⚡
20171 篇
20183 篇
2019 H12 篇
2019 H24 篇
2020 H15 篇
2020 H21 篇
2021 H14 篇
2021 H22 篇
2022 H18 篇
2022 H27 篇
2023 H117 篇
2023 H212 篇
2024 H113 篇
2024 H210 篇
2025 H114 篇
2025 H211 篇
2026 H115 篇
2026 H24 篇
2018-02P2
Deep Contextualized…AI2
Deep Contextualized Word Representations
AI2
贡献：双向LSTM深度上下文词表示，动态嵌入替代静态向量
2018-06P2
GPT-1OpenAI
Improving Language Understanding by Generative Pre-Training
OpenAI
贡献：确立生成式预训练+判别式微调两阶段范式
2019-04P2
Sparse AttnOpenAI
Generating Long Sequences with Sparse Transformers
OpenAI
贡献：strided/fixed稀疏注意力，复杂度O(L√L)
2019-09P2
MegatronNVIDIA
Megatron-LM: Training Multi-Billion Parameter Language Models Using Model Parallelism
NVIDIA
贡献：模型并行+张量切分，可训练多亿级参数LLM
2020-05P3
RAGMeta AI
Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks
Meta AI
贡献：参数记忆+非参数检索的seq2seq生成，命名RAG范式
2020-02P2
REALMGoogle
REALM: Retrieval-Augmented Language Model Pre-Training
Google
贡献：将MIPS检索嵌入预训练目标端到端训练
2022-01P2
GrokkingOpenAI
Grokking: Generalization Beyond Overfitting on Small Algorithmic Datasets
OpenAI
贡献：发现长时训练后从记忆到泛化的延迟突变现象
2022-03P3
ChinchillaDeepMind
Training Compute-Optimal Large Language Models
DeepMind
贡献：证明Kaplan低估数据：算力最优应参数数据约等比扩
2022-03P3
InstructGPTOpenAI
Training language models to follow instructions with human feedback
OpenAI
贡献：SFT+奖励模型+PPO三段RLHF，1.3B胜175B
2022-01P3
CoTGoogle
Chain-of-Thought Prompting Elicits Reasoning in Large Language Models
Google
贡献：中间推理步骤示例大幅提升复杂推理
2022-12P3
CAIAnthropic
Constitutional AI: Harmlessness from AI Feedback
Anthropic
贡献：用原则驱动AI自我批评修订，RLAIF替代人工标注
2022-12P3
Self-InstrUW/Microsoft
Self-Instruct: Aligning Language Models with Self-Generated Instructions
UW/Microsoft
贡献：LLM自举生成指令数据，175种子扩至52K
2022-10P2
GPTQIST Austria+HF
GPTQ: Accurate Post-Training Quantization for Generative Pre-trained Transformers
IST Austria+HF
贡献：基于OBQ的逐层权重量化，4bit近无损
2022-11P2
SpecDecodeGoogle Research
Fast Inference from Transformers via Speculative Decoding
Google Research
贡献：小模型草稿+大模型并行验证的无损加速解码
2023-03P2
AlpacaStanford
Stanford Alpaca: An Instruction-following LLaMA Model
Stanford
贡献：52K Self-Instruct数据+LLaMA低成本复刻指令能力
2023-05P3
DPOStanford
Direct Preference Optimization: Your Language Model is Secretly a Reward Model
Stanford
贡献：奖励重参数化进目标，偏好优化变监督分类损失
2023-03P2
Self-RefineCMU+AI2
Self-Refine: Iterative Refinement with Self-Feedback
CMU+AI2
贡献：同一模型生成-自评-修改循环，无需外部奖励
2023-05P2×2
GQAGoogle Research
GQA: Training Generalized Multi-Query Transformer Models from Multi-Head Checkpoints
Google Research
贡献：分组共享KV，MQA效率与MHA质量折中
2023-06P2
PosInterpMeta
Extending Context Window of Large Language Models via Positional Interpolation
Meta
贡献：训练时线性内插位置索引，低成本扩窗至32k
2023-06P2
AWQMIT韩松组
AWQ: Activation-aware Weight Quantization for LLM Compression and Acceleration
MIT韩松组
贡献：按激活分布识别保护约1%关键权重通道
2023-01P2
BLIP-2Salesforce
BLIP-2: Bootstrapping Language-Image Pre-training with Frozen Image Encoders and Large Language Models
Salesforce
贡献：Q-Former轻量桥接视觉与冻结LLM
2023-02P2
ToolformerMeta AI
Toolformer: Language Models Can Teach Themselves to Use Tools
Meta AI
贡献：自监督生成API调用数据教会LLM用工具
2023-03P2
ReflexionNortheastern+MIT
Reflexion: Language Agents with Verbal Reinforcement Learning
Northeastern+MIT
贡献：失败后生成语言化反思存入记忆，下轮重试
2023-12P3
MambaCMU/Princeton
Mamba: Linear-Time Sequence Modeling with Selective State Spaces
CMU/Princeton
贡献：选择性SSM：数据依赖参数+硬件感知扫描
2023-07P2×2
FlashAttn-2Princeton/Together AI
FlashAttention-2: Faster Attention with Better Parallelism and Work Partitioning
Princeton/Together AI
贡献：沿序列维并行+warp级分工重排
2023-09P2×2
LongLoRA上海交大等
LongLoRA: Efficient Fine-tuning of Long-Context Large Language Models
上海交大等
贡献：Shifted Sparse Attention+放宽LoRA，低成本训100k窗口
2023-09P2
StreamingLLMMIT
Efficient Streaming Language Models with Attention Sinks
MIT
贡献：attention sink机制让LLM稳定流式4M+超长对话
2023-12P1×2
InternVL上海AI实验室等
InternVL: Scaling up Vision Foundation Models and Aligning for Generic Visual-Linguistic Tasks
上海AI实验室等
贡献：6B视觉编码器对齐LLM的V1系列开端
2023-07P3×2
Llama-2Meta
Llama 2: Open Foundation and Fine-Tuned Chat Models
Meta
贡献：开源7B-70B基础+RLHF聊天模型，开放商用
2024-01P2×2
DS-MoEDeepSeek-AI
DeepSeekMoE: Towards Ultimate Expert Specialization in Mixture-of-Experts Language Models
DeepSeek-AI
贡献：细粒度专家切分+共享专家隔离双机制
2024-03P2
JambaAI21 Labs
Jamba: A Hybrid Transformer-Mamba Language Model
AI21 Labs
贡献：Mamba-Attention-MoE层间混合生产级架构
2024-05P2
Mamba-2Princeton/CMU
Transformers are SSMs: Generalized Models and Efficient Algorithms Through Structured State Space Duality
Princeton/CMU
贡献：SSD框架统一SSM与线性注意力对偶视角
2024-01P2
DS-LLMDeepSeek-AI
DeepSeek LLM: Scaling Open-Source Language Models with Longtermism
DeepSeek-AI
贡献：67B稠密开源，重估缩放定律提出数据参数新配比
2024-02P1
KTOContextualAI/MIT
KTO: Model Alignment as Prospect Theoretic Optimization
ContextualAI/MIT
贡献：基于前景理论，仅需二元好坏信号免配对偏好
2024-06P2×2
DS-Coder-V2DeepSeek-AI
DeepSeek-Coder-V2: Breaking the Barrier of Closed-Source Models in Code Intelligence
DeepSeek-AI
贡献：236B MoE代码模型(21B激活)，338语言128K上下文
2024-11P2
Scaling Laws…Harvard/MIT/Stanford/Databricks
Scaling Laws for Precision
Harvard/MIT/Stanford/Databricks
贡献：精度感知scaling law：低精度训练折损有效参数量
2024-08P3
TTC ScalingOpenAI/UMD
Scaling LLM Test-Time Compute Optimally can be More Effective than Scaling Model Parameters
OpenAI/UMD
贡献：测试时计算可更高效，小模型+test-time超14x大模型
2024-07P2×2
FlashAttn-3Together AI/Princeton等
FlashAttention-3: Fast and Accurate Attention with Asynchrony and Low-precision
Together AI/Princeton等
贡献：Hopper异步流水(warp专用/tma)+FP8量化
2025-01P2×2
MiniMax-01MiniMax
MiniMax-01: Scaling Foundation Models with Lightning Attention
MiniMax
贡献：456B-A45.9B闪电注意力混合MoE，原生1M上下文
2025-02P2
NSADeepSeek-AI+北大/UW
Native Sparse Attention: Hardware-Aligned and Natively Trainable Sparse Attention
DeepSeek-AI+北大/UW
贡献：压缩+选择+滑窗三分支原生可训练稀疏注意力
2025-02P2
MoBAMoonshot AI+清华
MoBA: Mixture of Block Attention for Long-Context LLMs
Moonshot AI+清华
贡献：块级top-k门控路由，把MoE思想引入注意力
2025-01P2
s1Stanford+UW+AI2
s1: Simple test-time scaling
Stanford+UW+AI2
贡献：1000样本SFT+budget forcing，最小代价复现o1测试时扩展曲线
2025-01P2
Kimi k1.5Moonshot AI
Kimi k1.5: Scaling Reinforcement Learning with LLMs
Moonshot AI
贡献：长短链联合RL与long2short，非MoE达o1级
2025-01P3
DS-R1DeepSeek-AI
DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning
DeepSeek-AI
贡献：纯RL涌现长CoT反思验证，GRPO+规则可验证奖励
2025-03P1×2
QwQ-32B阿里Qwen团队
QwQ-32B: Embracing the Power of Reinforcement Learning
阿里Qwen团队
贡献：官方博客：32B两阶段RL(规则验证器+通用RM)逼近R1并集成工具调用
2025-03P1
LongCoT Survey哈工大/中南/HKU/复旦
Towards Reasoning Era: A Survey of Long Chain-of-Thought
哈工大/中南/HKU/复旦
贡献：Long CoT首个系统综述：分类法+overthinking/test-time scaling梳理
2025-06P2×3
MiniMax-M1MiniMax
MiniMax-M1: Scaling Test-Time Compute Efficiently with Lightning Attention
MiniMax
贡献：首个开源混合注意力推理模型，CISPO算法80K思考预算
2025-06P2×2
Gemini 2.5Google DeepMind
Gemini 2.5 Technical Report
Google DeepMind
贡献：思考预算+原生多模态+1M上下文全档家族
2025-07P3×2
Kimi K2Moonshot AI
Kimi K2: Open Agentic Intelligence
Moonshot AI
贡献：1.04T总参32B激活MoE，MuonClip优化器+agentic数据合成
2025-10P2
Kimi LinearMoonshot AI
Kimi Linear: An Expressive, Efficient Attention Architecture
Moonshot AI
贡献：KDA(Kimi Delta Attention)线性混合架构
2025-11P2
Kimi K2 ThinkingMoonshot AI
Kimi K2 Thinking（技术博客/模型发布）
Moonshot AI
贡献：思考-行动交错推理，连续两百余步工具调用，原生INT4
2025-09P2×2
Hunyuan 3.0腾讯混元
HunyuanImage 3.0: Unified Multimodal Understanding and Generation
腾讯混元
贡献：原生多模态自回归统一理解+生成，MoE图像生成+原生CoT
2025-08P3×2
GLM-4.5Z.ai/智谱
GLM-4.5: Agentic, Reasoning, and Coding (ARC) Foundation Models
Z.ai/智谱
贡献：专家迭代+slime异步解耦RL基建的agentic后训练
2025-09P1
WebExplorerMiniMax
WebExplorer: Explore Progressive Web Information Acquisition
MiniMax
贡献：长程信息搜索Agent数据管线与框架
2026-02P3×3
GLM-5Z.ai/智谱
GLM-5: from Vibe Coding to Agentic Engineering
Z.ai/智谱
贡献：744B-A40B MoE，引入DSA稀疏注意力+异步Agent RL
2026-02P2×3
Qwen3.5阿里通义
Qwen3.5: Towards Native Multimodal Agents
阿里通义
贡献：Gated DeltaNet线性注意力+稀疏MoE 397B-A17B原生多模态智能体旗舰
2026-04P1
mHCDeepSeek-AI
Manifold-Constrained Hyper-Connections (mHC)
DeepSeek-AI
贡献：流形约束的超连接升级残差结构，随V4报告发表
2026-04P3×3
DS-V4DeepSeek-AI
DeepSeek-V4 Technical Report (Pro/Flash)
DeepSeek-AI
贡献：CSA+HCA混合注意力+mHC超连接+Muon优化器+FP4量化感知训练，1M上下文
2026-04P3×2
VisPrimitivesDeepSeek-AI+北大+清华
Thinking with Visual Primitives
DeepSeek-AI+北大+清华
贡献：点框空间标记作为最小思维单位嵌入CoT，弥合Reference Gap；CSA视觉KV压缩7056倍
2026-02P2×2
ForgeMiniMax
Forge: Scalable Agent RL Framework and CISPO
MiniMax
贡献：白盒黑盒Agent统一RL环+CISPO复合奖励
2026-02P2×2
Kimi K2.5Moonshot AI
Kimi K2.5: Visual Agentic Intelligence
Moonshot AI
贡献：文视联合训练+Agent Swarm并行编排(PARL)
2026-05P2×2
MiniMax-M2MiniMax
The MiniMax-M2 Series Technical Report
MiniMax
贡献：230B总参/A10B激活极致轻旗舰系列+Forge大规模Agent RL
2026-08P2×3
GLM-5.3-Flash智谱Z.ai
GLM-5.3-Flash（ox-alpha/牛来，原生多模态低延迟版）
智谱Z.ai
贡献：320B/18B激活、45层原生多模态(text+image+video)，首个稀疏注意力+线性注意力混合开源前沿模型(linear局部递归+DSA/IndexShare全局索引+mHC流形超连接+IndexPool索引缓存压缩)，1M上下文，INT4/fp8，MIT
2026-07P3×2
Kimi K3Moonshot AI
Kimi K3: Open Frontier Intelligence
Moonshot AI
贡献：2.8T总参/104B激活首个开源3T级，KDA+AttnRes+MoonClip二阶优化器
2026-08P2×2
Qwen3.8-Max阿里通义
Qwen3.8-Max: A New Bar for Coding and Cowork
阿里通义
贡献：编程与协同办公新旗舰；宣布开源 Max 本尊(2.4T总参/A95B激活)