CATok Poster — Main Results

CATok

NeurIPS 2026 · arXiv:2609.35469

Main Results

Better success rate, faster inference, fewer steps

95.9%

Avg. success on LIBERO

1.7×

Faster VLA inference than FAST

3.6×

VRR×CR of FAST at equal tokens

~50%

Steps to match the best baseline

We evaluate CATok on three simulation benchmarks—LIBERO, SimplerEnv (Simpler-Bridge), and RoboTwin 2.0—against three representative tokenizer baselines: BIN (uniform per-dimension discretization), FAST (DCT-based compression with BPE tokenization), and OAT (learned fixed-length tokenizer with prefix-based decoding).

Method	LIBERO					Simpler-Bridge					RoboTwin 2.0
Method	Spatial	Object	Goal	Long	Avg.	Spoon	Carrot	Stack	Eggplant	Avg.	Clean	Rand.
BIN	0.586	0.878	0.680	0.604	0.687	0.542	0.333	0.208	0.708	0.448	0.213	0.221
FAST	0.960	0.998	0.962	0.901	0.955	0.500	0.375	0.375	0.625	0.469	0.478	0.478
OAT	0.428	0.876	0.704	0.276	0.571	0.417	0.208	0.167	0.667	0.365	0.229	0.233
CATok (Ours)	0.978	0.994	0.954	0.910	0.959	0.458	0.417	0.542	0.542	0.490	0.489	0.531

Table 1 · Simulation benchmarking results. Accent bold marks the best result per column.

Training Efficiency — **Figure 3 · Training efficiency on Simpler-Bridge.** CATok achieves higher success rates with fewer training steps.

Inference Latency — **Figure 4 · VLA inference latency.** CATok (453 ms) is 1.7× faster than FAST (778 ms) and 3.9× faster than Binning (1747 ms).

Intrinsic Tokenizer Properties — **Figure 5 · Intrinsic properties of action tokenizers.** (a–c) CATok₈ achieves the highest VRR×CR score (16.85), outperforming all baselines at the same compression level. (d–e) Encode and decode latencies on log scale—CATok achieves lower latency than FAST with comparable or better downstream performance.