A — Year-by-Year Milestones (Master Merge)

A single merged chronology consolidating the "year-by-year milestones" timeline from every cheat sheet in this knowledge base that has one (27 of 36 files). Each source's timeline is preserved verbatim under its topic; paper links are kept intact.

Compiled August 2026. A synthesized unified field timeline (the cross-cutting big beats, 2012–2026) is up top; per-topic timelines follow, grouped into six thematic parts.

Contents

• The Unified Timeline (synthesized)
• Part I — LLMs: Architecture, Tokenization & Efficiency
  • Transformer / xFormer Catalogue
  • KV Cache
  • Mixture-of-Experts (MoE)
  • Quantization
  • Pruning
  • Distillation
  • Parameter-Efficient Fine-Tuning (PEFT)
  • Scaling Laws
• Part II — LLMs: Training, Reasoning, RL & Agents
  • Reasoning Technologies
  • Policy Optimization
  • Agentic Intelligence
  • Prompt · Context · Harness · Graph Engineering & Self-Improving AI
  • Test-Time & Training-Free Optimization
• Part III — Data & Evaluation
  • Data Collection & Curation
  • Metrics & Evaluations
• Part IV — Generative Models: Diffusion, Video, 3D & World
  • Diffusion Models
  • Video Generation
  • 3D & Multi-View Generation
  • World Models
• Part V — Neural Rendering & 3D Reconstruction
  • NeRF
  • Gaussian Splatting
  • Neural Rendering
  • Structure from Motion
  • Delighting & Relighting
  • Photorealistic Avatars
• Part VI — Core Vision, Robotics & Autonomy
  • Vision-Language-Action (VLA) Models
  • Autonomous Driving

The Unified Timeline (synthesized big beats, 2012–2026)

One merged field-level chronology distilled from all 27 per-topic timelines below — the moments that mattered across every subfield. Deep-dive dates and citations are in the per-topic sections.

• 2012 — AlexNet wins ImageNet: the deep-learning era begins.
• 2014 — GANs; VAEs; seq2seq + additive attention; Adam; VGG/GoogLeNet.
• 2015 — ResNet; BatchNorm; U-Net; DQN on Atari; diffusion's thermodynamic seed (Sohl-Dickstein).
• 2016 — AlphaGo; WaveNet; TRPO → the policy-gradient-stability line.
• 2017 — the Transformer; PPO; sparsely-gated MoE; Mask R-CNN; PointNet.
• 2018 — BERT & GPT; StyleGAN/BigGAN; PointPillars; self-supervised pretraining takes off.
• 2019 — GPT-2; nuScenes/Waymo Open; Test-Time Training; Megatron; scaling gets serious.
• 2020 — GPT-3 & few-shot ICL; DDPM; NeRF; ViT; RAG; ColBERT; SimCLR; Kaplan scaling laws; Lift-Splat-Shoot; TENT.
• 2021 — CLIP & DALL·E; diffusion beats GANs (classifier guidance); LoRA; Swin; MAE; Decision Transformer; Codex/GSM8K verifiers; Instant-NGP-era fast NeRF.
• 2022 — Chinchilla; InstructGPT/RLHF; Chain-of-Thought & Self-Consistency; Stable Diffusion/Imagen/DALL·E 2; classifier-free guidance; FlashAttention; ReAct; Constitutional AI; DreamFusion; Whisper; ChatGPT (Nov).
• 2023 — GPT-4; Llama/Llama-2; DPO; Tree-of-Thoughts/Reflexion; Mixtral-era MoE; 3D Gaussian Splatting; SDXL/ControlNet; SAM/DINOv2; QLoRA; vLLM/PagedAttention; Mamba; Voyager; Let's Verify.
• 2024 — o1 (test-time reasoning); GPT-4o/Claude 3/Llama-3; DeepSeek-V2/V3 (MLA); Sora; MCP; GraphRAG; SWE-agent; MM-DiT/SD3/FLUX; TTT layers; ADAS; Self-Rewarding LMs; DUSt3R/MASt3R; FlashAttention-3; compute-optimal test-time scaling.
• 2025 — DeepSeek-R1 & the open reasoning-model era (GRPO); s1 budget forcing; TTRL; GEPA; Titans/ATLAS; SEAL/Darwin Gödel Machine/AlphaEvolve; VGGT/π³ feed-forward geometry; Absolute Zero/R-Zero self-play from zero; Agent Skills; temporal-graph memory (Zep); Veo 3 / Sora 2 / Genie 3; SAM 3; FP4 training.
• 2026 — FlashAttention-4 (Blackwell); DeepSeek Sparse Attention (V3.2); Kimi Linear; foundation-model backbones standard for perception; world-model closed-loop sim production-grade; Waymo 6th-gen robotaxi scaling; self-improving agent stacks — the open question is whether self-improvement compounds past the base-model ceiling.

Part I — LLMs: Architecture, Tokenization & Efficiency

Transformer / xFormer Catalogue

Merged from "Year-by-year shortlist" — XFormer_Catalogue_SOTA_Updated.md.

• 2017: Transformer.
• 2018: BERT, GPT-1.
• 2019: GPT-2, RoBERTa, T5 (released 2020), ALBERT, XLNet, BART.
• 2020: GPT-3, ELECTRA, DETR, ViT (released 2021), iGPT.
• 2021: CLIP, Swin, DALL-E, ALIGN, GLIDE, Codex.
• 2022: PaLM, Flamingo, BLIP-2, Stable Diffusion (LDM), Whisper, Diffusion Transformer (DiT), DreamFusion, NeRF → Instant-NGP.
• 2023: GPT-4, LLaMA, LLaMA 2, Mistral, BLIP-2, MiniGPT-4, LLaVA, SDXL, Segment Anything, Control-Net, NeRF → 3DGS.
• 2024: Gemini 1.5, GPT-4o, Claude 3, LLaMA 3, Mixtral \(8\times 22\mathrm{B}\), DeepSeek-V2 (MLA + MoE), Sora, SD3, FLUX, \(\pi_0\), MM-DiT, Chameleon.
• 2025: GPT-5, Claude 4 / Opus 4.6, Gemini 2.5, LLaMA 4, DeepSeek-V3, DeepSeek-R1 (pure-RL reasoning), Qwen 3, FLUX Kontext, Veo 3, Wan 2.1, VGGT, GR00T, \(\pi_{0.5}\), Cosmos.
• 2026: native multimodal default, GRPO standard for reasoning, MM-DiT standard for diffusion, MLA / GQA / MoE composed everywhere.

KV Cache

Merged from "Year-by-Year KV Cache Milestones" — KV_Cache_SOTA_Updated.md.

• 2017: Original Transformer; KV cache implicit in autoregressive decoding.
• 2019–2020: MQA (Shazeer); Multi-Query as first cache reduction.
• 2022: ALiBi enables length extrapolation without re-rotating KV; Mistral introduces sliding window.
• 2023: GQA standardized via Llama 2; vLLM (PagedAttention) debuts; StreamingLLM (sinks); H2O eviction.
• 2024: KIVI (INT4 KV); SnapKV; FastV (vision-token drop); MLA in DeepSeek-V2; SGLang RadixAttention; Anthropic prompt caching API; OpenAI prompt cache; DistServe; Mooncake released; cross-layer KV sharing (CLA, YOCO, MiniCache); MInference dynamic sparse prefill; KVQuant sub-3-bit KV.
• 2025: DeepSeek-V3 (MLA + FP8 KV at 671B); Qwen3 GQA + long context; tree-decoding production (EAGLE-3); native sparse attention (DeepSeek NSA); MiniMax-01 hybrid attention; Quest dynamic block retrieval; head-specialized (DuoAttention) and offload-retrieval (ShadowKV) KV.
• 2026: MLA / fine-grained KV-sharing standard; FP8 / INT4 KV mainstream; disaggregated serving widespread; multimodal KV optimization (FastV-style) standard in VLMs; trained sparse attention in flagships (DSA, Kimi Linear); future-query-aware (Expected Attention) and cross-layer (XQuant) KV compression.

Mixture-of-Experts (MoE)

Merged from "Year-by-Year MoE Milestones" — MoE_SOTA_Updated.md.

• 1991: Original "adaptive mixtures of local experts" (Jacobs et al.).
• 2017: Sparsely-Gated MoE (Shazeer et al.) — modern MoE in DL is born.
• 2020: GShard (Google); top-2 + auxiliary loss + EP at scale.
• 2021: Switch Transformer (top-1, simple); V-MoE (vision); ST-MoE (stability).
• 2022: GLaM (1.2T params); LIMoE (multimodal).
• 2023: Soft MoE (vision); Mixtral 8×7B (open MoE breakthrough); Expert Choice routing.
• 2024: Mixtral 8×22B; DeepSeekMoE / V2 (fine-grained + shared experts + MLA); DBRX; Grok-1 open; Snowflake Arctic; MoE-LLaVA; CuMo; Aria; DeepSeek-V3 (auxiliary-loss-free + FP8 + MTP).
• 2025: DeepSeek-R1 (MoE + RL reasoning); Llama 4 (MoE); Qwen 3 MoE; MiniMax-01 (hybrid attention + MoE); ktransformers consumer-MoE serving; Mixture-of-Depths.
• 2026: MoE standard at frontier; auxiliary-loss-free training default; FP8 / FP4 MoE inference; multimodal MoE mainstream.

Quantization

Merged from "Year-by-Year Quantization Milestones" — Quantization_SOTA_Updated.md.

• 2018–2019: Mixed-precision training (Apex / FP16); BERT INT8 PTQ.
• 2020–2021: TensorRT INT8 PTQ + QAT mature; ZeroQuant; Q-BERT.
• 2022: BF16 mainstream for LLM training; ZeroQuant-V2; SmoothQuant; LLM.int8().
• 2023: GPTQ; AWQ; QLoRA / NF4; bitsandbytes 4-bit; KIVI; FP8 on H100 mainstream.
• 2024: SpinQuant; QuaRot; SVDQuant; HQQ; OmniQuant; GGUF K-quants + IQ-quants; AutoAWQ + AutoGPTQ standardized; FP8 KV cache mainstream; MXFP4 standard.
• 2025: DeepSeek-V3 FP8 native MoE training; BitNet b1.58 follow-ups; FP4 inference on Blackwell; ktransformers consumer MoE; KV cache quantization standard in serving frameworks.
• 2026: FP4 training maturing; microscaling (MXFP) standard hardware; per-expert MoE quantization standard; INT2 / 1.58-bit research; quantization compositions (Q × distill × sparsity) Pareto-frontier.

Pruning

Merged from "Year-by-Year Pruning Milestones" — Pruning_SOTA_Updated.md.

• 1989: Optimal Brain Damage (LeCun et al.).
• 1993: Optimal Brain Surgeon (Hassibi & Stork).
• 2015: Deep Compression (Han et al.) — prune + quantize + Huffman.
• 2018: Lottery Ticket Hypothesis (Frankle & Carbin); SNIP.
• 2019–2020: GraSP, SynFlow; movement pruning; structured pruning for BERT.
• 2020: NVIDIA 2:4 sparsity announced (Ampere).
• 2021–2022: DynamicViT; A-ViT; structured pruning for ViT.
• 2023 (early): ToMe (Bolya & Hoffman); ToMeSD.
• 2023 (mid): SparseGPT; Wanda; H2O; LLM-Pruner.
• 2024: Scissorhands; StreamingLLM (sinks + window); SnapKV; Pyramid KV; FastV (vision-token drop); ShortGPT (layer pruning); SliceGPT (Microsoft); LightGaussian; CompGS; Compact3D; Mini-Splatting; RDOGS; Mixture-of-Depths (Raposo et al.); Quest (query-aware KV).
• 2025: Native Sparse Attention (DeepSeek NSA) — end-to-end sparse training; LLM-Streamline; combined pruning + FP8 + 2:4 standard for LLM serving; FastV-style drop standard in production VLMs; Wanda++, 2SSP, DarwinLM, EvoPress (evolutionary / structured post-training pruning); DuoAttention, ThinK (head/channel KV compression); PyramidDrop (staged visual-token drop).
• 2026: pruning + quantization + distillation as the unified production default; adaptive per-input compute mainstream; native sparse pretraining viable; 3DGS web deployment with 20× compression normal.

Distillation

Merged from "Year-by-Year Milestones" — Distillation_SOTA_Updated.md.

• 2014: FitNets (Romero) — feature distillation.
• 2015: Hinton et al. — the original KD with soft targets and temperature.
• 2016–2018: Sequence-KD (Kim & Rush), Born-Again Networks (BAN), Mean Teacher, Attention Transfer.
• 2019: DistilBERT, RKD, MobileBERT.
• 2020: TinyBERT, MiniLM, DeiT (with distillation token), Reformer-distill.
• 2021: BYOL, SimSiam, DINO — self-distillation as SSL becomes a standard pre-training paradigm.
• 2022: Progressive distillation for diffusion (Salimans & Ho); DINOv2; Distil-Whisper later.
• 2023: Consistency Models, Latent Consistency Models, SDXL Turbo (LADD), DMD; Vicuna / WizardLM (open-instruction); Phi-1/1.5; GKD.
• 2024: DMD2, Hyper-SD, PCM, FLUX schnell; Phi-3 / 3.5; Gemma 2 (distill from Gemini); Distil-Whisper-v2; MobileSAM / FastSAM; ULD cross-tokenizer.
• 2025: R1-Distill series (DeepSeek) — reasoning distillation goes mainstream; Phi-4; Gemma 3; multimodal R1-distill (Vision-R1, VLM-R1 distill); cold-start + GRPO standard.
• 2025 (cont.): continuous-time consistency matures (sCM, SANA-Sprint); Inductive Moment Matching and MeanFlow push one-step generation; one-step video via adversarial post-training; on-policy LLM KD refined (TAID, DistiLLM-2); dataset distillation scales to ImageNet (SRe2L, RDED).
• 2026: Distillation + RL pipelines standard; cross-architecture (Transformer → Mamba); native multimodal distillation; constitutional distillation in alignment pipelines; on-policy distillation as a cheap alternative to RL; score-regularized consistency scales to large video (rCM).

Parameter-Efficient Fine-Tuning (PEFT)

Merged from "Year-by-Year PEFT Milestones" — PEFT_SOTA_Updated.md.

• 2019: Houlsby Adapter; Pfeiffer Adapter; AdapterHub.
• 2020: Aghajanyan et al. intrinsic dimension; BitFit; prompt tuning.
• 2021: LoRA (Hu et al.); prefix tuning; P-tuning; Compacter.
• 2022: AdapterFusion; IA3; CoOp / CoCoOp for CLIP; CLIP-Adapter / Tip-Adapter.
• 2023: QLoRA / NF4 (Dettmers); LyCORIS (LoCon / LoHa / LoKr); P-tuning v2; AdaLoRA.
• 2024 (early): DoRA; rsLoRA; LoRA+; S-LoRA / LoRAX (multi-LoRA serving); OFT / BOFT for diffusion.
• 2024 (mid–late): Unsloth (consumer-GPU PEFT); VeRA; LoRA-FA; LoRA-XS; ReLoRA; Civitai LoRA marketplace mainstream; multi-LoRA serving production-grade.
• 2025: PEFT + GRPO (R1-style RL via LoRA); cold-start SFT-LoRA standard; LoRA-MoE / X-LoRA active research; FLUX LoRAs ubiquitous; SD 3.5 LoRAs.
• 2026: PEFT mainstream for fine-tuning at all scales; multi-LoRA per-user customization standard in commercial APIs; DoRA / OFT mainstream alongside LoRA; PEFT theory reaching practical maturity (capacity bounds, conflict-free composition).

Scaling Laws

Merged from "Year-by-Year Scaling Milestones" — Scaling_Laws_SOTA_Updated.md.

• 2017: Hestness et al. universal scaling laws.
• 2018: Larger ImageNet networks; first hints of LLM scaling.
• 2019: GPT-2; OpenAI begins scaling-laws research.
• 2020: Kaplan et al. scaling laws; GPT-3 (175B); \(C = 6ND\) approximation.
• 2021: Henighan et al. multi-modal scaling.
• 2022: Chinchilla (Hoffmann et al.) corrects Kaplan; 20 tokens/param.
• 2023: µP / µTransfer mainstream; Llama 1; emergent abilities debate.
• Late 2023–2024: Llama 2 / 3 over-training; Schaeffer "Emergence is a Mirage".
• 2024 (mid–late): DeepSeek MoE laws; inference-aware scaling (Sardana et al.); FineWeb-Edu data-quality story; Snell test-time compute scaling.
• Late 2024–2025: DeepSeek-R1 pure-RL reasoning emergence; Sora 2 / Veo 3 video scaling; o1 / o3 inference compute paradigm; Llama 3.1 / 3.2 / 4 over-training.
• 2025: Stargate-class clusters; B200 deployment; FP8 / FP4 training; data quality debates intensify; multimodal scaling formalized.
• 2026: pretraining plateau debates; test-time compute scaling laws formalized; multi-axis scaling (pretrain + RL + inference) standard.

Part II — LLMs: Training, Reasoning, RL & Agents

Reasoning Technologies

Merged from "Year-by-Year Reasoning Milestones" — Reasoning_Technologies_SOTA_Updated.md.

• 2017–2020: Pretrained LMs; in-context learning emerges with GPT-3.
• 2022: Chain-of-Thought (Wei), Self-Consistency (Wang), Zero-shot CoT (Kojima), ReAct (Yao), Toolformer.
• 2023: Tree-of-Thought, PAL/PoT, Self-Refine, Reflexion, Plan-and-Solve, Step-Back, code interpreters in production (GPT-4).
• 2024: PRM800K, Math-Shepherd, OmegaPRM, AlphaGeometry, rStar, Mulberry, Marco-o1, Snell et al. inference-time scaling laws.
• Late 2024: OpenAI o1 (RL-trained reasoning), DeepSeek-R1 paradigm published.
• 2025: R1 / R1-Zero open-source release; QwQ, Skywork-o1, Qwen3-Reasoner; Vision-R1, MM-EUREKA, VLM-R1; o3 announced.
• 2026: GRPO + verifiable rewards is the standard reasoning recipe; reasoning capabilities mainstream in open + closed; multimodal R1-style native; reasoning agents in production (Claude Code, Cursor, Devin, OpenAI Operator).

Policy Optimization

Merged from "Year-by-Year Milestones" — Policy_Optimization_SOTA_Updated.md.

• 1992: REINFORCE (Williams).
• 1999: Policy Gradient Theorem (Sutton et al.).
• 2014: DPG (Silver et al.).
• 2015: TRPO (Schulman et al.); DQN (Mnih et al.).
• 2016: A3C, Dueling/Double DQN, GAE, DDPG.
• 2017: PPO (Schulman et al.); IMPALA / V-trace.
• 2018: SAC (Haarnoja); TD3 (Fujimoto); Rainbow DQN; AlphaZero.
• 2019: SAC v2 (auto \(\alpha\)); MuZero.
• 2020: BCQ, CQL; Decision Transformer (released 2021).
• 2021: Decision Transformer / Trajectory Transformer; IQL; PEARL.
• 2022: PPO-RLHF (InstructGPT recipe); Dreamer V3; TD-MPC.
• 2023: DPO; SimPO antecedents; Diffusion Policy; ACT (ALOHA); EfficientZero V2.
• 2024: GRPO (DeepSeek); IPO, KTO, ORPO; Diffusion-DPO; \(\pi_0\) flow-matching policy; Eureka, DrEureka.
• 2025: DeepSeek-R1 (pure RL reasoning); DAPO; Dr. GRPO; RLOO mainstream; VLM-R1, MM-EUREKA, Vision-R1; \(\pi_{0.5}\).
• 2026: GRPO-family standard; multimodal R1-style standard; RL fine-tuning of diffusion routine; world-model + RL converging in robotics.

Agentic Intelligence

Merged from "Year-by-Year Agentic Milestones" — Agentic_Intelligence_SOTA_Updated.md.

• 2022: ReAct, Toolformer, AutoGPT (early autonomous), LangChain v0.
• 2023: Voyager, Reflexion, MetaGPT/ChatDev, Self-Refine, GPT-4 with Code Interpreter, function calling becomes standard.
• 2024: Devin (Cognition), MCP launch (Anthropic), Claude Computer Use, OpenAI Swarm/Agents SDK, AutoGen 0.4, LangGraph maturation, Cursor / Aider / Cline mainstream.
• 2025: Frontier reasoning agents (R1-style + tools), Operator (OpenAI), Claude Code at scale, OpenAI Apps SDK, native multimodal agents (Gemini 2.5), Browser-Use mainstream, MLE-bench, OpenHands/Magentic-One, agentic-search RL (Search-o1, R1-Searcher), economic-value evals (GDPval).
• 2026: Agents in production across SWE, research, customer support, ops; MCP ecosystem; multimodal computer use mainstream; long-horizon agents (hours, not minutes) reliable.

Prompt · Context · Harness · Graph Engineering & Self-Improving AI

Merged from "Year-by-Year Milestones" — Prompt_Context_Harness_Graph_Engineering_SOTA_Updated.md.

• 2020: GPT-3 few-shot ICL; RAG; ColBERT.
• 2022: CoT, Zero-shot CoT, Self-Consistency, STaR, ReAct, induction heads, Constitutional AI; prompt-injection named.
• 2023: ToT/GoT, DSPy, Self-Refine/Reflexion, MemGPT, Lost in the Middle, Self-RAG, ReST, STOP, Promptbreeder, "can't self-correct yet"; self-authored skill/tool libraries (Voyager, CREATOR, Tool Makers).
• 2024: MCP; GraphRAG, HippoRAG, LightRAG; SWE-agent/OpenHands; Self-Rewarding LMs, SPIN, ReST-EM; ADAS, Gödel Agent; TextGrad; context-window scaling.
• 2025: DeepSeek-R1 & the reasoning-model era; context engineering formalized (Anthropic); ★ Context Rot/NoLiMa; ★ temporal-graph memory (Zep, Mem0); ★ GEPA; ★ self-play from zero (Absolute Zero, R-Zero), ★ TTRL; ★ Darwin Gödel Machine, ★ AlphaEvolve, ★ SEAL; self-evolving-agent surveys (1, 2); ★ Agent Skills (progressive-disclosure capability layer + plugin ecosystem).
• 2026: the four engineering layers converge into self-improving agent stacks; graph memory + agentic GraphRAG mainstream; harness-level RL and benchmark-gated self-modification at the frontier; the open question is whether self-improvement compounds past the base model's ceiling.

Test-Time & Training-Free Optimization

Merged from "Year-by-Year Milestones" — Test_Time_and_Training_Free_Optimization_SOTA_Updated.md.

• 2020: TTT (rotation SSL); TENT; BN-adapt; SHOT.
• 2021: Classifier guidance; Fisher merging; TTT++ (feature alignment).
• 2022: Self-Consistency; CFG; Model Soups; Contrastive Decoding; Task Arithmetic; MEMO; CoTTA; TPT; TTT-MAE; GSM8K verifiers; spec decoding.
• 2023: DoLa; CAD; RepE/ITI; TIES/DARE; Let's Verify; Math-Shepherd; ToT; SAR; universal guidance.
• 2024: ★ Snell compute-optimal; ★ Large Language Monkeys; Medusa/EAGLE/lookahead; PAG; ★ TTT layers; ★ ARC-TTT; TDA.
• 2025: ★ DeepSeek-R1; ★ s1; ★ rStar-Math; ★ TTRL; ★ diffusion test-time scaling; ★ Titans/ATLAS; ★ TPO; TTT provable ICL.
• 2026: test-time compute a first-class scaling axis alongside pretraining; training-free inference (steering, merging, guidance, spec-decoding) standard production tooling; TTT-as-memory architectures maturing for long context; the binding constraint everywhere remains the verifier / signal.

Part III — Data & Evaluation

Data Collection & Curation

Merged from "Year-by-Year Data Milestones" — Data_Collection_Curation_SOTA_Updated.md.

• 2018–19: Common Crawl + WikiText for BERT-era; Books3 controversially included in The Pile.
• 2020: The Pile (EleutherAI); C4 (T5); LAION-400M.
• 2021: LAION-5B; multilingual extensions (mC4); CLIP demonstrates scale + curation.
• 2022: Self-Instruct; InstructGPT demonstrates RLHF data flywheel; HH-RLHF released.
• 2023: Phi-1 synthetic-data winning; LLaMA recipe; RefinedWeb (Falcon); Vicuna pattern; Stanford Alpaca; OBELICS; Open-X-Embodiment announced.
• 2024: FineWeb / FineWeb-Edu (HuggingFace); DCLM-baseline (DataComp-LM); Dolma / OLMo data toolkit; Phi-3; UltraFeedback; LLaMA 3 15T tokens; PRM800K / Math-Shepherd PRM data; Tulu 3 + OpenHermes 2.5; NuminaMath; DROID (76k robot demos); Re-captioning standard for image / video gen (DALL-E 3, Sora).
• 2024 (cont.): Cosmopedia open synthetic textbooks; WRAP web-rephrasing; ProX model-based data refinement; Shumailov et al. model-collapse result (Nature).
• 2025: R1-Distill traces (800k); Phi-4; Llama 4 30T+; AgiBot World; Wan / Hunyuan video data recipes; verifiable-reward problem sets become a curation focus (OpenCodeReasoning, OpenMathReasoning, OpenThoughts, s1); FineWeb2 multilingual (1000+ languages); Ultra-FineWeb verification filtering; Essential-Web / Common Pile open corpora; Anthropic / OpenAI labeler ops scale to thousands; data-engine flywheel standard at frontier labs.
• 2026: synthetic data generation → filter → verify pipelines mature; multi-modal data scarcity increasingly addressed via sim + diffusion augmentation; per-jurisdiction compliance baked into pipeline; data-quality benchmarking standardized via DataComp-LM / DataComp-1B; reasoning data + verifiable rewards driving frontier RL.

Metrics & Evaluations

Merged from "Year-by-Year Eval Milestones" — Metrics_Evaluations_SOTA_Updated.md.

• 2014–2017: ImageNet / COCO mAP / BLEU as gold standards; FID 2017 for GAN.
• 2018: GLUE; LPIPS perceptual metric.
• 2019: SuperGLUE; BERT-era saturation.
• 2020: BIG-Bench started; HellaSwag / ARC.
• 2021: HELM (Stanford); MMLU.
• 2022: HumanEval / MBPP for code; lm-eval-harness mainstream.
• 2023: MT-Bench (LMSYS); GSM8K / MATH for math; Chatbot Arena launched; saturating MMLU.
• 2024: MMLU-Pro; GPQA; LiveCodeBench; SWE-bench-Verified; AlpacaEval 2; Arena-Hard; FineWeb-Edu eval correlation studies; VBench / VBench-2; Image Arena; HAL benchmark hub; PRM800K / Math-Shepherd PRM data.
• 2025: Humanity's Last Exam; FrontierMath (hidden); ARC-AGI-2; LiveBench monthly; SWE-bench Multimodal; MLE-bench; verifiable-reward benchmarks dominant for reasoning; per-capability multi-suite reporting.
• 2026: continually-updated benchmarks standard; private test sets prominent; arena-style for every modality; per-task per-segment reporting normalized; closed-loop downstream eval mandatory for production claims.

Part IV — Generative Models: Diffusion, Video, 3D & World

Diffusion Models

Merged from "Year-by-year Highlights" — Diffusion_Models_SOTA_Updated.md.

• 2015: Sohl-Dickstein, deep unsupervised learning via non-equilibrium thermodynamics (the seed).
• 2019: NCSN (Yang Song), score matching with Langevin dynamics.
• 2020: DDPM (Ho et al.); image quality competitive with GANs.
• 2021: GLIDE, CFG, Improved DDPM, score SDE, DALL-E 1, NCSN++.
• 2022: LDM / Stable Diffusion, DALL-E 2, Imagen, EDM, DiT, DreamBooth, ControlNet, classifier-free guidance ubiquitous.
• 2023: SDXL, ControlNet ecosystem, LoRA standard, MVDream / Zero123, AnimateDiff, SVD, Consistency Models, MAGVIT-v2.
• 2024: SD3 (MM-DiT + RF), FLUX.1, PixArt-\(\alpha/\Sigma\), Sora, Veo 2, Kling, MovieGen, \(\pi_0\) flow-matching policy, Trellis, DMD2, Hyper-SD.
• 2025: FLUX Kontext (editing), SD 3.5, Wan 2.1, Hunyuan Video, Sora 2, Veo 3 (with audio), Cosmos, native 3D diffusion mainstream, \(\pi_{0.5}\).
• 2026: flow matching standard, MM-DiT standard, distilled few-step ubiquitous, video / audio / 3D joint generation in production.

Video Generation

Merged from "Year-by-Year Video Generation Milestones" — Video_Generation_SOTA_Updated.md.

• 2022: Make-A-Video, Imagen Video, Phenaki — early T2V research.
• 2023: ModelScope T2V (open baseline); AnimateDiff (SD + temporal modules); SVD / SVD-XT (Stability I2V); ZeroScope; Show-1.
• Feb 2024: Sora (OpenAI) — transformative quality jump; "world simulator."
• May 2024: Veo 1 (Google).
• 2024 (mid): Kling 1 (Kuaishou); Pika 1.0; Runway Gen-3; Luma Dream Machine; CogVideoX 5B (Tsinghua, open); Open-Sora; AnimateAnyone (Alibaba); EMO; MimicMotion.
• 2024 (late): Mochi 1 (Genmo, open); Hunyuan Video (Tencent, 13B open frontier); LTX-Video (real-time); Allegro; Step-Video; MovieGen (Meta, joint video+audio); Cosmos World Foundation Models (NVIDIA); Genie 2 (DeepMind, playable 3D).
• 2025 (early): Veo 2 with camera control; Wan 2.1 (Alibaba, open frontier); MAGI-1 (chunk-wise AR); SkyReels-V2 / FramePack (long video); Runway Gen-4; Marey (Moonvalley, licensed); Hunyuan Video Avatar variants.
• May 2025: Veo 3 (Google) — native audio.
• Jun 2025: Seedance 1.0 (ByteDance) — multi-shot T2V/I2V frontier.
• Aug 2025: Genie 3 (DeepMind) — real-time general world model.
• Sep 2025: Sora 2 (OpenAI) — audio + longer + physics improvement; Wan-Animate / Wan-S2V.
• Oct 2025: Veo 3.1 (Google); Self-Forcing++ / LongLive (minute-scale AR); Krea Realtime 14B (open real-time).
• 2025 (late): Wan 2.2 (audio support); HunyuanVideo 1.5 (lightweight open); Cosmos Predict 2; LTX-Video 13B.
• 2026: native multimodal video + audio standard; world-model framing mainstream; long-form (> 1 min) viable; real-time on consumer hardware via distillation; open frontier (Wan, Hunyuan) competitive with closed.

3D & Multi-View Generation

Merged from "Year-by-Year Milestones" — 3D_MultiView_Generation_SOTA_Updated.md.

• 2020: NeRF; PolyGen (autoregressive mesh).
• 2021: PixelNeRF; Mip-NeRF; Plenoxels; DVGO.
• 2022: DreamFusion (SDS); Magic3D; Latent-NeRF; Instant-NGP; Score-Jacobian Chaining; ScanNet++.
• 2023: 3D Gaussian Splatting (Kerbl SIGGRAPH); ProlificDreamer (VSD); Zero-1-to-3; MVDream; ImageDream; Wonder3D; SyncDreamer; LRM (Adobe); InstantMesh; DreamGaussian; LucidDreamer; Fantasia3D; Objaverse + Objaverse-XL.
• 2024 (early): MeshLRM; GS-LRM; CRM; Long-LRM; PF-LRM; DUSt3R (Naver); MASt3R; SV3D; Splatter Image; LGM; LucidDreamer-Scene; MASt3R-SfM.
• 2024 (mid–late): Trellis (Microsoft); Hunyuan3D / Hunyuan3D-2 (Tencent); CLAY; Direct3D; 3DTopia-XL; SF3D; SPAR3D; MeshGPT; MeshXL; MeshAnything / V2; EdgeRunner; BPT; Spann3R; Splatt3R; NoPoSplat; Fast3R; Cube (Roblox); Rodin Gen-1.5; PixMo / Molmo; Genie 2.
• 2025 (early): VGGT (Meta) — feed-forward 3D dominant; π3 pose-free; CUT3R; LLaMA-Mesh; DepthCrafter; StreamingT3R; MoGe / MoGe-2; TripoSG; Hi3DGen; LHM; Bolt3D; texture-aware native 3D models.
• 2025 (mid–late): native 3D diffusion mainstream; Hunyuan3D 2.1 (PBR) + Hunyuan3D 2.5; Step1X-3D; StreamVGGT; SpatialTrackerV2; HunyuanWorld 1.0 / Voyager; Genie 3 (DeepMind); Tripo / Meshy / CSM commercial maturity; Codec Avatars 3.0; World Labs founded.
• 2026: feed-forward 3D + native 3D diffusion replace classical pipelines for asset gen; Marble (World Labs) commercial world model; AR/VR worldbuilding standard via 3DGS; web 3DGS deployment; PBR-aware native generation; physics-aware emerging.

World Models

Merged from "Year-by-Year Milestones" — World_Models_SOTA_Updated.md.

• 1990: Sutton's Dyna; Schmidhuber early world-model proposals.
• 2011: PILCO (Deisenroth & Rasmussen) — GP world model.
• 2018: PETS; Ha & Schmidhuber's "World Models" paper (VAE+RNN+controller).
• 2019–2020: Dreamer V1/V2; PlaNet; Plan2Explore.
• 2022: TD-MPC; Iso-Dream; SimMIM-style passive video pretraining (V-JEPA precursor).
• 2023: Dreamer V3; IRIS; DIAMOND; GAIA-1 (Wayve); MagicDrive; DriveDreamer.
• Feb 2024: Sora announced (OpenAI); positioned as "world simulator."
• 2024: Veo 1; Genie 1 (DeepMind, 2D playable); Cosmos announced (NVIDIA); GameNGen (DOOM); Oasis (Decart, playable Minecraft); MovieGen; Kling 1.6; DIAMOND-CS (CS:GO).
• Dec 2024: Genie 2 (3D playable, 1 min interactive).
• 2025: Sora 2, Veo 3 (with audio), Kling 2; Cosmos World Foundation Models released by NVIDIA (open + commercial); GAIA-2; 1X World Model; World Labs founded; Genie 3 rumored; Wayve LINGO.
• 2026: World models as standard substrate for robotics + AV; multi-minute coherent generation; playable worlds advancing toward production game-engine replacement; Cosmos / GR00T as platform-style adoption; physical commonsense + Cosmos-Reason as evaluation standard.

Part V — Neural Rendering & 3D Reconstruction

NeRF

Merged from "year-by-year milestones" — NeRF_SOTA_Updated.md.

Key

The NeRF arc (\(2020 \to 2026\)) introduced volume-rendering MLPs; 2021–2022 solved aliasing, speed, surfaces, dynamics, and 3D-aware GANs; 2023 delivered SOTA quality (Zip-NeRF, NeuralAngelo) just as 3DGS arrived; 2024–2026 repositioned NeRF as the accuracy-first 3D representation — relighting, dense mesh, generative pipelines, hybrids — while 3DGS owns real-time NVS.

Gaussian Splatting

Merged from "year-by-year milestones" — Gaussian_Splatting_SOTA_Updated.md.

Key

The arc: 2023 introduced 3DGS; 2024 built every major variant (geometry, dynamics, compression, generation, SLAM, avatars, AV); 2025 pushed toward feed-forward foundation models and on-device serving; 2026 fuses 3DGS with world-models and physics so a single representation supports rendering, simulation, generation, and learning.

Neural Rendering

Merged from "Year-by-Year Milestones" — Neural_Rendering_SOTA_Updated.md.

• 2020: NeRF (Mildenhall et al.) — the foundational paper.
• 2021: PixelNeRF (generalizable), Mip-NeRF (anti-aliasing), Plenoxels (no MLP), DVGO; Nerfies (dynamic).
• 2022: Instant-NGP (hash grids; seconds-to-train); Mip-NeRF 360 (unbounded); TensoRF; Block-NeRF (city-scale); DreamFusion (SDS).
• 2023: 3D Gaussian Splatting (Kerbl SIGGRAPH best paper); Zero123 (multi-view diffusion); LRM (feedforward 3D); ProlificDreamer (VSD); MVDream / Wonder3D / SyncDreamer.
• 2024: 3DGS variants (Mip-Splatting / 2D-GS / Scaffold-GS / 4D-GS / SuGaR / GS-IR); InstantMesh / MeshLRM; DUSt3R; MASt3R; LCM-3D; Trellis released; MoGe; Gaussian Avatars; GS-SLAM / MonoGS / SplaTAM; AnimateAnyone / EMO.
• 2025: VGGT (Meta) — feed-forward 3D dominant; Hunyuan3D-2 / CLAY / Direct3D — native 3D diffusion mainstream; PhysGaussian; Cosmos World Foundation Models (NVIDIA); MeshAnything V2 / EdgeRunner / BPT — autoregressive mesh; long-clip video gen aligned with 3D rendering.
• 2026: 3DGS standard for new projects; native 3D diffusion replaces SDS; feed-forward 3D replaces SfM as default; multimodal 3D scene reasoning (LangSplat, Feature 3DGS); web 3DGS deployment (gsplat / Brush); avatars in production (Apple, Meta).

Structure from Motion

Merged from "Year-by-Year SfM Milestones" — Structure_from_Motion_SOTA_Updated.md.

• 1981: Lucas-Kanade optical flow; Fischler & Bolles RANSAC.
• 1988: Harris corner detector.
• 1992: Tomasi-Kanade factorization (orthographic SfM).
• 1997: Hartley 8-point + normalization.
• 1999: SIFT (Lowe).
• 2000: Hartley & Zisserman "Multiple View Geometry" (the textbook).
• 2004: Nistér 5-point algorithm; SURF.
• 2007: PMVS / PMVS-2.
• 2008: Zhang's calibration; Bundler (precursor to COLMAP).
• 2011: ORB; FAST + BRIEF.
• 2015: COLMAP first release; PatchMatch MVS standardized; ORB-SLAM.
• 2016: NetVLAD; ORB-SLAM2.
• 2017: ORB-SLAM3 / VINS-Fusion mainstream; SuperPoint precursors.
• 2018: SuperPoint (DeTone); MVSNet; D2-Net.
• 2020: SuperGlue (Sarlin); R2D2; MAGSAC++; LoFTR (released 2021).
• 2022: HLoc mainstream; PixSfM.
• 2023: LightGlue (production-default learned matcher); AnyLoc; XFeat (released 2024).
• 2024 (early): VGGSfM (differentiable SfM); DUSt3R (Naver); MASt3R; Spann3R; Splatt3R; NoPoSplat; Fast3R; MoGe; Efficient LoFTR.
• 2024 (mid–late): MASt3R-SfM replaces COLMAP for many use cases; GLOMAP; MonST3R (dynamic); MV-DUSt3R+; PixSfM; HLoc 1.5.
• 2025: VGGT (Meta) — feed-forward 3D dominant; π3; MASt3R-SLAM; SLAM3R; MegaSaM; Light3R-SfM; MUSt3R; depthcrafter; MoGe-2; cross-modal SfM (sat + ground).
• 2026: feed-forward 3D + classical hybrid as the production default; mobile capture mainstream; SfM increasingly "solved" as separate problem; integrated with neural rendering.

Delighting & Relighting

Merged from "Year-by-Year Milestones" — Delighting_Relighting_SOTA_Updated.md.

• 1971: Land's Retinex.
• 1980: Woodham's photometric stereo.
• 1986: Kajiya's rendering equation.
• 1990s: Cook-Torrance, Marschner hair BRDF; HDR / Radiance from Greg Ward.
• 2000: Debevec light stage + reflectance fields; Ramamoorthi-Hanrahan SH irradiance.
• 2003–2007: Pre-filtered IBL, GGX standardized; Disney Principled BRDF (Brent Burley) 2012.
• 2012–2015: Quixel Megascans launches; UE4 split-sum IBL.
• 2014–2018: Intrinsic Images CNN era (Bell, Li & Snavely, Direct Intrinsics, IIW).
• 2018: Single-image SVBRDF from flash (Deschaintre).
• 2019: Sun et al. deep portrait relighting (SIGGRAPH).
• 2019–2022: NeRD / NeRF-OSR / PhySG / NeRO — NeRF + inverse rendering.
• 2021: Total Relighting; Mandalorian's Stagecraft popularizes virtual production.
• 2023–2024: IC-Light (Lvmin Zhang); SwitchLight (Beeble.ai); Relightful Harmonization (Google); \(\mathbf{RGB}{\to}\mathbf{X}\) (Adobe); Relightable 3D Gaussians; GS-IR; IntrinsicAnything.
• 2025: Diffusion-based relighting mainstream in commercial tools (Adobe Firefly, Photoshop Generative Fill); 4D-GS + relight; Vision Pro / Quest 3 high-fidelity AR insertion.
• 2026: Real-time consistency-distilled relighting on mobile; multi-view consistent diffusion relighting; integrated diffusion + physical hybrid pipelines for VFX.

Photorealistic Avatars

Merged from "Year-by-Year Milestones" — Photorealistic_Avatars_SOTA_Updated.md.

• 2015: SMPL parametric body model.
• 2017: FLAME parametric face model; Codec Avatar precursors.
• 2018: Codec Avatars 1.0 (Lombardi et al., Meta) — pioneering photoreal VR.
• 2019: Neural Volumes; HMR / pose regression mainstream.
• 2020: NeRF; HumanNeRF; SMPL-X.
• 2021: ICT Light Stage X; InstantAvatar.
• 2022: Codec Avatars 2.0; Total Relighting (Google); SadTalker.
• 2023: 3D Gaussian Splatting (Kerbl); GaussianAvatars; FlashAvatar.
• 2023: Apple Vision Pro Persona announced.
• 2024 (early): EMO (Alibaba); MimicMotion; AnimateAnyone; Champ; MagicAnimate; Live Portrait; Hallo / Hallo-2.
• 2024 (mid): Apple Vision Pro launch with Persona; Codec Avatars 3.0; Audio2Photoreal (Meta); Audio2Face commercial; PuLID; CogVideoX I2V for avatars.
• 2024 (late): GauHuman; HumanGaussian; HUGS; LightGaussian for avatars; Gaussian Head Avatar + Relightable Gaussian Codec Avatars; audio-driven wave: VASA-1, Loopy, Sonic, FLOAT; SqueezeMe on-device.
• 2025 (early): Sora 2 / Veo 3 audio integration enables avatar-style video; Codec Avatars Full Body; Apple iOS 18 enhanced Persona; OmniHuman-1 (ByteDance) full-body audio-driven; URAvatar universal relightable head from phone scan; Avat3r feed-forward heads.
• 2025 (mid–late): commercial avatar marketplace (HeyGen / Synthesia / D-ID); generative virtual influencers mainstream; consent / watermarking frameworks (SynthID-style).
• 2026: production-grade single-image full-body avatars; real-time on-device standard; Vision Pro 2 / Quest 4 telepresence-grade; holographic communication early prototypes.

Part VI — Core Vision, Robotics & Autonomy

Vision-Language-Action (VLA) Models

Merged from "Year-by-Year VLA Milestones" — VLA_Models_SOTA_Updated.md.

• 2022: RT-1 (Google, 35M); ACT (Stanford); Diffusion Policy (Columbia).
• 2023: RT-2 (Google); PaLM-E (Google); Mobile ALOHA (Stanford); RDT-1B; AnyGrasp.
• 2024 (early): RT-X / Open-X-Embodiment; OpenVLA (Stanford); Octo (Berkeley/Stanford); π0 (Physical Intelligence) released.
• 2024 (late): GR00T announced (NVIDIA); Helix (Figure) demoed; AgiBot World; DexCap; SimplerEnv.
• 2025 (early): GR00T N1 released; π0.5 (long-horizon humanoid); Cosmos World Foundation Models (NVIDIA); GR00T integration with Isaac Lab.
• 2025 (mid): Helix multi-robot coordination demo; 1X World Model + VLA integration; SpatialVLA / RoboPoint mainstream.
• 2025–2026: Production deployment of π0 / π0.5 in customer sites; Figure BMW partnership; Tesla Optimus VLA roll-outs.
• 2026: Humanoid VLA standard; cross-embodiment foundation models; world-model + VLA hybrids; RL fine-tuning routine; Jetson Thor as the standard edge target.

Autonomous Driving

Merged from "Year-by-Year AV Milestones" — Autonomous_Driving_SOTA_Updated.md.

• 2009: Google self-driving project starts (now Waymo).
• 2014: Mobileye SDK in Tesla; AP1 launches.
• 2015: NVIDIA Drive PX1; Tesla Autopilot v7 with vision-driven driving.
• 2017: Waymo Phoenix early-rider program; deep learning + LiDAR mainstream.
• 2018: PointPillars; Apollo open source; Cruise / Zoox milestones.
• 2019: nuScenes dataset; CenterPoint; PETR-precursor work; Tesla HW3 chip.
• 2020: LSS (Lift-Splat-Shoot); Waymo Open Dataset; Mobileye REM at scale.
• 2021: BEVFormer; CenterPoint mainstream; Tesla AI Day reveals occupancy networks.
• 2022: BEVFormer / BEVDet / PETR; PointPillars production; HD-map decline begins; UniAD prototype.
• 2023: UniAD (CVPR best paper); VAD; Wayve GAIA-1; Tesla FSD v11; nuPlan benchmark; SparseBEV / SparseDrive.
• 2024: Tesla FSD v12 end-to-end (major shift); Wayve LINGO-2; GAIA-2; Cosmos World Foundation Models; FB-OCC SOTA; MapTRv2; Hydra-MDP; OmniRe; StreetGaussians; DriveVLM; Senna; EMMA; DiffusionDrive.
• 2025: end-to-end mainstream; VLM/VLA reasoning (AlphaDrive, AutoVLA); diffusion/flow planners (GoalFlow); Cosmos as AV platform; DINOv3 / SAM 3 backbones; Waymo robotaxi 250k+ rides/week; Tesla Robotaxi (Austin) launch; Mobileye SuperVision in production; Chinese OEMs catch up rapidly; NVIDIA Thor dev kit GA.
• 2026: foundation-model backbones standard; world-model closed-loop sim (HUGSIM, NAVSIM v2) production-grade; Waymo 6th-gen Driver; long-tail edge cases the central frontier; L4 robotaxi expansion across cities; safety verification still the hardest open problem.

Files without a year-by-year timeline

These sheets have no standalone year-by-year section: Attention, CV Principal Deep Dive, CV Principal Math, Diffusion Derivations, Foundation Models, Reward Functions, RL Training Strategies & Recipes, Tokenization & Context (v1 & v2).