<?xml version="1.0" encoding="UTF-8"?><rss version="2.0"><channel><title>HelloAI · 论文精读</title><description>Attention Is All You Need · ResNet · Diffusion · LoRA · Scaling Laws...</description><link>https://ai.xwebgame.com/</link><language>zh-cn</language><item><title>Scaling Laws for Neural Language Models</title><link>https://ai.xwebgame.com/papers/scaling-laws/</link><guid isPermaLink="true">https://ai.xwebgame.com/papers/scaling-laws/</guid><description>OpenAI 2020 年的奠基性发现——&quot;模型损失随参数、数据、算力呈幂律下降&quot;。这条曲线是 GPT-3、GPT-4 等大模型投资的理论基础。</description><pubDate>Fri, 28 Aug 2026 00:00:00 GMT</pubDate><category>Scaling Laws</category><category>理论</category><category>OpenAI</category><category>必读</category></item><item><title>LoRA: Low-Rank Adaptation of Large Language Models</title><link>https://ai.xwebgame.com/papers/lora/</link><guid isPermaLink="true">https://ai.xwebgame.com/papers/lora/</guid><description>Microsoft 提出 LoRA—只训 0.01% 参数 + 不损失性能 = 让&quot;消费级 GPU 微调大模型&quot;成为可能。开源 LLM 微调生态的关键技术。</description><pubDate>Thu, 27 Aug 2026 00:00:00 GMT</pubDate><category>LoRA</category><category>微调</category><category>PEFT</category><category>必读</category></item><item><title>Training Compute-Optimal Large Language Models (Chinchilla)</title><link>https://ai.xwebgame.com/papers/chinchilla/</link><guid isPermaLink="true">https://ai.xwebgame.com/papers/chinchilla/</guid><description>DeepMind 证明 GPT-3 等大模型&quot;参数太多、数据太少&quot;。给出了&quot;算力如何在参数和数据间最优分配&quot;的新法则——重塑了大模型训练。</description><pubDate>Wed, 26 Aug 2026 00:00:00 GMT</pubDate><category>Chinchilla</category><category>Scaling Laws</category><category>训练</category><category>必读</category></item><item><title>Direct Preference Optimization (DPO)</title><link>https://ai.xwebgame.com/papers/dpo/</link><guid isPermaLink="true">https://ai.xwebgame.com/papers/dpo/</guid><description>把 RLHF 简化成一个简单的损失函数——跳过奖励模型和 PPO，效果接近，工程简单 10 倍。开源 LLM 对齐的事实标准。</description><pubDate>Tue, 25 Aug 2026 00:00:00 GMT</pubDate><category>DPO</category><category>RLHF</category><category>对齐</category><category>必读</category></item><item><title>The Pile: An 800GB Dataset of Diverse Text for Language Modeling</title><link>https://ai.xwebgame.com/papers/the-pile/</link><guid isPermaLink="true">https://ai.xwebgame.com/papers/the-pile/</guid><description>EleutherAI 开源的 800GB 训练数据集——第一个真正可用的&quot;GPT-3 级别&quot;开源训练数据。开源 LLM 革命的&quot;砖头&quot;。</description><pubDate>Fri, 21 Aug 2026 00:00:00 GMT</pubDate><category>The Pile</category><category>数据集</category><category>开源</category><category>基础</category></item><item><title>Visual Instruction Tuning (LLaVA)</title><link>https://ai.xwebgame.com/papers/llava/</link><guid isPermaLink="true">https://ai.xwebgame.com/papers/llava/</guid><description>把 CLIP + LLaMA + 指令微调 缝合起来——开源多模态指令模型的起点。让&quot;图像+对话&quot;AI 进入开源社区。</description><pubDate>Thu, 20 Aug 2026 00:00:00 GMT</pubDate><category>LLaVA</category><category>多模态</category><category>开源</category><category>指令微调</category></item><item><title>Transformers are SSMs: Generalized Models and Efficient Algorithms (Mamba 2)</title><link>https://ai.xwebgame.com/papers/mamba-2/</link><guid isPermaLink="true">https://ai.xwebgame.com/papers/mamba-2/</guid><description>Mamba 团队的反击——证明 Transformer 和 SSM 在数学上等价，并提出更快的 Mamba 2 架构。SSM 路线的关键升级。</description><pubDate>Wed, 19 Aug 2026 00:00:00 GMT</pubDate><category>Mamba</category><category>SSM</category><category>架构</category><category>前沿</category></item><item><title>TruthfulQA: Measuring How Models Mimic Human Falsehoods</title><link>https://ai.xwebgame.com/papers/truthfulqa/</link><guid isPermaLink="true">https://ai.xwebgame.com/papers/truthfulqa/</guid><description>一个测 LLM &quot;是否真实&quot;的 benchmark。第一次系统揭示：模型越大，反而在某些常见误区上越错。</description><pubDate>Tue, 18 Aug 2026 00:00:00 GMT</pubDate><category>TruthfulQA</category><category>评估</category><category>Benchmark</category><category>幻觉</category></item><item><title>Gemini: A Family of Highly Capable Multimodal Models</title><link>https://ai.xwebgame.com/papers/gemini/</link><guid isPermaLink="true">https://ai.xwebgame.com/papers/gemini/</guid><description>Google 用 6 年时间 + 1 万张 TPU 训出的&quot;原生多模态&quot;大模型。1M+ 上下文窗口，是 GPT-4 的最大挑战者之一。</description><pubDate>Fri, 14 Aug 2026 00:00:00 GMT</pubDate><category>Gemini</category><category>Google</category><category>多模态</category><category>大模型</category></item><item><title>Segment Anything (SAM)</title><link>https://ai.xwebgame.com/papers/sam/</link><guid isPermaLink="true">https://ai.xwebgame.com/papers/sam/</guid><description>Meta 的&quot;图像分割基础模型&quot;——点一下就能分割任何物体。开源 + 1100 万张图 + 1 亿 mask，让&quot;通用分割&quot;成为现实。</description><pubDate>Thu, 13 Aug 2026 00:00:00 GMT</pubDate><category>SAM</category><category>分割</category><category>视觉</category><category>基础模型</category><category>必读</category></item><item><title>Robust Speech Recognition via Large-Scale Weak Supervision (Whisper)</title><link>https://ai.xwebgame.com/papers/whisper/</link><guid isPermaLink="true">https://ai.xwebgame.com/papers/whisper/</guid><description>OpenAI 用 68 万小时弱监督音频训出最强 ASR。开源后统治整个开源语音识别市场。99 种语言通吃。</description><pubDate>Wed, 12 Aug 2026 00:00:00 GMT</pubDate><category>Whisper</category><category>ASR</category><category>语音</category><category>开源</category><category>必读</category></item><item><title>DeepSeek-V3 / R1：开源推理模型的革命</title><link>https://ai.xwebgame.com/papers/deepseek/</link><guid isPermaLink="true">https://ai.xwebgame.com/papers/deepseek/</guid><description>DeepSeek 用 $5.6M 训出接近 GPT-4 的开源模型——震动了整个行业。证明&quot;开源 + 高效工程 + 创新算法&quot; 能挑战美国巨头。</description><pubDate>Tue, 11 Aug 2026 00:00:00 GMT</pubDate><category>DeepSeek</category><category>开源</category><category>推理</category><category>前沿</category><category>必读</category></item><item><title>Phi-3 Technical Report: A Highly Capable Language Model Locally on Your Phone</title><link>https://ai.xwebgame.com/papers/phi-3/</link><guid isPermaLink="true">https://ai.xwebgame.com/papers/phi-3/</guid><description>Phi-3 mini 仅 3.8B 参数——但在多项 benchmark 上接近 GPT-3.5。证明了&quot;小模型 + 极致数据质量&quot;是另一条路。</description><pubDate>Fri, 07 Aug 2026 00:00:00 GMT</pubDate><category>Phi-3</category><category>小模型</category><category>数据质量</category><category>前沿</category></item><item><title>The Llama 3 Herd of Models</title><link>https://ai.xwebgame.com/papers/llama-3/</link><guid isPermaLink="true">https://ai.xwebgame.com/papers/llama-3/</guid><description>Meta 公开了 Llama 3 405B 的完整训练细节——开源模型首次达到 GPT-4 级别。92 页技术报告揭秘大模型训练的工程实战。</description><pubDate>Thu, 06 Aug 2026 00:00:00 GMT</pubDate><category>Llama</category><category>开源</category><category>大模型</category><category>必读</category></item><item><title>Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training</title><link>https://ai.xwebgame.com/papers/sleeper-agents/</link><guid isPermaLink="true">https://ai.xwebgame.com/papers/sleeper-agents/</guid><description>Anthropic 证明：可以训练一个&quot;装好的&quot;AI——表面对齐，遇到特定触发词激活恶意行为。而且当前所有对齐方法都检测不出来。</description><pubDate>Wed, 05 Aug 2026 00:00:00 GMT</pubDate><category>Sleeper Agents</category><category>对齐</category><category>AI 安全</category><category>警告</category></item><item><title>Learning to Reason with LLMs (OpenAI o1)</title><link>https://ai.xwebgame.com/papers/openai-o1/</link><guid isPermaLink="true">https://ai.xwebgame.com/papers/openai-o1/</guid><description>推理时计算的范式转变——让 LLM 在回答前花更多时间&quot;思考&quot;，复杂问题准确率从 20% 升到 80%。开启了&quot;推理模型&quot;时代。</description><pubDate>Tue, 04 Aug 2026 00:00:00 GMT</pubDate><category>o1</category><category>Reasoning</category><category>CoT</category><category>前沿</category><category>必读</category></item><item><title>Video generation models as world simulators (Sora)</title><link>https://ai.xwebgame.com/papers/sora/</link><guid isPermaLink="true">https://ai.xwebgame.com/papers/sora/</guid><description>OpenAI 的视频生成模型 Sora——把视频切成&quot;时空 patch&quot;用 Transformer 做扩散。1 分钟高质量视频成为可能，&quot;AI 世界模拟器&quot;露端倪。</description><pubDate>Fri, 31 Jul 2026 00:00:00 GMT</pubDate><category>Sora</category><category>视频生成</category><category>Diffusion</category><category>Transformer</category><category>前沿</category></item><item><title>FlashAttention: Fast and Memory-Efficient Exact Attention with IO-Awareness</title><link>https://ai.xwebgame.com/papers/flash-attention/</link><guid isPermaLink="true">https://ai.xwebgame.com/papers/flash-attention/</guid><description>通过感知 GPU 内存层级，让注意力计算快 2-4 倍 + 显存少 10 倍——而且数学上完全相同。所有现代 LLM 都用它。</description><pubDate>Thu, 30 Jul 2026 00:00:00 GMT</pubDate><category>FlashAttention</category><category>GPU</category><category>系统优化</category><category>必读</category></item><item><title>Constitutional AI: Harmlessness from AI Feedback</title><link>https://ai.xwebgame.com/papers/constitutional-ai/</link><guid isPermaLink="true">https://ai.xwebgame.com/papers/constitutional-ai/</guid><description>Anthropic 提出的对齐新方法——让 AI 用&quot;宪法原则&quot;自评自改，跳过大量人类标注。Claude 的核心训练秘密。</description><pubDate>Wed, 29 Jul 2026 00:00:00 GMT</pubDate><category>Constitutional AI</category><category>对齐</category><category>Anthropic</category><category>必读</category></item><item><title>Mamba: Linear-Time Sequence Modeling with Selective State Spaces</title><link>https://ai.xwebgame.com/papers/mamba/</link><guid isPermaLink="true">https://ai.xwebgame.com/papers/mamba/</guid><description>挑战 Transformer 霸权的&quot;选择性状态空间模型&quot;——线性复杂度处理超长序列，理论上能取代 Transformer。2024 年最热的架构研究之一。</description><pubDate>Tue, 28 Jul 2026 00:00:00 GMT</pubDate><category>Mamba</category><category>SSM</category><category>架构</category><category>前沿</category></item><item><title>Training language models to follow instructions with human feedback (InstructGPT)</title><link>https://ai.xwebgame.com/papers/instruct-gpt/</link><guid isPermaLink="true">https://ai.xwebgame.com/papers/instruct-gpt/</guid><description>从 GPT-3 到 ChatGPT 的&quot;桥梁&quot;。提出 SFT + RLHF 三阶段训练让 LLM &quot;听话&quot;——这套流程定义了之后所有商业 LLM 的训练范式。</description><pubDate>Fri, 24 Jul 2026 00:00:00 GMT</pubDate><category>RLHF</category><category>InstructGPT</category><category>对齐</category><category>ChatGPT</category><category>必读</category></item><item><title>Highly Accurate Protein Structure Prediction with AlphaFold</title><link>https://ai.xwebgame.com/papers/alphafold2/</link><guid isPermaLink="true">https://ai.xwebgame.com/papers/alphafold2/</guid><description>DeepMind 用 Transformer 解决了 50 年的&quot;蛋白质折叠&quot;问题。预测了所有已知生物的 2 亿个蛋白质结构。2024 年诺贝尔化学奖。</description><pubDate>Thu, 23 Jul 2026 00:00:00 GMT</pubDate><category>AlphaFold</category><category>AI for Science</category><category>蛋白质</category><category>诺贝尔奖</category><category>必读</category></item><item><title>BERT: Pre-training of Deep Bidirectional Transformers</title><link>https://ai.xwebgame.com/papers/bert/</link><guid isPermaLink="true">https://ai.xwebgame.com/papers/bert/</guid><description>2018 年的 NLP 核爆。提出 Masked Language Modeling + 双向 Transformer，让&quot;预训练 + 微调&quot;成为 NLP 主流范式。</description><pubDate>Wed, 22 Jul 2026 00:00:00 GMT</pubDate><category>BERT</category><category>NLP</category><category>预训练</category><category>必读</category></item><item><title>Learning Transferable Visual Models From Natural Language Supervision (CLIP)</title><link>https://ai.xwebgame.com/papers/clip/</link><guid isPermaLink="true">https://ai.xwebgame.com/papers/clip/</guid><description>用 4 亿张&quot;图 + 描述&quot;对训练——让图像 encoder 和文本 encoder 在同一向量空间对齐。从此 AI 能&quot;看图说话&quot;，&quot;看图作画&quot;。</description><pubDate>Tue, 21 Jul 2026 00:00:00 GMT</pubDate><category>CLIP</category><category>多模态</category><category>对比学习</category><category>必读</category></item><item><title>Attention Is All You Need</title><link>https://ai.xwebgame.com/papers/attention-is-all-you-need/</link><guid isPermaLink="true">https://ai.xwebgame.com/papers/attention-is-all-you-need/</guid><description>提出 Transformer 架构——完全抛弃 RNN，只用注意力机制。这篇 8 页的论文催生了今天所有大模型。被引 12 万+。</description><pubDate>Fri, 17 Jul 2026 00:00:00 GMT</pubDate><category>Transformer</category><category>Attention</category><category>NLP</category><category>必读</category></item><item><title>Denoising Diffusion Probabilistic Models (DDPM)</title><link>https://ai.xwebgame.com/papers/ddpm/</link><guid isPermaLink="true">https://ai.xwebgame.com/papers/ddpm/</guid><description>提出 DDPM —— 用&quot;加噪 → 去噪&quot;的范式做图像生成。Stable Diffusion、Sora 都基于这个思路。</description><pubDate>Fri, 17 Jul 2026 00:00:00 GMT</pubDate><category>Diffusion</category><category>生成模型</category><category>图像</category><category>必读</category></item><item><title>Language Models are Few-Shot Learners (GPT-3)</title><link>https://ai.xwebgame.com/papers/gpt-3/</link><guid isPermaLink="true">https://ai.xwebgame.com/papers/gpt-3/</guid><description>175B 参数的 GPT-3 展示了&quot;in-context learning&quot;——不微调，只给几个例子就能学会新任务。这篇论文重新定义了人们对 LLM 的预期。</description><pubDate>Fri, 17 Jul 2026 00:00:00 GMT</pubDate><category>GPT-3</category><category>LLM</category><category>Few-shot</category><category>必读</category></item><item><title>Deep Residual Learning for Image Recognition</title><link>https://ai.xwebgame.com/papers/resnet/</link><guid isPermaLink="true">https://ai.xwebgame.com/papers/resnet/</guid><description>提出残差连接（skip connection），让神经网络能训到 100+ 层。CVPR 2016 最佳论文，引用 25 万+，至今所有大模型仍在用这个技巧。</description><pubDate>Fri, 17 Jul 2026 00:00:00 GMT</pubDate><category>CNN</category><category>ResNet</category><category>残差连接</category><category>视觉</category><category>必读</category></item></channel></rss>