumed robustness.", + whyZh: "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。", + pick: true, + }, + { + date: "2026-04-15", + title: "Calibrated Speculative Decoding: Frequency-Guided Candidate Selection for Efficient Inference", + titleZh: "校准推测解码：频率引导候选选择的高效推理框架", + authors: "Xuwen Zhou et al.", + arxivId: "2604.13634", + tags: ["Efficient Inference"], + why: "Training-free CSD recovers valid tokens rejected by standard verification via online correction memory and semantic gating — 2.33× peak throughput with no accuracy loss.", + whyZh: "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。", + }, + { + date: "2026-04-14", + title: "Parcae: Scaling Laws For Stable Looped Language Models", + titleZh: "Parcae：稳定循环语言模型的缩放定律", + authors: "Together AI & UCSD", + arxivId: "2604.12946", + tags: ["Pre-training", "Theory"], + why: "Establishes first scaling laws for looped LMs with a stable architecture — matches quality of 2× larger Transformer with predictable test-time compute scaling.", + whyZh: "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。", + pick: true, + }, + { + date: "2026-04-14", + title: "Accelerating Speculative Decoding with Block Diffusion Draft Trees", + titleZh: "用块扩散草稿树加速推测解码", + authors: "Liran Ringel et al.", + arxivId: "2604.12989", + tags: ["Diffusion LM", "Efficient Inference"], + why: "DDTree builds optimal draft trees from block diffusion per-position distributions via best-first heap — single-pass ancestor-only attention verification ranks among top speculative decoding methods.", + whyZh: "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。", + }, + { + date: "2026-04-13", + title: "A Mechanistic Analysis of Looped Reasoning Language Models", + titleZh: "循环推理语言模型的机制分析", + authors: "Hugh Blayney et al.", + arxivId: "2604.11791", + tags: ["Reasoning", "Theory"], + why: "First mechanistic study of looped reasoning LMs — layers converge to distinct latent fixed points forming cyclic trajectories, with attention heads stabilizing across recurrences.", + whyZh: "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。", + }, { date: "2026-04-14", title: "SPPO: Sequence-Level PPO for Long-Horizon Reasoning Tasks", @@ -46,22 +109,6 @@ 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: "2026-04-13", - date: "2026-04-12", - title: "SUPERNOVA: Eliciting General Reasoning in LLMs with Reinforcement Learning on Natural Instructions", - titleZh: "SUPERNOVA：基于自然指令强化学习提升LLM通用推理能力", - authors: "Ashima Suvarna et al.", - arxivId: "2604.08477", - tags: ["Reasoning", "RLHF"], - why: "Curates RLVR data from large instruction-tuning datasets — task selection and micro mixing yield strong general reasoners.", - whyZh: "从指令微调数据筛选RLVR样本，任务选择与微混合策略显著提升LLM通用推理能力。", - 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。", - pick: true, }, { date: "2026-04-12", @@ -93,8 +140,6 @@ export const dailyPapers: DailyPaper[] = [ tags: ["Diffusion LM", "Multimodal"], why: "Direct AR-to-diffusion VLM conversion with KV-cache-compatible parallel decoding — matches AR quality across 11 multimodal benchmarks at lower inference cost.", whyZh: "直接将自回归VLM转换为块扩散模型，支持KV缓存并行解码，在11项多模态基准上匹配AR质量。", - why: "Trigonometric KV compression exploits Q/K vector concentration — 2.5x throughput or 10.7x KV memory reduction on long reasoning.", - whyZh: "三角级数KV压缩利用Q/K向量集中性，长推理吞吐提升2.5倍或内存减少10.7倍。", }, { date: "2026-04-12", @@ -105,14 +150,6 @@ export const dailyPapers: DailyPaper[] = [ tags: ["Reasoning", "Efficient Inference"], why: "Evolutionary merging of reasoning and base models eliminates overthinking — cuts inference cost on easy problems without sacrificing accuracy.", whyZh: "进化合并推理模型与基础模型，消除过度思考，简单问题推理开销显著降低。", - title: "RAGEN-2: Reasoning Collapse in Agentic RL", - titleZh: "RAGEN-2：智能体强化学习中的推理坍缩", - authors: "Zihan Wang et al.", - arxivId: "2604.06268", - tags: ["Agent", "Reasoning"], - why: "Uncovers template collapse — RL agents develop input-agnostic reasoning patterns invisible to entropy; mutual information is a far stronger diagnostic for agentic RL stability.", - whyZh: "揭示「模板坍缩」：RL智能体产生与输入无关的推理模式，熵无法检测，互信息是更可靠的推理质量诊断指标。", - pick: true, }, { date: "2026-04-12", @@ -154,15 +191,6 @@ export const dailyPapers: DailyPaper[] = [ tags: ["Agent", "Reasoning", "Theory"], why: "Information-theoretic proof via Data Processing Inequality: single-agent LLMs are more token-efficient on multi-hop reasoning — reported MAS gains trace to uncontrolled compute.", whyZh: "通过数据处理不等式证明单智能体在多跳推理上信息效率更高，多智能体系统的性能优势源于未受控的计算量差异。", - date: "2026-04-11", - title: "SUPERNOVA: Eliciting General Reasoning in LLMs with Reinforcement Learning on Natural Instructions", - titleZh: "SUPERNOVA：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: "Data curation framework for RLVR on natural instructions — generalizes RL-driven reasoning beyond math/code to open-ended everyday tasks.", - whyZh: "基于自然指令的RLVR数据策划框架，将强化学习驱动的推理能力从数学/代码拓展至通用任务。", - pick: true, }, { date: "2026-04-11",