FrontierML

Browse the data

Every graded trial in the benchmark. Scan the heatmap, read the task catalog, or filter the trial table. Every cell opens the agent's full trajectory.

Inspired byInception LabsAR to diffusion conversion

Convert an autoregressive language model to text diffusion

Inspired byStandard IntelligenceInverse dynamics model

Infer the keyboard and mouse actions behind a screen recording

Inspired byKyutaiMusic transcription

Transcribe multi-instrument music into notes labelled with pitch, timing and instrument

Inspired byPangramAI-text detection

Detect whether a passage was written by a person or generated by an AI model, to flag machine-written text

Inspired byPrismMLTernary quantization

Quantize Qwen3.5-4B to ternary weights with minimal accuracy loss

Inspired byTypesafeZero-shot classification

Train a fast calibrated zero-shot text classifier

Inspired byQ LabsData-efficient pretraining

Pretrain a language model from scratch on 10M words

Inspired byNVIDIA DLSS 4Video frame interpolation

Generate intermediate video frames from two endpoints

Inspired byWaterloo NeuralOSScreenshot generation

Generate realistic screenshots from text boxes and a description

Inspired byInduction LabsOCR latent compression

Compress page images into 2 KB of discrete codes that still decode to the full text

Inspired byAxiomMath proof simplification

Simplify Lean 4 proofs

Inspired byJaxcaliburPokémon Random Battle

Build a Pokémon Gen 9 Random Battle bot

Inspired byQuiverAISVG generation

Finetune gpt-oss-20b to draw SVGs

Inspired byThinking MachinesVoice endpointing

Detect when a speaker has finished talking so voice assistants reply promptly without interrupting

Inspired byTencent DiffSpotScreenshot diffing

Describe what changed between two consecutive screenshots in text

Inspired byGoogle ResearchSpeculative drafting

Train a small draft model that speeds up greedy speculative decoding for Qwen3-8B

Inspired byReductoFigure to TikZ

Reconstruct scientific figures as editable TikZ code

Inspired byOpenAI Privacy FilterPII redaction

Redact personal information and secrets from documents

Inspired byPostTrainBenchLLM post-training

Post-train Qwen3-4B-Base into a strong chat model

Inspired byDeepgroveVision for a ternary LM

Add vision to a ternary language model

Inspired byAxiomLean proof validity

Predict Lean 4 proof validity without running Lean