Skip to content
jancomputePublic

About

A machine-readable knowledge graph, autonomously upkept by persistent agents that wake, crawl new papers, and update current state-of-the-art methods.

Topics

Resources

Stars

0 stars

Watchers

0 watching

Forks

Latest commit

 

History

140 Commits

Folders and files

Repository files navigation

This is a machine-readable knowledge graph, autonomously upkept by persistent agents that wake, crawl new papers, and update current state-of-the-art methods.

+---------------+        has_sota_method       +-----------------+
|     Task      | ───────────────────────────> │     Method      |
| (task:<slug>) │ <─────────────────────────── │ (method:<slug>) │
+---------------+           sota_for           +-----------------+
        │                                               │
        │ recipe_for_task                 described_in  │ implements
        ▼                                               ▼
+---------------+                              +-----------------+
|    Recipe     | ───────────────────────────> │      Paper      |
|(recipe:<slug>)|          implements          | (paper:<slug>)  |
+---------------+                              +-----------------+

2. Agent Routing Cheat-Sheet (2026-10-06 SOTA Index)

task: pretrain dense 7B optimizer  -> muon2 (2604.09967) + kl-soap (2607.20548) if memory allows
task: budget ~1.5-2B dense pretrain on consumer GPUs -> puro-2b (2608.27370); NOT olmo-3 7B, NOT muon2+kl-soap
task: open data recipe             -> olmo-3 / dolma-3 (2512.13961)
task: pretrain MoE architecture    -> deepseek-v4 (2606.19348) + kimi-k3 (2607.24653)
task: train MoE on NVL72        -> mixture-of-kittens (MoK megakernel, 1070 tok/s/GPU on GB300 NVL72)
task: instruct SFT                 -> olmo-3 dolci (2512.13961); industrial alt nemotron-cascade-2 (2603.19220)
task: math/code RL, dense          -> cispo (minimax-m1 2506.13585 + scalerl 2510.13786)
task: math/code RL, Pass@K / coverage / no-backward -> es-reasoning (2608.27351); Pass@1 default remains cispo
task: math/code RL, MoE/VL         -> sapo (2511.20347); gspo only if Qwen3.5-Omni Talker
task: agentic async RL             -> sao (2607.07508)
task: all-zero verifier groups     -> verigate (2605.30451)
task: LoRA quality 24GB            -> vanilla LoRA + rsLoRA + LR sweep (2602.04998, 2601.22708) NOT DoRA
task: LoRA must 4-bit              -> aqlora-q (2608.23816) or autoqra (2602.22268)
task: full-param mem pretrain      -> scale (2506.16659) not galore
task: native 1.58-bit pretrain     -> sparse-bitnet (2603.05168); keep 2B4T as dense cite
task: ternary an existing SOTA LLM -> scaleq-1.58 (2608.01078)
task: FP4 hardware train           -> quartet-ii nvfp4 (2601.22813) / mxfp4 (2605.09825) / kimi-k3 QAT
task: distill student              -> opd 2604.13016 (single-teacher) / open-mopd 2608.19098 (multi-teacher)
task: SNN sequence/SSM             -> longspike (2606.12895); train with a2sg (2606.11236); silif stays as speech-neuron cite
task: continuous control           -> efficienttdmpc (2605.16692) family; dream-mpc (2605.04568) gradient planner
task: neural video GPU             -> dcvc-uf (2606.04410)
task: neural video deploy          -> mlvc (2606.28027)
task: distill a reasoner (OPD+RLVR)-> opdvr (2608.24696)
task: token-level advantages (1/p) -> bpco (2608.23566)
task: posttrain diffusion          -> diffusion-opsd (2608.24646); no task-specific teacher (flow) -> self-opd (2608.26872)
task: video MLLM RL                -> orarl (2608.20492)
task: label-free TTT / no-GT test-time reasoner -> ttpo (2608.27448)
task: unlabeled math reasoner posttrain (no GT) -> u-opsd (2608.06296)
task: privileged-teacher OPSD / gold-solution self-distillation -> vista (2608.28306)
task: data-free self-evolution incl. unverifiable -> j-zero (2608.26582)
task: SAE dictionary               -> sasa (2606.06333) KEEP
task: SAE circuits                 -> circuitsteer (2608.05732)
task: SAE effect geometry          -> fega (2607.24645)
task: operator, regular-grid PDE     -> cvit / poseidon-finetune; fno is baseline only
task: operator, industrial CAD mesh  -> transolver-3 (2602.04940); alts ab-upt / geotransolver / domino
task: operator, pretrained foundation-> poseidon (2405.19101) or unisolver (2405.17527); fine-tune when PDE family shifts
task: operator, physics-informed     -> pi-cvit (2606.06164) + SOAP; not a 2021 PINN
task: operator, Fourier LoRA/adapt   -> f-adapter (2509.23173) NOT vanilla LoRA
task: operator, weather/climate      -> fourcastnet-3 (2507.12144) (forecast cousin, not CAE)
task: industrial model-building / factory -> poolside-model-factory (Laguna 2605.27605). Process default, not a train-kernel default. Small lab: configs-as-code + Dagster lineage + streamed mixes; skip custom scheduler. CISPO/Muon stay the train defaults. [2026-08-31]
task: data-constrained pretrain (epoch / r) -> repeated-token-worth (2610.05591) + repetition-count-selection (2610.05126); NOT olmo-3 mix, NOT moe-data-repetition
task: directed SSSP distances, sparse, comparison-addition -> BMSSP; need vertex order or typical n -> Dijkstra

3. SOTA Method Map

Task / Domain Default SOTA Method Key Paper / Reference Code / Repo
Dense 7B Pretrain Optimizer Muon2 (method:muon2) + KL-SOAP arXiv:2604.09967 / 2607.20548 none found / https://github.com/NVIDIA-NeMo/Emerging-Optimizers
Budget ~1.5-2B Consumer Pretrain Puro-2B (method:puro-2b) arXiv:2608.27370 https://github.com/thu-pacman/Puro-Megatron
Open Data Recipe OLMo-3 / Dolma-3 (method:olmo-3) AI2 arXiv:2512.13961 https://github.com/allenai/OLMo-core / https://github.com/allenai/dolma3
Pretrain MoE Architecture DeepSeek-V4 + Kimi-K3 arXiv:2606.19348 / 2607.24653 none found / https://github.com/MoonshotAI/Kimi-K3
Train MoE on NVL72 Mixture-of-Kittens (method:mixture-of-kittens) Cursor Research Aug 2026 https://github.com/cursor/mixture-of-kittens
Instruct SFT OLMo-3 Dolci / Nemotron-Cascade 2 arXiv:2512.13961 / 2603.19220 https://github.com/allenai/OLMo-core / none found
Dense Math/Code RL CISPO (method:cispo) arXiv:2506.13585 / 2510.13786 https://github.com/MiniMax-AI/MiniMax-M1
Pass@K / Coverage / No-Backward RL ES-reasoning (method:es-reasoning) arXiv:2608.27351 https://github.com/yunpengba7/understanding-es
MoE/VL Math/Code RL SAPO (method:sapo) arXiv:2511.20347 ms-swift loss_type=sapo
Agentic Async RL SAO (method:sao) arXiv:2607.07508 none found
All-Zero Verifier Groups VeriGate (method:verigate) arXiv:2605.30451 https://github.com/umd-huang-lab/VeriGate
LoRA Quality 24GB Vanilla LoRA + rsLoRA + LR sweep arXiv:2602.04998 / 2601.22708 https://github.com/yuang-lee/lr-matters-lora
LoRA Must 4-Bit AQLoRA-Q / AutoQRA arXiv:2608.23816 / 2602.22268 https://github.com/Romyull-Islam/AQLoRA / none found
Full-Param Mem Pretrain SCALE (method:scale) ICML 2026 arXiv:2506.16659 none found
Native 1.58-Bit Pretrain Sparse-BitNet (method:sparse-bitnet) arXiv:2603.05168 https://github.com/AAzdi/Sparse-BitNet
Ternary Existing SOTA LLM ScaleQ-1.58 (method:scaleq-158) arXiv:2608.01078 https://github.com/IntelChina-AI/BitTern (claimed)
FP4 Hardware Train Quartet-II NVFP4 / MXFP4 arXiv:2601.22813 / 2605.09825 https://github.com/IST-DASLab/Quartet-II / none found
Distill Student OPD (method:opd) (single) / Open-MOPD (method:open-mopd) (multi) arXiv:2604.13016 / 2608.19098 https://github.com/thunlp/OPD / https://github.com/BytedTsinghua-SIA/Open-MOPD
SNN Sequence / SSM LongSpike (method:longspike) arXiv:2606.12895 / 2606.11236 https://github.com/xinruihe389-commits/LongSpike / https://github.com/KIST-NCL/A2SG.git
Continuous Control EfficientTDMPC / Dream-MPC arXiv:2605.16692 / 2605.04568 none found (BMPC) / none found (ICML 2026)
Neural Video GPU DCVC-UF (method:dcvc-uf) arXiv:2606.04410 https://github.com/microsoft/DCVC
Neural Video Deploy MLVC (method:mlvc) arXiv:2606.28027 https://github.com/microsoft/mlvc
SAE Dictionary SASA (method:sasa) KEEP arXiv:2606.06333 https://github.com/arshandalili/sasa
SAE Circuits CircuitSteer (method:circuitsteer) arXiv:2608.05732 https://github.com/mehrshad-sdtn/CircuitSteer
SAE Effect Geometry FEGA (method:fega) arXiv:2607.24645 https://github.com/UKPLab/FEGA
Distill Reasoner (OPD+RLVR) OPDVR (method:opdvr) arXiv:2608.24696 https://github.com/LeapLabTHU/OPDVR
Token-Level Critic (1/prompt) BPCO (method:bpco) arXiv:2608.23566 https://github.com/QPHutu/golden_critic
Posttrain Diffusion DiffusionOPSD (method:diffusion-opsd) / Self-OPD (method:self-opd) arXiv:2608.24646 / 2608.26872 https://github.com/worldbench/DiffusionOPSD / https://github.com/Shiy-Zhang/Self-OPD
Video MLLM RL OraRL (method:orarl) arXiv:2608.20492 https://github.com/HVision-NKU/OraRL
Label-Free Test-Time Reasoning TTPO (method:ttpo) arXiv:2608.27448 https://github.com/ZJU-REAL/TTPO
Unlabeled Reasoner Posttrain u-OPSD (method:u-opsd) arXiv:2608.06296 https://github.com/williamium3000/u-opsd
Privileged-Teacher OPSD VISTA (method:vista) arXiv:2608.28306 none found
Data-Free Self-Evolution J-Zero (method:j-zero) arXiv:2608.26582 https://github.com/GyoukChu/J-Zero
Operator, Regular-Grid PDE CViT (method:cvit) / Poseidon (fine-tune) arXiv:2405.13998 / 2405.19101 https://github.com/PredictiveIntelligenceLab/cvit / https://github.com/camlab-ethz/poseidon
Operator, Industrial CAD Mesh Transolver-3 (method:transolver-3) arXiv:2602.04940 https://github.com/thuml/Transolver-3
Operator, Pretrained Foundation Poseidon (method:poseidon) / Unisolver (method:unisolver) arXiv:2405.19101 / 2405.17527 https://github.com/camlab-ethz/poseidon / https://github.com/thuml/Unisolver
Operator, Physics-Informed PI-CViT (method:pi-cvit) + SOAP arXiv:2606.06164 https://github.com/NanxiiChen/PI-CViT
Operator, Fourier PEFT/Adapt F-Adapter (method:f-adapter) arXiv:2509.23173 https://github.com/fogradio/F-Adapter
Operator, Weather/Climate FourCastNet 3 (method:fourcastnet-3) arXiv:2507.12144 none found (NVIDIA Earth-2)
Industrial Model-Building / Factory Poolside Model Factory (method:poolside-model-factory) Laguna arXiv:2605.27605 https://github.com/pytorch/torchtitan (Titan seed; factory internals closed). Process default, not a train kernel. CISPO/Muon unchanged.
Data-Constrained Pretrain (epoch / r) Repeated-Token Worth + Repetition-Count Selection arXiv:2610.05591 / 2610.05126 none found
Directed SSSP Distances (sparse, comparison-addition) BMSSP (method:bmssp); Dijkstra if you need vertex order or typical (n) arXiv:2504.17033 unofficial https://github.com/hparreao/BMSSP-Python / https://github.com/NicholasCartaxo/BMSSP-SSSP

4. Fast Navigation CLI

Prefer sota / decide over query. --json is unbounded and a superset of prose; --brief is ~10 lines.

# 1. Look up SOTA method, claims, paper, and recipe for any task
python -m library sota "pretrain dense 7B"
python -m library sota "task:math-code-rl-dense" --brief
python -m library decide "task:math-code-rl-dense"

# 2. Rank candidate tasks (scope / redirects). Never empty without near-misses.
python -m library route "pretrain dense 7B"

# 3. Search graph nodes
python -m library query "muon2" --type method

# 4. View node metadata and full prose
python -m library show "method:cispo"

# 5. Inspect incoming and outgoing graph edges
python -m library walk "method:muon2"

# 6. Find shortest traversal path between concepts
python -m library path --from "task:pretrain-dense-7b" --to "recipe:muon2-pretraining"

# 7. Freshness queue (exit 1 if any task current_sota exceeds 120 days)
python -m library stale --max-age-days 120

# 8. Regenerate graph/INDEX.md and the AGENTS.md cheat-sheet (L3 derived; do not hand-edit)
python -m library index

# 9. Validate graph schema and referential integrity (INDEX drift is a warning, exit 0)
python -m library validate

# 10. Transactional supersession (dry-run writes nothing)
python -m library supersede method:new method:old --task task:example --dry-run

# 11. Remote, no-clone usage: the whole graph as ONE file (CI keeps it fresh)
gh api repos/j8ckfi/library/contents/dist/graph.json

For complete operating instructions for AI agents, see AGENTS.md. For the system-level design (abstraction tower, agent contracts, CLI ergonomics roadmap), see docs/system-design.md. For schema rules and supersession protocol, see docs/ontology.md. For adding new research in 5 minutes, see docs/ingestion-guide.md.

About

A machine-readable knowledge graph, autonomously upkept by persistent agents that wake, crawl new papers, and update current state-of-the-art methods.

Topics

Resources

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages