- Rating
- 3
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
JAM replaces the visual-generation objective of an action policy with prediction of frozen future representations from its existing backbone features. The prediction head and target branch are discarded after training, preserving the corresponding action-only deployment path. The paper first compares objectives within FastWAM and ImageWAM architectures, then applies the recipe to self-supervised and vision-language backbones. Experiments cover LIBERO, a shared LIBERO-Plus subset, RoboTwin, target and head ablations, statistical comparisons, and training-cost estimates.
Critical strengths
-
The fixed-architecture comparisons address a consequential question: whether generators need to retain visual-generation supervision during policy fine-tuning. Table 1 separates the deployed architecture from the training objective and target representation. The best configurations improve LIBERO-Plus from 55.31 to 64.54 percent for Wan and from 65.27 to 79.95 percent for Ovis-U1, without adding inference machinery. The current-offset, future-offset, random-anchor, and VAE-direct controls are particularly useful: they show that future target information matters and that a large pretrained target transformer is not always necessary. This is a meaningful empirical contribution even though future-feature alignment itself has substantial precedent.
-
The main improvements have stronger support than isolated benchmark averages. Table 20 reports paired outcomes and task-bootstrap intervals, and Table 21 repeats the selected JAM configurations across three independent training runs with low observed variability. The paper also provides action-only Qwen and DINO controls, distinguishes local evaluations from published scores, explicitly identifies the 1,506-task OOD subset, and limits cross-backbone conclusions rather than claiming to isolate pretraining alone. These details make the central conclusion credible within the evaluated settings. The main value is the controlled evidence for simplifying generator-based policy training, rather than a fundamentally new alignment loss.
Critical weaknesses
- The headline training-cost reduction mixes the objective change with a caching advantage that may also be available to the generative baselines. In Table 9, FastWAM's 15.52 TFLOPs include 6.54 TFLOPs of online VAE encoding, while cached JAM uses only 0.93 TFLOPs for its current frame. Under the same fixed-preprocessing assumptions, the generative baseline's frozen VAE encodings of recorded futures also appear cacheable; flow-time and noise sampling could still be performed online. The paper transparently reports online and cached JAM costs, but does not provide an equivalently cached FastWAM/ImageWAM reference. Thus, 73 percent is supported for the specific reported pipelines, but does not isolate the saving attributable to replacing visual generation. This is a fixable limitation of an important efficiency claim, rather than evidence against the policy-performance gains.
Questions for authors
-
Can the generative baselines cache their frozen future-frame VAE encodings under the same preprocessing assumptions as JAM? Please report the online cost and measured update time with caching available to both methods, or explain what prevents this. Reporting one-time cache construction and storage would also clarify the practical trade-off.
-
How were the target type, readout offsets, and loss hyperparameters selected before the final LIBERO-Plus evaluation? Were they selected using an independent validation set? Table 20 characterizes fixed-checkpoint uncertainty and Table 21 repeats selected JAM configurations; please clarify whether the generative baselines also have independent training-run estimates. This would distinguish the stability of the training comparison from uncertainty conditional on the selected models and fixed OOD manifest.
Response to Reviewer W7DZ (1/1):
We sincerely thank Reviewer W7DZ for the recognition of our work. The further questions are addressed as follows.
W1, Q1: Can the generative baselines cache their future encodings?
A: Yes, and with both methods cached, foresight still cuts the estimated FLOPs by 57.7% for Wan and 31.9% for Ovis-U1. The abstract, introduction and conclusion now state these figures in place of 73%, and Table 9 adds the cached baselines. A cached FastWAM gives bit-identical latents and losses to online training. For Wan, replacing generation saves the same 5.72 TFLOPs per sample with or without caching: caching cuts FastWAM's VAE cost from 6.54 to 0.93 TFLOPs but leaves 8.93 TFLOPs of trainable visual computation, against 3.17 for JAM. FLOPs per sample and GPU time per update, profiled on one GH200 as in Table 9, each JAM row against its baseline in the same setting:
Building the Wan LIBERO cache once takes 22 minutes on eight GH200s and 16.2 GB of storage. Caching saves 0.27 s of GPU time per update for either method, so the cache is repaid within a quarter of one 20k-update run; a cached FastWAM reads the same cache, so this cost is shared by both methods.
Q2: Were the configurations selected on an independent validation set?
A: Following FastWAM [R1] and ImageWAM [R2], we did not hold out any part of LIBERO or LIBERO-Plus; LIBERO alone returns the selected configurations, and so do 1,246 LIBERO-Plus tasks that were not used in the selection. The baselines keep their published settings. Wan's loss weight, depth and head and Ovis-U1's λ_F=0.3 with [1] were selected on LIBERO, the latter before any LIBERO-Plus evaluation. Targets and readouts are compared in Table 1, and LIBERO picks the same two: each of the three training runs of VAE-direct [1,2] (98.00 to 98.25) and of Yak [1] (96.60 to 96.90) is above every alternative on LIBERO (at most 97.65 and 95.75). We also re-evaluated the generative and foresight checkpoints of Table 1 on a second manifest drawn with seed 43, without re-selection. LIBERO-Plus success (%), new Table 22:
Selecting on the 1,246 tasks outside the paper manifest returns the same two configurations, and there they lead the generative baselines by 10.43 and 13.40 points. Across 10,000 random splits of the 2,752 distinct tasks of the two manifests into validation and test halves, the validation half selects VAE-direct [1,2] and Yak [1] every time.
Q2: Independent training runs for the generative baselines.
A: We added two runs of each baseline. Table 20 measures uncertainty at fixed checkpoints; these runs measure training variability. Mean ± SD over three training runs (revised Table 21):
JAM leads its baseline on LIBERO-Plus by 11.16 points for Wan and 15.78 for Ovis-U1, above the single-run leads of the submission. Every JAM run beats every baseline run on both benchmarks (Table 21): on LIBERO-Plus the weakest JAM run is 8.97 points above the strongest FastWAM run and 13.61 above the strongest ImageWAM run.
References
[R1] Yuan, Tianyuan, et al. "Fast-WAM: Do World Action Models Need Test-time Future Imagination?" arXiv:2603.16666, 2026.
[R2] Zhang, Yuyang, et al. "ImageWAM: Do World Action Models Really Need Video Generation, or Just Image Editing?" CoRL, 2026.