Organizations: Xinjiang University · Tongji University · Institute of Automation, Chinese Academy of Sciences · University of Science and Technology of China
Recent world action models (WAMs) drop future prediction at inference, assuming that representations learned through auxiliary future supervision still benefit action generation. However, our layerwise analysis of Fast-WAM reveals a depth-dependent mismatch: while probes decode future state transitions more accurately from deeper visual layers, the gradients and visual focuses of the future and action objectives increasingly diverge with depth, leaving much of this predictive information unused by the action pathway. To address this issue, we introduce RelayWAM, which adds an explicit transition space between the video and action backbones. Conditioned on the task instruction, robot state, and current action estimate, recurrent transition layers extract step-relevant predictive cues across depth and relay them directly into the action attention, without generating future frames at inference. Experiments show that RelayWAM substantially amplifies the benefit of future supervision, translating into stronger generalization under distribution shifts in both simulation and real-world robot tasks. On LIBERO-Plus, RelayWAM achieves 72.03% mean success, compared to 56.42% without future supervision and 51.50% for Fast-WAM.
Figures & tables
Figure 1: Overview of RelayWAM and the representation–utilization gap. Left: Training-focus attribution reveals that existing future-supervised WAMs suffer from an objective mismatch—action relies on local gripper–object regions while future prediction attends to the whole scene—whereas RelayWAM aligns both focuses. Top right: Linear probing shows that future supervision writes rich predictive cues ( ΔR2 ) into deeper Video DiT layers. Bottom right: RelayWAM introduces an explicit transition space to relay these cues to the Action DiT without test-time future generation, reaching 72.03% success on LIBERO-Plus.
Figure 2: Causal erasure of future information in Fast-WAM vs. RelayWAM. We erase residual future directions from a sliding window of 4 consecutive Video DiT layers read by the action pathway, leaving the video backbone unchanged. Top: held-out R2 of the residual future decoded after future erasure (solid red) versus variance-matched random erasure of the same rank (dashed blue). Bottom: resulting action shift normalized by noise-resampling variance, plotted at the center of each erased window. While Fast-WAM (left) is insensitive to future erasure, RelayWAM (right) shows a pronounced action shift across intermediate-to-deep layers (layers 12–20), confirming active utilization of predictive cues. Shaded bands indicate 95% confidence intervals over episodes.
Figure 3: Action and future supervision diverge across depth in Fast-WAM. Top left: gradient energy projected onto top- k action and future directions shows that the two objectives share subspaces in shallow layers (layer 4) but become nearly orthogonal in deep layers (layer 28). Top right: held-out future R2 decoded from action-sensitive subspaces drops well below random subspaces of equal dimension as depth increases. Bottom: training-focus maps on “open the middle drawer” and top-10% focus concentration averaged over five tasks. Unlike Fast-WAM (solid), RelayWAM (dashed) keeps both objectives focused on task-relevant regions.
Figure 4: Architecture of RelayWAM. Attached to selected intermediate and deep layers of the Video DiT, recurrent transition layers extract step-relevant future information from visual representations conditioned on the task instruction, robot state, current action estimate, and flow time. The extracted cues are progressively accumulated across depth into action-aligned transition tokens Tl , projected into transition keys and values (K^l,V^l) , and injected directly into the native Action DiT attention. This enables action generation to leverage multi-depth predictive representations without generating future observations at inference time.
Paradigm
Method
LIBERO-Plus Success Rate (%)
Bg. ↑
Cam. ↑
Lang. ↑
Light ↑
Layout ↑
Robot ↑
Noise ↑
Total ↑
VLA
OpenVLA [ 26 ]
34.8
0.8
23.0
8.1
28.5
3.5
15.2
15.6
π0 [ 7 ]
81.4
13.8
58.8
85.0
68.9
6.0
79.0
53.6
OpenVLA-OFT [ 25 ]
93.3
56.4
79.5
88.7
74.2
31.9
75.8
69.6
Joint
Fast-WAM-Joint [ 70 ]
44.9
34.0
88.9
90.0
73.4
55.1
33.3
59.1
OpenWAM- α [ 59 ]
87.1
33.8
88.0
97.0
77.5
76.1
39.8
69.2
Table 1: Results on LIBERO-Plus [ 15 ] under distribution shifts – We report success rates (%) across seven perturbation types; Total is the success rate over all evaluation episodes, following [ 15 ] . Joint denotes joint future–action decoding at inference, while FFI denotes future-free inference without explicit future generation at test time. Darker and lighter blue cells mark the best and second-best result in each column.
Figure 5: Real-world apple picking on a Piper robotic arm. Left: qualitative rollouts under out-of-distribution apple placements with leaf occlusion (top) and colored ambient lighting perturbations (bottom). Right: success rates (%) across 20 trials per setting. Without any large-scale embodied pretraining (unlike DM0.5), RelayWAM outperforms Fast-WAM across all settings and maintains strong robustness under severe illumination shifts (70% vs. 15%).
Method
Paradigm
Spatial
Object
Goal
Long
Avg.
π0.5 [ 6 ]
VLA
98.8
98.2
98.0
92.4
96.9
LingBot-VA [ 29 ]
Joint
98.5
99.6
97.2
98.5
98.5
Motus [ 4 ]
Joint
96.8
99.8
96.6
97.6
97.7
Fast-WAM-Joint [ 70 ]
Joint
99.6
99.4
98.2
96.8
98.5
OpenWAM- α [ 59 ]
Joint
99.6
99.6
99.8
98.2
99.3
Light-WAM [ 32 ]
FFI
98.2
99.6
97.8
93.0
97.2
Table 2: Results on LIBERO [ 38 ] . VLA denotes vision-language-action policies without future prediction, Joint denotes joint future–action inference, and FFI denotes future-free inference.
Paradigm
Method
LIBERO-PRO Success Rate (%) ↑
Env.
Lang.
Obj.
Pos.
Task
Avg.
VLA
OpenVLA [ 26 ]
68.0
97.3
93.0
0.0
0.0
51.7
π0 [ 7 ]
38.8
90.5
90.5
0.0
0.0
44.0
π0.5 [ 6 ]
52.8
95.8
96.0
20.8
0.8
53.2
FFI
Fast-WAM [ 70 ]
32.1
95.1
73.8
0.4
2.3
40.7
RelayWAM (Ours)
34.3
93.9
75.2
1.4
18.8
44.7
Table 3: Results on LIBERO-PRO [ 78 ] under challenging distribution shifts – We report success rates (%) across five generalization dimensions. FFI denotes future-free inference without explicit future generation at test time.
Variant
Success Rate (%) ↑
LIBERO
LIBERO-Plus
LIBERO-PRO
A0 (Base)
81.9
46.3
37.0
A1 (+ Future Sup.)
97.6
51.5
40.7
B0 (+ Transition Space)
96.3
56.4
39.8
B1 (RelayWAM)
98.2
72.0
44.7
Table 4: Ablation study – Success rates (%) on LIBERO, LIBERO-Plus, and LIBERO-PRO, reported with the same metric as Tables 1 to 3 . Darker and lighter blue cells mark the best and second-best result in each column.
Figure 6: Controlled 2×2 ablation of future supervision and the transition space. LIBERO-Plus success rate (%) versus layer-20 future predictability ( R2 ) across all four controlled variants. Adding future supervision alone (w/o transition space) raises R2 but yields only a +5.2 -point success gain over the base model (w/o both), whereas introducing the transition space unlocks the utility of predictive features and amplifies the gain to +15.6 points ( 72.03% overall).
Appendix figures & tables2 assets
Supplementary material from the paper’s appendix.
Appendix
Layer
All features
Random
Action-sensitive
ρl
0
0.69
0.66
0.64
0.93
10
0.82
0.65
0.43
0.52
20
0.89
0.75
0.53
0.60
29
0.89
0.71
0.41
0.46
Appendix
Table 5: Future decodability ( Rl2(S) , Eq. ( C3 )) across 8-dimensional subspaces. Random reports the median over 12 covariance-matched draws.
Action shift δ
Error ratio κ
Layers
Future
Random
Future
Random
8–11
0.23
0.17
1.17
1.01
14–17
0.36
0.27
1.35
1.34
18–21
0.44
0.25
1.11
1.09
22–25
0.41
0.39
1.20
1.11
26–29
1.06
0.90
1.28
1.21
Appendix
Table 6: Causal erasure across four-layer windows in Fast-WAM.
College of Computer Science and Artificial Intelligence, Fudan University, Shanghai, China · Singapore Management University, Singapore · Institute of Trustworthy Embodied AI, Fudan University, Shanghai, China