Paper review

I priced five claims and ran the two I could afford. The released RFT dataset shows a cap rather than a rebalance and a strong horizon bias, and the released checkpoint scored 0 of 40 on the task the report puts at 1.00.

Honest and reproducible, but correlated-noise flow matching is asserted rather than ablated. Weak reject for a main track, accept as a competition report.

72% of BEHAVIOR goal predicates are spatial relations. Linear probes on frozen SigLIP sit at 0.51 to 0.53 on them at the 256-token budget the baselines run, and more tokens do not close the gap.

Strongest in data engineering and as the first open humanoid VLA. The dual-system framing is more marketing than architecture, and N1.5 and N1.6 walked away from it within nine months.

A useful open release judged on its authors’ own benchmark, with a three-trial robotics experiment and no ablation of the recipe. Weak accept.

The first VLA shown working in homes it never saw. Accept, with one standard benchmark added and the PaliGemma backbone moved out of the appendix.

On Berkeley cable routing, FAST+ keeps a median 81% of a custom-fit tokenizer’s compression, right at my threshold, and the answer turns on how the custom tokenizer is tuned.