← Researchへ戻る

Keywords

LLMキーワード一覧

キーワードをクリックすると、LLM技術知識体系の該当ページを開きます。

1-6｜基礎・学習・推論

Transformerp.5 encoder-onlyp.5 encoder-decoderp.5 decoder-onlyp.5 causal maskp.5 cross-attentionp.5 QKVp.5 scaled dot-product attentionp.5 Multi-Head Attentionp.5 MHAp.5 MQAp.5 GQAp.5 KV cachep.5 FlashAttentionp.5 FlashAttention-3p.5 FlashAttention-4p.5 FFNp.5 SwiGLUp.5 pre-normp.5 post-normp.5 RoPEp.5 ALiBip.5 QK-Normp.5 logit soft-cappingp.5 誤差逆伝播p.12 forward propagationp.12 backpropagationp.12 勾配降下法p.12 SGDp.12 Adamp.12 AdamWp.12 学習率スケジュールp.12 warmupp.12 cosine decayp.12 weight decayp.12 cross entropyp.12 label smoothingp.12 gradient clippingp.12 activation checkpointingp.12 BF16p.12 FP16p.12 FP32p.12 loss scalingp.12 tokenizerp.20 BPEp.20 WordPiecep.20 Unigramp.20 SentencePiecep.20 byte fallbackp.20 vocabularyp.20 special tokensp.20 BOS / EOSp.20 chat templatep.20 detokenizationp.20 multilingual tokenizerp.20 scaling lawsp.25 Chinchilla則p.25 compute-optimalp.25 pretraining datap.25 data mixturep.25 データ重複除去p.25 data contaminationp.25 Common Crawlp.25 FineWebp.25 RefinedWebp.25 Dolmap.25 Nemotron-CCp.25 synthetic datap.25 SFTp.32 instruction tuningp.32 preference datap.32 reward modelp.32 RLHFp.32 PPOp.32 DPOp.32 IPOp.32 KTOp.32 ORPOp.32 RLAIFp.32 RLVRp.32 GRPOp.32 DAPOp.32 rejection samplingp.32 knowledge distillationp.32 judge modelp.32 推論モデルp.39 Chain-of-Thoughtp.39 hidden CoTp.39 test-time computep.39 推論時計算p.39 inference-time scalingp.39 thinking budgetp.39 reasoning effortp.39 best-of-Np.39 self-consistencyp.39 tree searchp.39 verifierp.39 process rewardp.39 outcome rewardp.39 budget forcingp.39 DeepSeek-R1p.39

7-12｜アーキテクチャ・評価・安全性

MoEp.43 Mixture of Expertsp.43 dense modelp.43 routed expertp.43 shared expertp.43 expert parallelismp.43 expert routingp.43 top-k routingp.43 top-1 routingp.43 load balancingp.43 router z-lossp.43 capacity factorp.43 all-to-allp.43 DeepSeekMoEp.43 fine-grained expertp.43 node-limited routingp.43 長文脈p.50 long contextp.50 context lengthp.50 lost in the middlep.50 RoPE scalingp.50 YaRNp.50 NTK-aware scalingp.50 positional interpolationp.50 sparse attentionp.50 Native Sparse Attentionp.50 NSAp.50 DSAp.50 MLAp.50 Ring Attentionp.50 MoBAp.50 LongRoPEp.50 KV compressionp.50 SSMp.56 State Space Modelp.56 Mambap.56 Mamba-2p.56 RWKVp.56 RetNetp.56 Hyenap.56 Linear Attentionp.56 Gated DeltaNetp.56 KDAp.56 Diffusion LMp.56 hybrid architecturep.56 マルチモーダルp.60 VLMp.60 vision encoderp.60 image tokenp.60 patch embeddingp.60 CLIPp.60 cross-modal projectorp.60 Q-Formerp.60 OCRp.60 ASRp.60 TTSp.60 音声理解p.60 動画理解p.60 frame samplingp.60 temporal modelingp.60 ベンチマークp.65 MMLUp.65 MMLU-Prop.65 GPQAp.65 GSM8Kp.65 MATHp.65 AIMEp.65 HumanEvalp.65 MBPPp.65 LiveCodeBenchp.65 SWE-benchp.65 Terminal-Benchp.65 BrowseCompp.65 LongBenchp.65 RULERp.65 IFEvalp.65 MT-Benchp.65 Arenap.65 pass@kp.65 アライメントp.71 alignmentp.71 RLHF alignmentp.71 red teamingp.71 jailbreakp.71 prompt injectionp.71 data poisoningp.71 model extractionp.71 privacyp.71 PIIp.71 unlearningp.71 Constitutional AIp.71 guardrailsp.71 mechanistic interpretabilityp.71 activation patchingp.71 sparse autoencoderp.71

13-18｜効率化・分散学習・推論システム

量子化p.76 quantizationp.76 PTQp.76 QATp.76 GPTQp.76 AWQp.76 SmoothQuantp.76 OmniQuantp.76 bitsandbytesp.76 INT8p.76 INT4p.76 FP8p.76 FP4p.76 NF4p.76 GGUFp.76 weight-only quantizationp.76 activation quantizationp.76 KV cache quantizationp.76 mixed precisionp.76 calibrationp.76 outlierp.76 枝刈りp.82 pruningp.82 structured pruningp.82 unstructured pruningp.82 magnitude pruningp.82 N:M sparsityp.82 block sparsityp.82 知識蒸留p.82 teacher-studentp.82 logit distillationp.82 feature distillationp.82 PEFTp.88 LoRAp.88 QLoRAp.88 AdaLoRAp.88 DoRAp.88 LoHap.88 LoKrp.88 IA3p.88 prefix tuningp.88 prompt tuningp.88 adapterp.88 full fine-tuningp.88 モデルマージp.88 TIES mergingp.88 DAREp.88 SLERPp.88 分散学習p.95 data parallelismp.95 tensor parallelismp.95 pipeline parallelismp.95 sequence parallelismp.95 3D parallelismp.95 ZeROp.95 FSDPp.95 DeepSpeedp.95 Megatron-LMp.95 NCCLp.95 all-reducep.95 all-gatherp.95 fused kernelp.95 CUDA Graphsp.95 Tritonp.95 vLLMp.101 Text Generation Inferencep.101 TGIp.101 TensorRT-LLMp.101 SGLangp.101 llama.cppp.101 Ollamap.101 ONNX Runtimep.101 PagedAttentionp.101 continuous batchingp.101 prefix cachingp.101 speculative decodingp.101 draft modelp.101 chunked prefillp.101 disaggregated servingp.101 TTFTp.101 tokens per secondp.101 throughputp.101 latencyp.101 GPUp.110 TPUp.110 H100p.110 H200p.110 B200p.110 Blackwellp.110 Hopperp.110 Tensor Corep.110 HBMp.110 NVLinkp.110 NVSwitchp.110 CPU offloadp.110 NPUp.110 NVFP4p.110 MXFP4p.110 E4M3p.110 E5M2p.110

19-23｜モデル・企業・OSSエコシステム

frontier modelp.120 GPTp.120 Claudep.120 Geminip.120 DeepSeek-V4-Prop.214 DeepSeek-V4-Flashp.120 DeepSeek-R1p.120 Qwenp.126 Llamap.126 Mistralp.126 Grokp.126 Kimip.126 GLMp.126 MiniMaxp.126 Command Rp.126 open weightsp.126 open source modelp.126 dense modelp.126 MoE modelp.126 Nemotron 3p.215 Nemotron 3 Nanop.215 Nemotron 3 Superp.215 Nemotron 3 Ultrap.215 Nemotron-Hp.126 日本語LLMp.131 LLM-jpp.131 Swallowp.131 PLaMop.131 Rakuten AIp.131 cotomip.131 Sarashinap.131 モデルルーティングp.136 hyperscalerp.136 AI labp.136 NVIDIAp.136 AMDp.136 Googlep.136 Microsoftp.136 AWSp.136 Metap.136 OpenAIp.136 Anthropicp.136 Hugging Facep.143 Model Hubp.143 safetensorsp.143 Transformersp.143 Datasetsp.143 Tokenizersp.143 Acceleratep.143 PEFT libraryp.143 TRLp.143 vLLM ecosystemp.143 llama.cpp ecosystemp.143 Ollama ecosystemp.143 OpenRouterp.143 LiteLLMp.143 Open Model Licensep.143 model cardp.143 data cardp.143

24-27｜RAG・エージェント・コーディング

Embeddingp.151 埋め込みp.151 dense retrievalp.151 sparse retrievalp.151 BM25p.151 ベクトルデータベースp.151 ANNp.151 HNSWp.151 IVFp.151 PQp.151 FAISSp.151 Milvusp.151 Qdrantp.151 Weaviatep.151 pgvectorp.151 RAGp.151 chunkingp.151 chunk overlapp.151 metadata filterp.151 rerankerp.151 cross-encoderp.151 query rewritingp.151 HyDEp.151 multi-query retrievalp.151 RAG Fusionp.151 contextual compressionp.151 GraphRAGp.151 agentic RAGp.151 groundednessp.151 citationp.151 AIエージェントp.159 agentp.159 tool callingp.159 function callingp.159 ReActp.159 planningp.159 reflectionp.159 workflowp.159 state machinep.159 MCPp.159 Model Context Protocolp.159 Skillsp.159 agent memoryp.159 short-term memoryp.159 long-term memoryp.159 episodic memoryp.159 semantic memoryp.159 procedural memoryp.159 sandboxp.159 permissionsp.159 JSON Schemap.159 multi-agentp.159 subagentp.159 orchestrationp.159 AgentCorep.168 LangChainp.168 LlamaIndexp.168 Haystackp.168 DSPyp.168 Semantic Kernelp.168 AutoGenp.168 CrewAIp.168 PydanticAIp.168 managed agent platformp.168 Bedrock Agentsp.168 Vertex AI Agent Builderp.168 コーディングエージェントp.175 code agentp.175 repo-level codingp.175 terminal agentp.175 shell toolp.175 patch generationp.175 test loopp.175 browser toolp.175 human-in-the-loopp.175 approval flowp.175 sandbox executionp.175 Git integrationp.175 pull requestp.175 code searchp.175 diffp.175 Codexp.175 Claude Codep.175 OpenCodep.175 DeepSeek Harnessp.214

28-31｜プロンプト・LLMOps・実務設計

prompt engineeringp.181 プロンプト工学p.181 system promptp.181 instruction hierarchyp.181 zero-shotp.181 few-shotp.181 in-context learningp.181 Chain-of-Thought promptingp.181 structured outputp.181 JSON modep.181 function calling schemap.181 context engineeringp.181 context packingp.181 context pruningp.181 conversation summarizationp.181 context compactionp.181 retrieval planningp.181 prefix cachingp.181 attention sinkp.181 delimiterp.181 prompt injection defensep.181 LLMOpsp.186 model registryp.186 prompt registryp.186 tracingp.186 observabilityp.186 evaluation pipelinep.186 monitoringp.186 model driftp.186 quality monitoringp.186 latency monitoringp.186 cost managementp.186 token usagep.186 cache hit ratep.186 rate limitp.186 retryp.186 fallbackp.186 model routingp.186 canary releasep.186 A/B testp.186 SLOp.186 SLAp.186 MIT licensep.196 Apache-2.0p.196 Llama licensep.196 Open Model Licensep.196 OpenMDWp.196 RAIL licensep.196 CC-BYp.196 CC-BY-NCp.196 model cardp.196 copyrightp.196 privacyp.196 AI Actp.196 NIST AI RMFp.196 Japan AI guidelinesp.196 export controlsp.196 requirements definitionp.202 workload designp.202 quality targetp.202 latency targetp.202 budget planningp.202 model selectionp.202 RAG vs fine-tuningp.202 prototypep.202 productionizationp.202 failure modesp.202 human evaluationp.202 rollbackp.202 security reviewp.202 incident responsep.202