Literature
Display
← All papers

Rethinking Causal Action Tokenization with Conditional Annealing in Flow Matching

Chenyu Zhang · Yuhang Cao · Daru Du · Yingxi Lu · Jing Shao · Ruoqu Chen · Jiajun Liu · Liu Cao · Yicheng Liu · Hang Zhao · Mengdi Xu

Preprint posted · NeurIPS 2026 (accepted) ↗

Robotics10 min text · CondensedOriginal paper ↗
Contents

At a glance

CATok compresses continuous action chunks into discrete tokens and reconstructs them with a flow-matching decoder. During tokenizer training, a time-dependent suffix mask assigns token positions different roles along the noise-to-action trajectory.

Frozen tokens/decoder then support next-token VLA training. CATok reports the strongest aggregate success among tested tokenizers on three simulation benchmarks and a 20-percentage-point hardware gain over FAST in three suites. Evidence concerns this learned ordering and tested systems; it does not establish physical causal discovery or guarantee preservation of all pretrained language knowledge.

1 Introduction

Uniform action binning creates long sequences; compression methods shorten them but may not provide an order suited to autoregressive prediction. CATok uses conditional annealing to encourage early-token global guidance and later-token refinement.

The policy predicts token prefixes autoregressively. The tokenizer itself uses a bidirectional action/query encoder, followed by quantization and a time-masked decoder. Thus the claimed causal structure concerns learned token order and decoder influence, rather than a causal encoder that sees only earlier physical timesteps.

Complete original Figure 1 : Flow-Matching to Action Tokens. CATok grounds discrete action tokens in the continuous flow matching trajectory. Each token is learned as a stage-wise information increment, imbuing the sequence with a natural coarse-to-fine hierarchy. By construction, these increments follow the temporal evolution of the flow, ensuring a causally-ordered structure that seamlessly aligns with autoregressive generation.
Figure 1: Flow-Matching to Action Tokens. CATok grounds discrete action tokens in the continuous flow matching trajectory. Each token is learned as a stage-wise information increment, imbuing the sequence with a natural coarse-to-fine hierarchy. By construction, these increments follow the temporal evolution of the flow, ensuring a causally-ordered structure that seamlessly aligns with autoregressive generation.

2 Related Works

VLAs can predict discrete action tokens or use continuous action heads. The paper compares binning, FAST’s DCT/BPE compression and OAT’s learned ordered tokens. CATok retains a discrete policy interface with a continuous flow decoder.

Its differences combine a dual-stream encoder, bottleneck quantizer, flow matching and conditional masking. The appendix includes controlled comparisons to separate part of the masking benefit from decoder architecture [13] [15].

3 Method

The method first trains an action-only tokenizer, then freezes it for VLA learning. Observations, language and optional proprioception condition the policy; discrete action codes connect that policy to the frozen decoder.

3.1 Problem Formulation

An encoder maps an H-step action chunk to K finite-codebook indices. A decoder reconstructs continuous actions, while the policy learns teacher-forced next-token prediction on the frozen codes.

The initial reconstruction objective formalizes the goal. Later equations specify the actual flow, spectral and quantization losses; the original numbered and unnumbered displays are preserved.

Equations
ℰ⁡(𝒜)=C=(c1,…,cK),ck∈{1,…,|𝒞|},𝒟⁡(C)≈𝒜,\mathcal{E}(\mathcal{A})=C=(c_{1},\dots,c_{K}),\quad c_{k}\in\{1,\dots,|\mathcal{C}|\},\qquad\mathcal{D}(C)\approx\mathcal{A},
minℰ,𝒟⁡𝔼𝒜​[‖𝒟⁡(ℰ⁡(𝒜))−𝒜‖2].\min_{\mathcal{E},\mathcal{D}}\mathbb{E}_{\mathcal{A}}\big[\|\mathcal{D}(\mathcal{E}(\mathcal{A}))-\mathcal{A}\|^{2}\big].
minθ𝔼(ot,𝒜)[−∑k=1Klogπθ(ck∣ot,c<k)].\min_{\theta}\mathbb{E}_{(o_{t},\mathcal{A})}\Big[-\sum_{k=1}^{K}\log\pi_{\theta}(c_{k}\mid o_{t},c_{<k})\Big].

3.2 CATok Architecture

Quantile-normalized actions pass through a weight-normalized 2D convolutional encoder with positional embeddings. A symmetric multimodal transformer lets action features and K learned queries attend bidirectionally; final query states are quantized.

Each query is projected into a smaller codebook space, assigned its nearest Euclidean code and projected back to the model width. This separates codebook dimensionality from transformer hidden size.

The decoder interpolates from noise at t=0 to action at t=1. At time t, it removes the first floor(tK) token embeddings and retains the suffix. All tokens are available at the start; earlier tokens disappear first, while later ones remain available for refinement. At the terminal endpoint all are masked.

The L1 velocity objective trains the network under that schedule. Stage-dependent token support is partly imposed by construction; learned coarse/fine content still requires the empirical probes below [17] [18].

Complete original Figure 2 : An Overview of CATok . (a) CATok Pipeline. CATok encodes an action chunk with a Dual-Stream Encoder into latent queries, discretizes them via a Bottleneck VQ , and reconstructs actions using a Flow-Matching Decoder with Conditional Annealing .
(b) Conditional Annealing Mechanism progressively masks the first κ ⁡ ( t ) \kappa(t) token embeddings according to flow matching timestep t: darker colors indicate tokens earlier in the sequence, and lighter colors indicate later tokens, assigning each token to a different stage of the flow trajectory.
Figure 2: An Overview of CATok. (a) CATok Pipeline. CATok encodes an action chunk with a Dual-Stream Encoder into latent queries, discretizes them via a Bottleneck VQ, and reconstructs actions using a Flow-Matching Decoder with Conditional Annealing. (b) Conditional Annealing Mechanism progressively masks the first κ⁡(t)\kappa(t) token embeddings according to flow matching timestep t: darker colors indicate tokens earlier in the sequence, and lighter colors indicate later tokens, assigning each token to a different stage of the flow trajectory.
Equations 1–3
ck=arg⁡minj∈{1,…,|𝒞|}⁡‖𝐞j−ψ⁡(𝐪k)‖2,c_{k}=\mathop{\arg\min}_{j\in\{1,\dots,|\mathcal{C}|\}}\;\bigl\|\mathbf{e}_{j}-\psi(\mathbf{q}_{k})\bigr\|_{2}, (1)
κ⁡(t)=⌊t​K⌋,t∈[0,1],\kappa(t)=\lfloor tK\rfloor,\quad t\in[0,1], (2)
ℒFM=𝔼t,x0,x1​[‖vθ​(xt,t,𝒬^>κ⁡(t))−(x1−x0)‖1].\mathcal{L}_{\mathrm{FM}}=\mathbb{E}_{t,x_{0},x_{1}}\left[\left\|v_{\theta}(x_{t},t,\hat{\mathcal{Q}}_{>\kappa(t)})-(x_{1}-x_{0})\right\|_{1}\right]. (3)

3.3 CATok Training and Inference

Tokenizer training combines L1 flow-velocity error, a DCT-domain velocity discrepancy and a commitment term. Codebook entries use exponential-moving-average updates, with rarely used entries reinitialized from batch embeddings.

Equation 6 subtracts the bottleneck projection of a query from its post-projected quantized embedding. The preceding definitions place those terms in different dimensions when the bottleneck is smaller. This notation is retained as printed; the paper does not resolve which space the commitment comparison uses.

Equation 7 describes continuous ODE decoding with the changing suffix mask. The VLA section and architecture caption state single-step decoding. The numerical integration choice and how a single deployment step uses the multistage conditioning schedule are not detailed. Exact loss weights, architectural sizes and codebook-reset thresholds are also not fully specified.

Equations 4–7
ℒ=ℒFM+α​ℒsmooth+β​ℒVQ,\mathcal{L}=\mathcal{L}_{\text{FM}}+\alpha\mathcal{L}_{\text{smooth}}+\beta\mathcal{L}_{\text{VQ}}, (4)
ℒsmooth=𝔼t,x0,x1​‖DCT⁡(x1−x0)−DCT⁡(vθ​(xt,t,𝒬^>κ⁡(t)))‖1,\mathcal{L}_{\text{smooth}}=\mathbb{E}_{t,x_{0},x_{1}}\bigl\|\mathrm{DCT}(x_{1}-x_{0})-\mathrm{DCT}(v_{\theta}(x_{t},t,\hat{\mathcal{Q}}_{>\kappa(t)}))\bigr\|_{1}, (5)
ℒVQ=∑i=1K‖ψ⁡(𝐪i)−sg⁡(𝐪^i)‖22,\mathcal{L}_{\text{VQ}}=\sum_{i=1}^{K}\bigl\|\psi(\mathbf{q}_{i})-\mathrm{sg}(\hat{\mathbf{q}}_{i})\bigr\|_{2}^{2}, (6)
𝐱1=𝐱0+∫01vθ​(𝐱t,t∣𝐐~t)​𝑑t,𝐱0∼𝒩⁡(𝟎,𝐈),\mathbf{x}_{1}=\mathbf{x}_{0}+\int_{0}^{1}v_{\theta}\!\left(\mathbf{x}_{t},t\mid\tilde{\mathbf{Q}}_{t}\right)dt,\qquad\mathbf{x}_{0}\sim\mathcal{N}(\mathbf{0},\mathbf{I}), (7)

3.4 VLA- CATok

Visual, language and proprioceptive tokens enter a VLM that predicts action codes. The tokenizer/codebook/decoder remain frozen, while cross-entropy updates the autoregressive backbone.

This blocks continuous decoder-loss gradients from reaching the VLM during policy training. Action-token cross-entropy still changes the backbone; the experiments do not include an independent language/vision knowledge-retention benchmark. The paper’s broader knowledge-insulation claim should be scoped to this training interface [26].

Complete original Figure 3 : An Overview of VLA- CATok . VLA- CATok feeds visual, language, and proprioceptive tokens into a VLM backbone to autoregressively predict causal action tokens. These tokens are mapped via the learned codebook and processed by a frozen MMDiT flow-matching decoder to produce the final action in a single denoising step. Discrete tokens serve as the sole interface, naturally insulating the VLM’s pretrained knowledge from action-specific gradients.
Figure 3: An Overview of VLA-CATok. VLA-CATok feeds visual, language, and proprioceptive tokens into a VLM backbone to autoregressively predict causal action tokens. These tokens are mapped via the learned codebook and processed by a frozen MMDiT flow-matching decoder to produce the final action in a single denoising step. Discrete tokens serve as the sole interface, naturally insulating the VLM’s pretrained knowledge from action-specific gradients.
Equation 8
ℒA​R=−∑i=1Klogpθ(ci|It,st,l,c<i)\mathcal{L}_{AR}=-\sum^{K}_{i=1}{\log{p_{\theta}\left(c_{i}|I_{t},s_{t},l,c_{<i}\right)}} (8)

4 Experiments

Experiments evaluate downstream simulation/hardware policies, representation probes, tokenizer metrics and ablations. All aggregate and per-task results remain alongside the original experimental figures.

4.1 Experimental Setup

BIN uses 256 bins per scalar, FAST applies DCT/BPE and OAT learns ordered VQ codes. LIBERO policies finetune π0-FAST-base; SimplerEnv and RoboTwin use the same Qwen3-VL-4B-Instruct backbone trained from scratch.

LIBERO evaluates 40 tasks with 50 rollouts each. SimplerEnv tests four WidowX tasks using Bridge training data and 24 rollouts per task. RoboTwin evaluates 50 clean/randomized tasks with 20 rollouts per task and condition.

Hardware uses a Franka robot, 362 picking and 220 stacking demonstrations, and shared Qwen3.5-VL training protocols. CATok tokenizer pre-training additionally uses InternA1 joint data to match FAST reconstruction quality. This is a comparison of the resulting trained pipelines rather than identical tokenizer pre-training data [37] [38] [39].

Complete original Figure 4 : Real-world scene layouts for the three task suites. From left to right: Pick-Spatial , testing spatial generalization by varying cup and plate positions; Pick-Color , testing color generalization with different unseen cup colors;
and Stack-Long , testing long-horizon manipulation under different spatial layouts.
Figure 4: Real-world scene layouts for the three task suites. From left to right: Pick-Spatial, testing spatial generalization by varying cup and plate positions; Pick-Color, testing color generalization with different unseen cup colors; and Stack-Long, testing long-horizon manipulation under different spatial layouts.
Original panel captions
(a) Pick-Spatial
(b) Pick-Color
(c) Stack-Long
Complete original Table 1 : Comparison on robotic manipulation simulation benchmarks. CATok achieves the best overall performance across all benchmarks, with particularly significant improvements on long-horizon tasks such as Long in LIBERO and StackGreenCubeOnYellowCube in SimplerEnv, demonstrating the effectiveness of causal action tokenization for modeling long-range temporal dependencies.
Table 1: Comparison on robotic manipulation simulation benchmarks. CATok achieves the best overall performance across all benchmarks, with particularly significant improvements on long-horizon tasks such as Long in LIBERO and StackGreenCubeOnYellowCube in SimplerEnv, demonstrating the effectiveness of causal action tokenization for modeling long-range temporal dependencies.
Accessible table and original citation links
LIBERO SimplerEnv RoboTwin 2.0
Method Spatial Object Goal Long Avg. Spoon Carrot Stack Eggplant Avg. Clean Randomized
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 0.978 0.994 0.954 0.910 0.959 0.458 0.417 0.542 0.542 0.490 0.489 0.531

4.2 Can CATok Help VLA Training?

Aggregate simulation success is 95.9% on LIBERO, 49.0% on SimplerEnv, and 48.9/53.1% on clean/randomized RoboTwin. Corresponding FAST values are 95.5%, 46.9% and 47.8/47.8%. Gains vary substantially across settings.

CATok does not win every task: FAST is stronger on LIBERO Object/Goal and SimplerEnv Spoon/Eggplant. SimplerEnv stacking rises from 37.5% to 54.2%, a 16.7-percentage-point gain, about 44.5% relative. The source’s percentage wording should not merge those quantities.

The smoothed training curve reaches FAST’s best plotted result at about half the steps. Figure 5 reports 453 ms VLA latency versus FAST’s 778 ms, about 1.7× faster; OAT is faster still at 266 ms but much less successful. Step counts are not a full wall-clock training-cost comparison.

Hardware scores 65/50/47.5% versus FAST’s 45/30/27.5%. The project specifies 20 spatial, ten unseen-red-cup and 40 stacking rollouts. These are small tests on one robot, without reported confidence intervals or independent training-seed variability.

Complete original Figure 5 : (a) Training efficiency on the SimplerEnv benchmark. Curves are Gaussian-smoothed to reduce evaluation stochasticity without changing the relative ranking of different methods. Notably, CATok reaches the best baseline performance using only around 50% of the training steps. (b) VLA inference latency. By jointly considering autoregressive generation and decoding overhead, CATok achieves near-minimal end-to-end inference latency while attaining the best success rate among all methods.
Figure 5: (a) Training efficiency on the SimplerEnv benchmark. Curves are Gaussian-smoothed to reduce evaluation stochasticity without changing the relative ranking of different methods. Notably, CATok reaches the best baseline performance using only around 50% of the training steps. (b) VLA inference latency. By jointly considering autoregressive generation and decoding overhead, CATok achieves near-minimal end-to-end inference latency while attaining the best success rate among all methods.
Complete original Table 2 : Real-world manipulation results. CATok consistently outperforms FAST across spatial, color, and long-horizon generalization tasks.
Table 2: Real-world manipulation results. CATok consistently outperforms FAST across spatial, color, and long-horizon generalization tasks.
Accessible table and original citation links
Method Pick-Spatial Pick-Color Stack-Long
FAST 0.450 0.300 0.275
CATok 0.650 0.500 0.475

4.3 Does CATok Produce Causal Tokens with Generative Semantics?

An action-only autoregressive probe measures next-token entropy on held-out mixed action chunks. CATok entropy falls in forward order and rises in reverse; this supports a directional predictive organization. Entropy also depends on positional code distributions and model fitting, so that trend alone is not a proof of causal dependence.

t-SNE of embeddings from 32,768 normalized LIBERO chunks shows slot-dependent geometry. Prefix reconstructions at 1, 6, 11 and 16 tokens show structured trajectory changes. These are representation/decoder probes rather than guarantees that every token has a human-interpretable semantic meaning.

Native-decoder swaps and removals in Appendix C.1 test positional influence directly. Their triangular time support follows the masking schedule; variation within active stages supplies the additional learned-behavior evidence.

Complete original Figure 6 : CATok Produces Causal Tokens with Generative Semantics . (a) Causal Ordering CATok ’s entropy decreases in the forward order ( CATOK fwd ) and increases in the reverse order ( CATOK rev ), while FAST and Bin show weaker positional structure. (b) Token embedding geometry. T-SNE map shows slot-dependent organization, progressing from smooth early-token regions to compact late-token clusters. (c) Prefix reconstruction. Increasing token prefixes produce structured trajectory updates, with dashed curves showing incremental contributions.
Figure 6: CATok Produces Causal Tokens with Generative Semantics. (a) Causal Ordering CATok ’s entropy decreases in the forward order (CATOK fwd) and increases in the reverse order (CATOK rev), while FAST and Bin show weaker positional structure. (b) Token embedding geometry. T-SNE map shows slot-dependent organization, progressing from smooth early-token regions to compact late-token clusters. (c) Prefix reconstruction. Increasing token prefixes produce structured trajectory updates, with dashed curves showing incremental contributions.

4.4 Intrinsic Properties of CATok

Valid Reconstruction Rate counts chunks below a reconstruction threshold, and Compression Rate divides raw scalar count by token count. Their product summarizes one chosen fidelity/compression trade-off.

CATok with eight tokens reaches the best plotted product, 16.85; sixteen tokens are used in the main LIBERO policy. These are different operating points. Figure 7’s 3.6× comparison concerns the product metric rather than a 3.6× improvement in task success or reconstruction accuracy alone.

CATok tokenization is faster than FAST in the plotted setting, while binning is intrinsically cheaper and OAT decoding remains competitive. End-to-end policy latency additionally includes autoregressive generation. Exact metric and timing limitations remain in Appendix B.

Complete original Figure 7 : Intrinsic properties of action tokenizers. (a–c) Fidelity vs. compression efficiency. VRR measures reconstruction quality; CR measures compression ratio; VRR × \times CR jointly captures the fidelity–compression tradeoff. CATok 8 achieves the highest VRR × \times CR ( 16.85 16.85 ), outperforming all baselines at the same compression level. (d–e) Inference efficiency. Encode and decode latencies are shown on a log scale.
The subscript k k in OAT k and CATok k denotes the number of tokens.
Figure 7: Intrinsic properties of action tokenizers. (a–c) Fidelity vs. compression efficiency. VRR measures reconstruction quality; CR measures compression ratio; VRR×\timesCR jointly captures the fidelity–compression tradeoff. CATok8 achieves the highest VRR×\timesCR (16.8516.85), outperforming all baselines at the same compression level. (d–e) Inference efficiency. Encode and decode latencies are shown on a log scale. The subscript kk in OATk and CATokk denotes the number of tokens.

4.5 Ablations and Analysis

Ablations compare annealing, random masking, decoder architecture and token count. Full numerical controls and all token-budget plots are retained in Appendix C.

5 Conclusion

CATok combines compact autoregressive action prediction with a frozen flow decoder, with gains on the evaluated aggregate policy metrics and a useful stage-ordered inductive bias.

Limits include one hardware platform, constrained datasets, no broad cross-embodiment scaling study, limited uncertainty reporting and incomplete implementation specifications. The source’s causality and knowledge-preservation language is stronger than the specific measurements establish.

Original project comparisons

The authors’ project page includes six original successful rollouts, comparing CATok and FAST on spatial picking, unseen-red-cup picking and stacking. All files are unchanged and stored locally with generated preview frames.

These selected successes illustrate execution rather than supply additional unbiased trials. The public code and project’s detailed donor-swap explanation remain linked.

CATok · Pick-Spatial · Selected successful rollout
CATok · Pick-Color (red) · Selected successful rollout
CATok · Stack-Long · Selected successful rollout
FAST · Pick-Spatial · Selected successful rollout
FAST · Pick-Color (red) · Selected successful rollout
FAST · Stack-Long · Selected successful rollout

Appendix A Implementation Details

LIBERO/Bridge tokenizer training mixes the datasets at weights 1:5 before separate policy training; RoboTwin trains its tokenizer and policy on its full data. Action horizons are 20/10/20 and token counts 16/16/32, with codebook size 4,096.

The complete table retains tokenizer steps, batch sizes, learning rates, warm-up, weight decay and four-H100 policy training configuration. It does not supply a complete latency measurement setup or total training time.

Complete original Table 3 : Benchmark-specific implementation details.
Table 3: Benchmark-specific implementation details.
Accessible table and original citation links
Setting LIBERO Simpler-Bridge RoboTwin 2.0
Benchmark and embodiment
Embodiment Franka WidowX Bimanual ALOHA
Action DoF 7 7 14
Training data Official LIBERO dataset Bridge dataset Clean and randomized RoboTwin 2.0 trajectories
Evaluation protocol 4 suites
10 tasks per suite
50 rollouts per task
4 WidowX tasks
in SimplerEnv
24 rollouts per task
50 tasks
20 rollouts per task
Action tokenizer configuration
Action horizon HH 20 10 20
Number of action tokens KK 16 16 32
Codebook size |𝒞||\mathcal{C}| 4096 4096 4096
Tokenizer training steps 300 000 300 000 200 000
Tokenizer batch size 512 512 512
Tokenizer learning rate 1×10−41\times 10^{-4} 1×10−41\times 10^{-4} 1×10−41\times 10^{-4}
Tokenizer weight decay 1×10−21\times 10^{-2} 1×10−21\times 10^{-2} 1×10−21\times 10^{-2}
Tokenizer warmup steps 1000 1000 1000
VLA training configuration
VLA backbone π0\pi_{0}-FAST-base Qwen3-VL-4B-Instruct Qwen3-VL-4B-Instruct
VLA training steps 30 000 60 000 40 000
VLA batch size 32 64 32
VLA learning rate 2.5×10−52.5\times 10^{-5} 1×10−51\times 10^{-5} 1×10−51\times 10^{-5}
VLA weight decay 1×10−101\times 10^{-10} 1×10−81\times 10^{-8} 1×10−81\times 10^{-8}
Number of GPUs 4×\timesH100 GPUs 4×\timesH100 GPUs 4×\timesH100 GPUs

Appendix B Evaluation Metrics

The appendix defines the intrinsic tokenizer metrics and timing protocol. These quantities evaluate action reconstruction and computation, rather than physical task success.

B.1 Valid Reconstruction Rate (VRR)

VRR uses the norm of the complete reconstructed action-chunk error divided by horizon H, then thresholds it at σ. The norm spans both time and action coordinates.

The expression is not simply a mean of per-timestep norms or a dimension-normalized error. This version does not state the numerical threshold σ, limiting reproduction and comparison across action dimensions or horizons.

Equation 9
VRR=NvalidNtotal,Nvalid=∑i=1Ntotal𝟏​(‖Aipred−Aigt‖2H<σ),\displaystyle\mathrm{VRR}=\frac{N_{\mathrm{valid}}}{N_{\mathrm{total}}},\quad N_{\mathrm{valid}}=\sum_{i=1}^{N_{\mathrm{total}}}\mathbf{1}\left(\frac{\|A_{i}^{\mathrm{pred}}-A_{i}^{\mathrm{gt}}\|_{2}}{H}<\sigma\right), (9)

B.2 Compression Rate (CR)

CR divides B×H×D raw action scalars by the discrete token count. Higher values mean fewer tokens per scalar representation.

This is an element/token count ratio, not a byte-storage or bitrate comparison: scalar precision, code-index bits, codebook storage and decoder parameters are outside the formula.

Equation 10
CR=NrawNtoken,Nraw=B×H×D,\displaystyle\mathrm{CR}=\frac{N_{\mathrm{raw}}}{N_{\mathrm{token}}},\quad N_{\mathrm{raw}}=B\times H\times D, (10)

B.3 Encoding and Decoding Latency

Encoding measures action-to-code conversion and decoding measures code-to-action reconstruction. The paper averages timings across batches under a shared setting.

It does not report the exact intrinsic-timing batch size, hardware/device configuration, synchronization or warm-up procedure. Table 3’s four-H100 training allocation does not itself specify those measurement details.

Appendix C Ablations and Analysis

The full appendix separates conditional annealing from selected architectural controls, then examines token budget and decoder interventions. All four ablation rows and every plotted token count are retained.

Conditional Annealing and Architecture Ablations.

Removing conditional annealing lowers LIBERO mean success from 95.9% to 91.4%. A parameter-comparable VQ-VAE/MMDiT with random masking scores 73.7%, versus 48.6% without masking.

The A-versus-B comparison is the clearest direct annealing ablation. The additional VQ-VAE comparisons change reconstruction/training design as well as masking, so the entire gap should not be assigned to token causality alone.

Complete original Table 4: Controlled ablation of conditional annealing, random masking, and decoder architecture.
Table 4: Controlled ablation of conditional annealing, random masking, and decoder architecture.
Accessible table and original citation links
Method Spatial Object Goal Long Avg.
A - Ours 0.978 0.994 0.954 0.910 0.959
B - Ours w/o conditional annealing 0.920 0.980 0.918 0.834 0.914
C - VQ-VAE + MMDiT + random masking 0.724 0.924 0.756 0.542 0.737
D - VQ-VAE + MMDiT 0.614 0.796 0.296 0.240 0.486

Number of Tokens.

The study sweeps 8, 16, 32, 64 and 128 tokens, holding other settings fixed. Sixteen gives the strongest aggregate LIBERO performance; larger budgets degrade it.

The authors attribute this to harder autoregressive modeling of long sequences. This is a plausible interpretation of the tested sweep, rather than proof that sixteen is optimal across datasets, action horizons or model scales. Some individual suites do not improve uniformly from eight to sixteen.

Complete original Figure 8 : Effect of the number of action tokens on downstream VLA performance on the LIBERO benchmark. Performance improves from 8 8 to 16 16 tokens but degrades with larger token budgets, with 16 16 tokens achieving the best overall performance.
Figure 8: Effect of the number of action tokens on downstream VLA performance on the LIBERO benchmark. Performance improves from 88 to 1616 tokens but degrades with larger token budgets, with 1616 tokens achieving the best overall performance.

C.1 Causality Validation

Frozen-native-decoder interventions replace one token while holding the others fixed, measuring changes in generated actions or velocity. This establishes sensitivity within the trained decoder, with its masking rules fixed.

Donor-Swap Intervention.

Donor swaps substitute another chunk’s embedding at the same slot and evaluate the change over flow time. Same-slot donors preserve the positional embedding distribution better than arbitrary replacement.

The project further describes fixed state/time/noise and normalization within each token’s active interval. A token cannot influence stages where the architecture masks it, so triangular support alone verifies the schedule.

Single-Token Removal.

Single-token removal uses a neutral/zero embedding and compares with complete-code reconstruction. The heatmap shows position-dependent joint/gripper influence.

A zero embedding can fall outside normal encoded examples. These controlled model interventions characterize decoder roles without proving causal relations among physical actions or environment states.

Results.

Early slots show stronger coarse task/gripper effects and earlier-stage influence; later slots contribute joint/trajectory refinements. The combined geometry, reconstruction and intervention evidence supports a useful ordered representation.

That evidence should be distinguished from a universal semantic interpretation, an autoregressive encoder factorization or an identified real-world causal model.

Complete original Figure 9 : Causality validation via native-decoder interventions. Both donor-swap (a) and single-token removal (b) exhibit a clear position-dependent structure: early tokens have stronger influence on coarse action components and earlier generation stages, whereas later tokens increasingly affect fine-grained action details, confirming that the learned token ordering corresponds to meaningful causal roles rather than an arbitrary positional ordering.
Figure 9: Causality validation via native-decoder interventions. Both donor-swap (a) and single-token removal (b) exhibit a clear position-dependent structure: early tokens have stronger influence on coarse action components and earlier generation stages, whereas later tokens increasingly affect fine-grained action details, confirming that the learned token ordering corresponds to meaningful causal roles rather than an arbitrary positional ordering.
Original panel captions
(a) Donor-swap influence matrix Δ⁡(t)\Delta(t). Replacing token kk with a donor’s same-position embedding yields a position- and flow-stage-dependent influence map with triangular support, confirming the conditional-annealing schedule.
(b) Single-token removal. Zeroing token kk and decoding the intervened sequence isolates each position’s generative contribution; the per-dimension footprint progresses from coarse gripper/task-state information (early tokens) to fine joint/trajectory-detail information (late tokens).

Appendix D Full Robotwin 2.0 Benchmark

The full 50-task table preserves clean/randomized success for all four methods, including many ties, reversals and zero-success tasks. CATok’s aggregate advantage coexists with weaker performance than FAST on several individual tasks.

The source prints FAST randomized Place Container Plate at 96%, although 20 single-run trials would permit only five-percentage-point increments. It also prints the resulting average as 47.82%, rounded to 47.8% in Table 1. Values remain unchanged; no pooling/repetition explanation is supplied.

Complete original Table 5 : Per-task success rates on RoboTwin 2.0.
Table 5: Per-task success rates on RoboTwin 2.0.
Accessible table and original citation links
Simulation Task BIN FAST OAT CATok
Clean Rand. Clean Rand. Clean Rand. Clean Rand.
Adjust Bottle 80% 95% 90% 100% 100% 100% 90% 100%
Beat Block Hammer 15% 20% 25% 50% 15% 5% 65% 65%
Blocks Ranking RGB 0% 0% 10% 10% 5% 0% 5% 0%
Blocks Ranking Size 0% 0% 0% 5% 5% 0% 5% 10%
Click Alarmclock 65% 60% 90% 75% 60% 70% 65% 80%
Click Bell 95% 85% 80% 80% 75% 50% 95% 95%
Dump Bin Bigbin 30% 30% 95% 75% 15% 20% 65% 70%
Grab Roller 65% 50% 100% 100% 55% 60% 100% 100%
Handover Block 0% 0% 0% 0% 0% 0% 10% 10%
Handover Mic 10% 10% 55% 50% 5% 10% 60% 65%
Hanging Mug 0% 0% 10% 15% 0% 0% 15% 5%
Lift Pot 25% 0% 45% 45% 0% 0% 30% 40%
Move Can Pot 0% 0% 25% 50% 15% 15% 50% 55%
Move Pillbottle Pad 0% 10% 55% 40% 15% 25% 50% 70%
Move Playingcard Away 30% 65% 65% 80% 45% 60% 90% 85%
Move Stapler Pad 0% 5% 10% 10% 5% 0% 15% 10%
Open Laptop 0% 15% 70% 55% 20% 20% 75% 90%
Open Microwave 0% 5% 15% 20% 20% 15% 20% 35%
Pick Diverse Bottles 0% 0% 40% 45% 0% 5% 35% 40%
Pick Dual Bottles 0% 0% 30% 55% 0% 5% 60% 40%
Place A2B Left 40% 30% 85% 80% 40% 40% 75% 80%
Place A2B Right 35% 25% 65% 70% 45% 35% 65% 70%
Place Bread Basket 25% 20% 70% 40% 20% 20% 45% 65%
Place Bread Skillet 40% 35% 50% 50% 15% 10% 45% 60%
Place Burger Fries 5% 10% 80% 70% 25% 40% 80% 80%
Place Can Basket 0% 5% 20% 30% 0% 5% 30% 50%
Place Cans Plasticbox 0% 0% 50% 25% 0% 0% 30% 35%
Place Container Plate 55% 60% 85% 96% 65% 50% 90% 90%
Place Dual Shoes 5% 0% 10% 20% 0% 0% 0% 15%
Place Empty Cup 35% 35% 80% 85% 35% 10% 50% 75%
Place Fan 0% 10% 40% 40% 5% 5% 25% 30%
Place Mouse Pad 5% 10% 30% 30% 5% 5% 30% 55%
Place Object Basket 5% 5% 45% 25% 20% 25% 50% 60%
Place Object Scale 10% 30% 70% 45% 20% 30% 40% 50%
Place Object Stand 35% 45% 80% 90% 35% 50% 75% 85%
Place Phone Stand 25% 0% 65% 55% 10% 0% 40% 75%
Place Shoe 25% 25% 50% 40% 25% 50% 75% 50%
Press Stapler 45% 45% 60% 55% 55% 75% 85% 85%
Put Bottles Dustbin 0% 0% 5% 0% 0% 0% 5% 0%
Put Object Cabinet 0% 0% 25% 35% 0% 0% 45% 40%
Rotate QRcode 10% 10% 30% 35% 0% 5% 60% 35%
Scan Object 0% 10% 20% 25% 0% 5% 25% 20%
Shake Bottle 95% 85% 100% 100% 85% 100% 90% 95%
Shake Bottle Horizontally 90% 95% 100% 95% 90% 85% 100% 85%
Stack Blocks Three 0% 0% 0% 10% 0% 0% 0% 0%
Stack Blocks Two 10% 10% 50% 40% 15% 5% 40% 50%
Stack Bowls Three 0% 0% 10% 10% 0% 0% 20% 15%
Stack Bowls Two 25% 25% 40% 75% 45% 20% 50% 60%
Stamp Seal 5% 15% 25% 40% 15% 20% 40% 40%
Turn Switch 25% 15% 40% 20% 20% 15% 40% 40%
Average 21.3% 22.1% 47.8% 47.82% 22.9% 23.3% 48.9% 53.1%
Concept demonstration

Open image ↗