NF-VLA

VLA from the University of Pennsylvania and the Australian National University that generates action chunks with an autoregressive normalizing flow conditioned on VLM features, as a likelihood-based alternative to diffusion and flow-matching heads.

University of Pennsylvania & ANU · Nov 15, 2026 · VLAReport an error
98.3%LIBERO · #35 of 83
74.2%LIBERO+ · #27 of 37
2Benchmarks reported
Nov 15, 2026Released

Benchmark results2

#27 of 3774.2%, self-reported by the authors, not a leaderboard entry
LIBEROSaturatedSimulation
#35 of 8398.3%, self-reported by the authors, not a leaderboard entry
Not reported on 13 benchmarks
Robustness drop: 98.3% on LIBERO → 74.2% on LIBERO-Plus. The perturbed variants show how much of the LIBERO score is memorization.

Running the model

HardwarePrecisionConfigLatency / chunkStepsVRAMChunkControlSource
RTX 4080 SUPER–FP32 with BF16 autocast74.1 ms––10–Nov 2026 ↗
Wrong or missing number? Report an error