- Rating
- 1
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
This paper proposes DeltaWAM, a World-Action Model that replaces full-frame or full-latent future prediction with a single delta token per future step, reusing the representation from DeltaWorld [25]. Two components are added: Gated AdaLN injection of frozen-T5 instruction embeddings into every DeltaWorld block, making the originally unconditional world model task-conditioned, and a layer-wise KV adapter that feeds the current observation's DINO patch features together with the predicted future delta tokens into a flow-matching Action DiT. The model is evaluated on the four standard LIBERO suites, on the task-perturbation dimension of LIBERO-Pro, and on a subset of LIBERO-Plus, with emphasis on its small parameter count, low latency and low peak memory.
Critical strengths
-
The motivation is clear and matches the form of the method. Consecutive frames in manipulation do share most of their context and differ in low-dimensional structured ways, and representing foresight as transitions rather than scenes follows from that observation directly.
-
Bringing DeltaWorld's delta-token representation into world-action modeling is a worthwhile direction. Because the representation is already transition-centric, the pipeline avoids intermediate machinery that comparable methods require, such as the latent-action distillation stage in LaWAM [10].
Critical weaknesses
-
The ablations in Section 4.4 do not test the claimed contribution. The third contribution states that compressed delta tokens beat full state predictions, but the decoded-future-DINO variant uses the same DeltaWorld predictions merely decoded back through the frozen tokenizer, so it differs in two uncontrolled ways at once, decoder reconstruction error (which Figure 10 shows is nonzero even given ground-truth tokens) and an order-of-magnitude longer conditioning sequence, and is not full-state prediction in either case. The test-time ablation in Figure 6(b) has the same problem: zeroing the future tokens of a policy trained to consume them corrupts its input distribution, so its collapse is expected rather than evidence that foresight is necessary, and the paper's own current-DINO-only variant, trained without future conditioning, retains most of the performance.
-
LIBERO performance is far below the field and the closest baseline is cited but never compared against. The reported average falls below every world-action model in Table 1 and below pi0. LiLa-WAM [50] is absent from all comparison tables although it is also a latent WAM built on frozen DINOv3 without embodied pretraining, and it reaches roughly 97 percent at a smaller total parameter count and lower latency, with its largest margin on LIBERO-Long; Light-WAM [31] is likewise omitted at a comparable success rate and a peak memory footprint close to this submission's, which removes the distinctiveness of the memory claim. Given a gap of several points against methods of comparable or smaller size, the submission sits off the accuracy-cost Pareto frontier and the parameter savings do not buy a usable trade-off.
-
Reporting on both robustness benchmarks is selective and unexplained. LIBERO-Pro defines five perturbation dimensions and Section 4.3 reports one (Task); LIBERO-Plus defines seven and Appendix B reports four, omitting language, background and noise. On the official LIBERO-Pro leaderboard, baselines remain largely robust on the object and semantic dimensions, so the claim in Section 4.3 that near-zero baseline scores reveal mechanical memorization holds on only two of the five. Appendix B states that its baselines come from [15, 54], and [54] reports all seven LIBERO-Plus dimensions, so the omission is an editorial choice rather than a data limitation.
-
The relationship between delta tokens and latent actions is never discussed, although the two are close and the relevant work is cited. UniVLA [7] learns latent actions with an IDM encoder and FDM decoder over pairs of DINOv2 feature maps under an embedding-reconstruction objective with a quantized bottleneck; DeltaTok is an encoder-decoder over pairs of DINOv3 feature maps under a feature-reconstruction objective with a continuous single-token bottleneck. UniVLA appears in Appendix B as a baseline, yet neither Section 2 nor Section 3.1 states what distinguishes the two beyond continuity and token count. Without that discussion it is unclear whether the first contribution is a new factorization or a re-parameterization of an established one.
Questions for authors
-
Will you provide a token-matched full-state control, for instance future DINO patch features pooled or projected to the same dimensionality and token count as the delta tokens, holding sequence length and decoder involvement fixed? Without it, does the paper still claim that a change-centric inductive bias is superior rather than that compression helps?
-
Why are LiLa-WAM [50] and Light-WAM [31] excluded from Table 1, given that both are in the bibliography, both predate the submission by months, and both exceed DeltaWAM at comparable or smaller size without embodied pretraining? If you consider the protocols incomparable, please say so explicitly.
-
What determined the reported subsets of LIBERO-Pro (one of five dimensions) and LIBERO-Plus (four of seven)? Since [54] already supplies all seven LIBERO-Plus dimensions, can you report the omitted ones, in particular background, object and semantic, which most directly test the claim that delta tokens filter out shared static context?
-
How does a delta token differ conceptually from a latent action as learned in UniVLA [7] or distilled in LaWAM [10], given that all are low-dimensional codes extracted from consecutive VFM feature pairs under a reconstruction objective? If the distinction is that delta tokens are autoregressively forecast as a world model rather than consumed as a policy action space, please state that and position the first contribution accordingly.