ThinkWAM

World-action model from Xi’an Jiaotong University, Phantrix Research and others that proposes an action chunk, corrects its predicted near, mid and far future latents with that proposal, then re-predicts the executed chunk. No action pretraining. Real-robot results are self-reported: 64.4% mean success across three out-of-distribution conditions on a Galaxea A1Z. The model card is Apache-2.0 and the code MIT, but the repo’s third-party notices put the SIGReg component under a noncommercial license (CC BY-NC 4.0).

Xi’an Jiaotong University et al. · Nov 16, 2026 · World-actionReport an error
98.7%LIBERO · #20 of 83
84.8%LIBERO+ · #13 of 37
1.5BParameters
2Benchmarks reported

Benchmark results2

#13 of 3784.8%, self-reported by the authors, not a leaderboard entry
LIBEROSaturatedSimulation
#20 of 8398.7%, self-reported by the authors, not a leaderboard entry
Not reported on 13 benchmarks
Robustness drop: 98.7% on LIBERO → 84.8% on LIBERO-Plus. The perturbed variants show how much of the LIBERO score is memorization.

Running the model

HardwarePrecisionConfigLatency / chunkStepsVRAMChunkControlSource
A100 80GB–LIBERO checkpoint, two action passes (proposal + refine), batch size 1153.48 ms––16–Nov 2026 ↗
Wrong or missing number? Report an error