Reasoningtracked 0
General reasoning, scientific QA, long-chain inference and planning
Nothing tracked yet
GPQA, MMLU-Pro, BBH, ARC-AGI
ReasoningSOTA RADAR
SOTA
Thirteen capability domains in four families. Each cell answers three things: what ruler measures this domain, who sits on top of the ladder right now, and how far that claim can be trusted. We do not keep a full tool directory — the long tail stays behind the inclusion gate, and only the ladder shows.
- Tracked assets
- 0
- Domains with a ladder
- 0/13
Ladder order: our pick first, then maturity, then recency of the measurement.
Language & Reasoning
4 domains · 0 trackedCodetracked 0
Code generation, repair, refactoring, software engineering tasks
Nothing tracked yet
SWE-bench Verified, LiveCodeBench
CodeMathtracked 0
Competition math, proof assistance, numerical reasoning
Nothing tracked yet
AIME, FrontierMath, MathArena
MathMultimodaltracked 0
Image-text understanding, chart/doc QA, visual reasoning
Nothing tracked yet
MMMU, MathVista, ChartQA, DocVQA
MultimodalVision & Generation
4 domains · 0 trackedImage Generationtracked 0
Text-to-image, image editing, consistency and controllability
Nothing tracked yet
GenAI-Bench, DrawBench, arena preference, controllability
Image GenerationVideo Generationtracked 0
Text/image-to-video, physical and temporal coherence
Nothing tracked yet
VBench, EvalCrafter, physical & temporal coherence
Video Generation3D & World Modelstracked 0
3D asset generation, scene reconstruction, long-horizon world prediction
Nothing tracked yet
Geometric consistency, scene reconstruction, long-horizon prediction
3D & World ModelsSpeech & Audiotracked 0
TTS naturalness, voice cloning, music generation, ASR
Nothing tracked yet
TTS naturalness, voice cloning, music generation, ASR WER
Speech & AudioAgents & Harness
3 domains · 0 trackedAgentstracked 0
Multi-step tool use, web/OS tasks, long-horizon autonomy
Nothing tracked yet
WebArena, OSWorld, GAIA, multi-step tool tasks
AgentsHarness & Evaluationtracked 0
Evaluation infra, reproducibility, failure taxonomy, cost & latency
Nothing tracked yet
Completion rate, failure taxonomy, reproducibility, cost & latency
Harness & EvaluationGames & Simulationtracked 0
Long-horizon policy in resettable environments, open-world game agents
Nothing tracked yet
Atari 100k, Procgen, MineDojo, BALROG
Games & SimulationEmbodied & Safety
2 domains · 0 trackedEmbodied AItracked 0
Real-robot manipulation, navigation, VLA and whole-body control
Nothing tracked yet
CALVIN, LIBERO, Open X-Embodiment, real robots
Embodied AISafety & Alignmenttracked 0
Jailbreak resistance, hallucination rate, tool misuse, long-task boundaries
Nothing tracked yet
Jailbreak rate, hallucination rate, tool misuse, long-task boundaries
Safety & AlignmentHow confidence is graded
Applied only to SOTA claims and benchmark evidence. The letter on the seal says how far a claim is from something we reproduced ourselves — C and D are deliberately drawn dashed and grey so they never read as hard as an A.
ConfirmedTwo or more independent sources, or reproduced by our harness
ContestedSources disagree (data leakage, protocol mismatch)
Vendor ClaimOfficial model card or keynote only, no independent re-test
Needs ReproductionSingle source, not yet verified by us