{"id":"c0718bfe-ff60-4d0a-8cd9-b181bf3ba8f2","arxiv_id":"2610.18568","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"WAVE-0 trains a 257M latent world model on ground-truth robot actions and uses its predicted future visual representations to condition the action head of a pretrained VLA, reporting gains of 11 points on RoboTwin 2.0 and 35 points on real-world tasks over the base policy.","lead":"Robotics researchers present WAVE-0, a way to give a vision-language-action robot policy a small world model that predicts the visual consequences of its planned actions before the action head generates commands. On simulation benchmarks and three real robot tasks it reports consistent gains over the base policy it wraps, including a jump from 38% to 73% success in real-world trials.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek/deepseek-v4-flash","headline":"Additive dynamics in a frozen SigLIP space is claimed but never measured; the loop may be a high-capacity shortcut rather than action-consequence prediction.","rationale":"The reader already flagged the unverified additivity and the missing L_future/L_delta, so this stress-test is an extension rather than a new objection. I agree on the other two reasons for CONDITIONAL (no error bars, baseline provenance), but the strongest single point of the paper is the causal claim that predicted futures 'directly guide action generation.' For that mechanism to be true, the world model must genuinely predict future SigLIP states; L_align and L_rank can both be satisfied while the predicted future is a poor approximation of the true future. In the extreme, the model could learn a high-capacity, action-sensitive but temporally trivial map and the expert would still benefit from extra conditioning capacity, exactly as pi0.5+WM benefits from auxiliary supervision. The paper's own Fig. 4 is consistent with this: it shows action sensitivity but never prediction accuracy. Note also that a low L_future is a necessary but not sufficient condition: if the base representation changes little over the horizon (e.g., short horizons), even a trivial near-identity map would achieve low loss, so a copy-baseline comparison is required. I did not find circularity or missing definition of Eq. (3)-(12); every loss uses observed ground-truth future frames through the frozen encoder or ground-truth actions. There is no formal verification and the paper's own classifier labels no limitations other than the latency caveat, suggesting the authors should be asked to add the fidelity metrics I propose.","tokens_in":20890,"tokens_out":1714,"duration_ms":12438,"concrete_test":"Fixed checkpoint, compare cosine similarity between bS_{t+h} and S_{t+h} versus S_t and S_{t+h} at the same offsets (e.g., h=9, 31, 32) on held-out episodes of RoboTwin 2.0 and LIBERO. If the prediction adds nothing over the current state (the gain is under ~5-10%), the mechanism is not anticipating consequences. If it materially beats copying the current state at the same h and tracked across tokens, the additive-reachability concern is dissolved.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central causal claim is that predicted futures condition the action expert and that the resulting gains come from anticipating action consequences. The mechanism depends on Eq. (4): bS_{t+h} = S_t + W_phi(S_t, z_s, Z_a). Since the target is a fixed SigLIP representation, the world model is only useful if this additive residual parametrization actually captures the action-dependent part of the representation change. The paper never reports L_future, L_delta, or achieved cosine distances between predicted and ground-truth futures; Fig. 4(a,b) reports only ranking gaps. The reader correctly identified that the counterfactual ranking gap (0.1648 vs. 0.0020) proves sensitivity to actions but does not prove accuracy of the predicted future. The claim in Section 4.3 that gains on long-horizon tasks substantiate 'anticipating the consequences of latent action plans' is therefore a causal attribution, not an established fact: the measured alternative explanation is that Stage-2 gains come from added capacity and auxiliary supervision (pi0.5+WM already gains 89.51% on RoboTwin 2.0), and from shortcut-free training signal rather than from a faithful future. One way the mechanism could be false while results hold: the base representation collapses to a near-constant |bS| in the relevant future window, so the residual W_phi degenerates to a high-capacity action-conditioned feature vector that is a good regression target for L_align but carries no future information; the expert then uses it as extra capacity. The absence of any reported absolute prediction-fidelity number makes this indistinguishable from the intended mechanism.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek/deepseek-v4-flash","summary":"WAVE-0 is a two-stage framework that couples a pretrained VLA (π0.5) with a lightweight (257M-parameter) latent world model. In Stage 1, a world model predicts future SigLIP representations from current observations, proprioception, and ground-truth action sequences through an additive residual (Eq. 4), trained with future-representation, residual, and counterfactual ranking losses (Eqs. 5-8). In Stage 2, the world model is frozen; the PaliGemma planner is trained to emit latent action plans aligned with the Stage-1 action encoder (Eq. 10), the frozen world model converts these plans into predicted future representations, and a resampler produces conditioning tokens for the action expert, which is trained with a flow-matching objective plus world-model and alignment auxiliary losses (Eq. 12). The predicted futures condition the expert at both training and inference. On RoboTwin 2.0, WAVE-0 improves average success from π0.5's 80.38% to 91.47%; on LIBERO from 96.9% to 98.7%; on LIBERO-Plus, LIBERO-Pro, and LIBERO-Para it reports 88.0%, 61.4%, and 75.0%; and on three real-world tasks it reports an improvement from 38.3% to 73.3%. The paper also argues that WAVE-0 is markedly more sample-efficient than LaWAM, processing 33.28M versus 345.60M examples across the three real-world tasks.","tokens_in":21202,"tokens_out":23977,"duration_ms":141927,"significance":"If the central mechanism holds, this is a strong and practically useful contribution. It shows that a modest world model trained only on in-domain action-labeled demonstrations can substitute for expensive video-based future prediction, and it provides one of the cleanest available ablations of future-conditioning versus auxiliary world-model supervision: on RoboTwin 2.0, π0.5 scores 80.38%, π0.5+WM with auxiliary supervision 89.51%, and WAVE-0 with explicit future conditioning 91.47% (Table 1b); the Stage-2 objective ablation in Fig. 4(d) matches the prose. The counterfactual ranking objective is concrete and falsifiable, and its effect is directly measured (ranking gap 0.1648 with L_rank versus 0.0020 without; margin satisfaction 89.5% versus 0%). The paper ships a repository and project page, reports full per-task results on 50 RoboTwin tasks (Table 13), and gives concrete GPU-hour and sample-efficiency accounting for the LaWAM comparison.","major_comments":[{"comment":"The central causal claim—that 'anticipated action consequences directly guide action generation' (abstract) and that long-horizon gains reflect 'anticipating the consequences of latent action plans' (§4.3)—requires that the predicted future Ŝ_{t+h} be an accurate function of the ground-truth future S_{t+h} in the frozen SigLIP space. The only direct evidence offered, Fig. 4(a-b), is the ranking gap g = d_wrong − d_correct (0.1648 with L_rank versus 0.0020 without) and the margin-satisfaction rate; this establishes action-sensitivity but not predictive accuracy, because d_correct itself is never reported. The achieved values of L_future and L_delta (Eq. 5), the cosine distance between predicted and ground-truth futures, the relative residual norm |ΔŜ|/|S_t|, and the Stage-2 L_align value (Eq. 10) are all absent. Please report these at the end of Stage-1 and Stage-2 training, alongside baselines of the identity predictor Ŝ_{t+h} = S_t and an action-agnostic residual. Without such measurements, the alternative explanation—that the residual W_φ degenerates into a high-capacity action-conditioned feature vector that is a good L_align regression target but carries little genuine future signal—is not excluded, and the attribution of the remaining +1.96 points over π0.5+WM (§4.7.1) to future prediction per se is not established.","section":"§3.2, Eqs. (4)-(5); §4.7.2","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"Even taking the Stage-1 world model as accurate, the paper does not isolate whether the Stage-2 gains come from the information content of the predicted future or from the added pathway itself. WAVE-0 differs from π0.5+WM (Table 1b) not only in conditioning but also in the resampler and conditioning tokens, the latent-action query/planner branch, the alignment loss, and the L_world term evaluated on planner-generated plans. The Fig. 4(d) ablation shows that L_world and L_align each contribute, but it lacks a matched-capacity control, such as conditioning the expert on (i) the current representation with the same token budget or (ii) predicted futures generated from mismatched actions; such a control would test whether future-specific information, rather than extra capacity or the alignment-supervised planning signal, drives the remaining gain. Relatedly, the achieved value of L_align (Eq. 10) is never reported, so there is no coverage measure for whether the planner's latent plans stay inside the frozen world model's training distribution. Finally, because the query tokens pass through the shared PaliGemma stack, the claim that the planning branch guides the expert 'only through its predicted visual consequences' (end of §3.2) would benefit from an inference-time ablation that removes or corrupts the predicted futures.","section":"§4.7.1; Fig. 4(d)","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The real-world headline rests on 20 trials per task with no confidence intervals or significance tests (§4.5, Table 3). For USB Storage the comparison is 9/20 for π0.5 versus 13/20 for WAVE-0, a difference consistent with sampling noise; Book Organization (8/20 versus 17/20) and Keyboard Pushback (6/20 versus 14/20) are more supportive, but the abstract states the averaged claim (38.3% to 73.3%) as a headline result. Beyond noting 'diverse initial configurations,' the section gives no per-trial protocol (sampling of initial configurations, evaluation ordering, operator blinding), and the real-robot dataset is not released. The same absence of variance reporting applies to Tables 1 and 2: all entries are single runs without seeds or intervals, and several headline differences (e.g., LIBERO-Plus +3.5 overall; its −3.7 on the Noise category) are small enough that sampling variability is material. Please provide confidence intervals or standard deviations (at least three seeds where feasible), exact trial counts, and an explicit protocol for the real-world evaluation.","section":"§4.5, Table 3; Tables 1-2","authors_acknowledge":null,"authors_acknowledge_p":null}],"minor_comments":[{"comment":"The abstract calls the module a '257M world model,' but §4.1 says the 257M figure covers the world model, action encoder, and state encoder together; please report the world model's standalone parameter count.","section":"Abstract; §4.1","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"It is not stated which baseline numbers were produced in-house under identical data, initialization, and evaluation settings and which were quoted from the cited papers; the appendix gives training details only for WAVE-0, π0.5, LaWAM, and ImageWAM. Please clarify this for each baseline.","section":"§4.2-4.3; Tables 1-2","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The attention analysis is qualitative and based on two tasks; either quantify the attention concentration over held-out episodes or soften the claim that WAVE-0 exhibits 'more localized attention.'","section":"§4.6; Fig. 3(b)","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The long-horizon attribution cites two per-task gains ('Stack Blocks Three +23.0', 'Stack Bowls Three +33.0') selected from 50 tasks; please report the distribution of per-task gains and, if the horizon claim is retained, test it across all tasks rather than by example.","section":"§4.3; Table 13","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The paper has no limitations or broader-impact discussion; given the gap between the mechanism claim and the reported measurements, a short discussion of what the predicted futures do and do not certify would help calibrate readers' interpretation of the results.","section":"§5","authors_acknowledge":null,"authors_acknowledge_p":null}],"recommendation":"major_revision"},"author_rebuttal":null,"desk_editor":{"model":"deepseek/deepseek-v4-flash","letter":"Two things to know. The ablation ladder is genuinely good: pi0.5 at 80.38, pi0.5+WM auxiliary at 89.51, WAVE-0 at 91.47 on RoboTwin 2.0, plus the Stage-2 loss ablation in Fig 4d that matches the prose. And the paper's load-bearing mechanism — that predicted future representations carry real action-consequence signal into the action expert — is asserted, not measured. I mostly agree with the reader's conditional verdict; the stress-test note lands but needs proportion, since it names an unexcluded alternative rather than an established flaw.\n\nWhat's new: training the latent world model directly on ground-truth robot actions rather than latent actions inferred from visual transitions, the counterfactual ranking objective, and conditioning the action expert on predicted futures at both train and inference. The idea space is populated — the related work honestly shows VLA-JEPA, JEPA-WAM, LaWAM, ImageWAM, and Fast-WAM already doing latent future prediction — but this specific recipe, and especially the ranking loss, is genuinely theirs. The arithmetic is consistent (150k steps at batch 512 gives 76.8M examples), the efficiency comparison with LaWAM is fairly drawn, and there is no circularity: all objectives terminate in observed future frames from a frozen SigLIP encoder and in ground-truth actions. The Stage-2 alignment target is the paper's own Stage-1 action encoder, which is itself grounded in observed futures; that is a legitimate training target, not a loop.\n\nSoft spots, in proportion. The ranking gap (0.1648 vs 0.0020) proves action-sensitivity, not prediction accuracy. The paper never reports achieved L_future, L_delta, or cosine distances between predicted and ground-truth futures, so attributing the long-horizon gains to anticipating consequences (Section 4.3) is a causal interpretation, not a demonstrated fact. The stress-test scenario — the residual degenerating into a high-capacity action-conditioned feature the expert uses as extra capacity — cannot be ruled out from what is reported. That said, the pi0.5+WM ablation is the right control and it weakens the pure-capacity story: future-conditioning adds roughly two points on RoboTwin, and the ranking loss adds 2.6–3.0 points on LIBERO-Plus/Pro. The mechanism is partially supported, just not pinned down. Missing confidence intervals and 20 real-world trials per task are real but standard for this field, and the 38.3-to-73.3 gain is larger than the noise you would expect at N=20. The paper has no limitations section; the only caveat it volunteers concerns latency comparisons. A referee should ask for one.\n\nBaseline-number provenance is not a soft spot in my view: simulation numbers come from published, cited work, and the real-world baselines were trained in-house with full appendix detail.\n\nWho this is for: anyone working on world models for robot policies, latent dynamics, or conditioning mechanisms inside VLAs. It deserves a serious referee; one revision asking for prediction-fidelity numbers, error bars, and a limitations paragraph would tighten it substantially. Send it to review.","headline":"Well-executed, honestly framed latent-world-model work with a strong ablation ladder; the causal story ('anticipating consequences') outruns the evidence because absolute future-prediction fidelity is never reported.","tokens_in":21920,"tokens_out":6983,"duration_ms":46684},"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek/deepseek-v4-flash","headline":"WAVE-0 claims that a 257M latent world model trained on ground-truth robot actions can give a VLA policy explicit foresight, lifting real-world success from 38.3% to 73.3% and RoboTwin 2.0 from 80.38% to 91.47%.","keywords":["vision-language-action policy","latent world model","future representation prediction","robot manipulation","world action model","counterfactual ranking","latent action planning","flow matching"],"falsifier":"Measure the achieved future-prediction error $\\mathcal{L}_{\\text{future}}$ and $\\mathcal{L}_{\\text{delta}}$ on held-out trajectories from Stage 1, which the paper does not report, and compare them with a no-action baseline that predicts $\\hat{S}_{t+h} = S_t$ plus the state-only term. If the predicted future under the demonstrated action is not substantially closer to the true future SigLIP representation than the baseline prediction, or if the large counterfactual margin (0.1648) collapses outside the 2,048 evaluation samples used for Figure 4, then the predicted consequences carry little action signal and the conditioning bottleneck is not doing the world-modeling work attributed to it.","tokens_in":22229,"feed_emoji":"🤖","tokens_out":14836,"duration_ms":78680,"temperature":0.7,"pith_summary":"WAVE-0 asks whether an existing vision-language-action (VLA) robot policy can be given the ability to anticipate the consequences of its actions by a small latent world model, without video generation or large-scale pretraining. The paper answers yes: a 257M-parameter model trained directly on ground-truth robot actions translates VLM-generated latent action plans into predicted future representations, and those predictions condition the action expert during both training and inference. The claim matters because it offers an inexpensive route from policies that react to observations to policies that plan with imagined futures. On the paper's numbers the recipe lifts the base $\\pi_{0.5}$ policy from 80.38% to 91.47% on RoboTwin 2.0, from 96.9% to 98.7% on LIBERO, and from 38.3% to 73.3% on three real-world tasks.","feed_headline":"No-video world model lifts robot success from 38% to 73%","feed_subtitle":"A 257M latent model predicts action consequences and feeds them to the action expert at train and inference time.","key_machinery":"The load-bearing object is the latent world model with its additive residual dynamics in the frozen SigLIP embedding space: $\\hat{S}_{t+h} = S_t + W_\\phi(S_t, z_s, Z_a)$, where $S_t$ is the spatially pooled current visual representation, $z_s$ the encoded proprioceptive state, and $Z_a$ the action encoding (Eq. 4). A counterfactual ranking loss compares the predicted future under the demonstrated action sequence with the prediction under a mismatched sequence, forcing the model to be action-dependent; without it, the action-sensitivity gap on the paper's measurement drops from 0.1648 to 0.0020 and the margin satisfaction rate from 89.5% to 0%. The second critical piece is the future-conditioning bottleneck: a resampler pools the predicted future into conditioning tokens prepended to the noisy action tokens, while the latent-plan tokens are withheld from the expert, so anticipated consequences rather than plans must carry the planning signal. The two-stage objective (Eqs. 10–12) couples the frozen world model to the PaliGemma VLM planner through alignment and world losses, so the planner learns to emit plans whose predicted futures are both aligned with demonstrated action encodings and predictive of the observed future.","core_discovery":"The paper's central claim is that a VLA policy should be told what its actions will do, not just what to do next: the VLM planning branch writes latent action plans, a lightweight world model converts them into predicted future visual representations in a frozen SigLIP embedding space, and a resampler compresses those predictions into conditioning tokens the action expert must attend to (Eq. 3). Because the latent plan tokens are withheld from the expert's context, the planning branch can influence action generation only through the predicted consequences — an explicit prediction bottleneck that the paper argues distinguishes it from world models using future prediction merely as auxiliary supervision. The world model is trained in two stages: first, on ground-truth action sequences with future-representation, latent-delta, and counterfactual ranking losses (Eqs. 5–8); then, frozen, it supervises a VLM planner that must learn to emit plans aligned with the learned action space (Eqs. 9–12). The paper concludes that grounding latent dynamics in executable actions rather than video generation or inferred latent actions is what produces the gains, reporting 91.47% on RoboTwin 2.0, 98.7% on LIBERO, 88.0% on LIBERO-Plus, 61.4% on LIBERO-Pro, and 75.0% on LIBERO-Para, with the largest improvements on sequential and out-of-distribution manipulations.","pith_inferences":["Because the paper reports ranking margins but not achieved $\\mathcal{L}_{\\text{future}}$ values, a decisive follow-up would condition the expert on uninformative future tokens (a learned constant or the current representation), matched for parameter count, to test whether gains come from genuine action-consequence signal or from extra capacity.","The counterfactual ranking objective could serve as a general diagnostic for how action-sensitive any latent representation is, independent of this architecture: a large ranking margin in a frozen embedding space indicates the representation actually tracks action effects.","If the additive-residual assumption holds in SigLIP space because of its contrastive training geometry, the same two-stage recipe may transfer to other frozen contrastive encoders and VLA bases; swapping the vision encoder and checking whether the ranking margin survives would test this.","The large robustness gains under robot-state and camera perturbations suggest the future-conditioning branch may act partly as an implicit state estimator; a natural extension is closed-loop replanning that compares predicted futures against incoming observations and corrects when they diverge."],"forward_implications":["Existing VLA policies can gain world-model-style reasoning by adding a roughly 257M-parameter latent dynamics module, with no video generation and no large-scale world-model pretraining.","The largest wins come on sequential manipulation: +5.2 points on the LIBERO Long suite, +23.0 on RoboTwin 2.0's Stack Blocks Three, and +33.0 on Stack Bowls Three.","Latent future conditioning improves robustness to distribution shift, topping all compared methods on LIBERO-Plus (88.0%) and LIBERO-Pro (61.4%), with the biggest gains under robot initial-state changes (+14.0) and camera viewpoint shifts (+6.5).","Training is far cheaper than the latent video-pretrained alternative LaWAM: 11.09M examples per real-world task versus 115.20M, with higher success rates.","The planning branch's influence on actions runs through an explicit prediction bottleneck during both training and inference, so the reported gains are attributable to conditioning on predicted consequences rather than on the latent plans themselves."],"supporting_citations":[{"why":"The pretrained $\\pi_{0.5}$ policy WAVE-0 builds on; it supplies the frozen SigLIP encoder, the PaliGemma VLM planner, and the flow-matching action expert.","marker":"[15]"},{"why":"SigLIP, the frozen visual embedding space in which the additive residual dynamics are learned and where the core additivity assumption must hold.","marker":"[41]"},{"why":"LaWAM, the latent world action model WAVE-0 is contrasted with; its inverse-dynamics latent actions and video-scale pretraining motivate grounding dynamics in ground-truth actions.","marker":"[11]"},{"why":"PaliGemma, the VLM backbone that is fine-tuned to emit latent action plans aligned with the learned action-representation space.","marker":"[3]"},{"why":"LIBERO, the benchmark whose four task suites yield the 98.7% average success rate.","marker":"[25]"},{"why":"RoboTwin 2.0, the 50-task bimanual benchmark where WAVE-0's largest gains over $\\pi_{0.5}$ appear.","marker":"[12]"},{"why":"VLA-JEPA, a latent world model that uses future prediction to shape representations without conditioning action generation on predicted futures.","marker":"[31]"},{"why":"JEPA-WAM, the joint-embedding world model that likewise leaves prediction and action selection loosely connected, motivating WAVE-0's explicit bottleneck.","marker":"[24]"},{"why":"LIBERO-Plus, the seven-category perturbation benchmark where WAVE-0 reports 88.0%, the best among compared methods.","marker":"[13]"},{"why":"LIBERO-Pro, the object, spatial, and language generalization benchmark where WAVE-0 reports 61.4%.","marker":"[45]"}],"fun_headline_variants":["No-video latent world model lifts robot success to 73%","WAVE-0: Latent futures guide robot actions, no video needed","Latent future prediction boosts robot success from 38% to 73%","257M latent world model boosts robot success without video pretraining","WAVE-0's latent world model predicts actions, raising success to 73%"],"cache_read_input_tokens":0,"weakest_assumption_plain":"The pipeline assumes additive residual dynamics in the fixed SigLIP embedding space — the predicted future is the current representation plus an action-conditioned residual (Eq. 4) — and it assumes the VLM planner can reach the action encoder's target space from observation and instruction alone; if either fails, the predicted futures that condition the expert carry little genuine action-consequence signal and the gains could come from added capacity and auxiliary losses.","fun_headline_variants_meta":{"raw":{"variants":["No-video latent world model lifts robot success to 73%","WAVE-0: Latent futures guide robot actions, no video needed","Latent future prediction boosts robot success from 38% to 73%","257M latent world model boosts robot success without video pretraining","WAVE-0's latent world model predicts actions, raising success to 73%"]},"model":"deepseek/deepseek-v4-flash","effort":"low","cost_usd":0.000985,"raw_usage":{"cost":0.00098532,"is_byok":false,"cost_details":{"upstream_inference_cost":0.00098532,"upstream_inference_prompt_cost":0.00015904,"upstream_inference_completions_cost":0.00082628},"total_tokens":4087,"prompt_tokens":1136,"completion_tokens":2951,"prompt_tokens_details":{"audio_tokens":0,"video_tokens":0,"cached_tokens":0,"cache_write_tokens":0},"completion_tokens_details":{"audio_tokens":0,"image_tokens":0,"reasoning_tokens":2866}},"tokens_in":1136,"tokens_out":2951,"cost_source":"openrouter_usage","duration_ms":15194,"image_pages":[],"temperature":1.0,"billed_attempts":1,"reasoning_tokens":2866,"transport_provider":"openrouter","cache_read_input_tokens":0,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-10-14T09:53:32.146772+00:00","model_set":{"reader":"deepseek/deepseek-v4-flash"},"pipeline_version":"pith-pipeline@v1.4.1-daily-deepseek-or-medium","input_kind":null,"input_chars":null,"input_total_chars":null,"passes_missing":null,"published_standing":null,"published_confidence":null,"objection_audit":null,"run_record":null,"basis":{"kind":"not_recorded","source":null,"chars_given":null,"chars_total":null},"receipt":null,"falsifier":"Measure the achieved future-prediction error $\\mathcal{L}_{\\text{future}}$ and $\\mathcal{L}_{\\text{delta}}$ on held-out trajectories from Stage 1, which the paper does not report, and compare them with a no-action baseline that predicts $\\hat{S}_{t+h} = S_t$ plus the state-only term. If the predicted future under the demonstrated action is not substantially closer to the true future SigLIP representation than the baseline prediction, or if the large counterfactual margin (0.1648) collapses outside the 2,048 evaluation samples used for Figure 4, then the predicted consequences carry little action signal and the conditioning bottleneck is not doing the world-modeling work attributed to it.","supporting_citations":[{"cited_title":"$\\pi_{0.5}$: a Vision-Language-Action Model with Open-World Generalization","cited_arxiv_id":"2504.16054","evidence_quote":"The pretrained $\\pi_{0.5}$ policy WAVE-0 builds on; it supplies the frozen SigLIP encoder, the PaliGemma VLM planner, and the flow-matching action expert."},{"cited_title":"Sigmoid loss for language image pre-training","cited_arxiv_id":null,"evidence_quote":"SigLIP, the frozen visual embedding space in which the additive residual dynamics are learned and where the core additivity assumption must hold."},{"cited_title":"LaWAM: Latent World Action Models for Efficient Dynamics-Aware Robot Policies","cited_arxiv_id":"2606.15768","evidence_quote":"LaWAM, the latent world action model WAVE-0 is contrasted with; its inverse-dynamics latent actions and video-scale pretraining motivate grounding dynamics in ground-truth actions."},{"cited_title":"PaliGemma: A versatile 3B VLM for transfer","cited_arxiv_id":"2407.07726","evidence_quote":"PaliGemma, the VLM backbone that is fine-tuned to emit latent action plans aligned with the learned action-representation space."},{"cited_title":"Libero: Benchmarking knowledge transfer for lifelong robot learning.Advances in Neural Information Processing Systems, 36:44776–44791, 2023","cited_arxiv_id":null,"evidence_quote":"LIBERO, the benchmark whose four task suites yield the 98.7% average success rate."},{"cited_title":"RoboTwin 2.0: A Scalable Data Generator and Benchmark with Strong Domain Randomization for Robust Bimanual Robotic Manipulation","cited_arxiv_id":"2506.18088","evidence_quote":"RoboTwin 2.0, the 50-task bimanual benchmark where WAVE-0's largest gains over $\\pi_{0.5}$ appear."},{"cited_title":"Vla-jepa: Enhancing vision-language-action model with latent world model","cited_arxiv_id":null,"evidence_quote":"VLA-JEPA, a latent world model that uses future prediction to shape representations without conditioning action generation on predicted futures."},{"cited_title":"JEPA-WAM: Learning Vision-Language-Action Policies with Joint-Embedding World Modeling","cited_arxiv_id":"2608.09381","evidence_quote":"JEPA-WAM, the joint-embedding world model that likewise leaves prediction and action selection loosely connected, motivating WAVE-0's explicit bottleneck."},{"cited_title":"LIBERO-Plus: In-depth Robustness Analysis of Vision-Language-Action Models","cited_arxiv_id":"2510.13626","evidence_quote":"LIBERO-Plus, the seven-category perturbation benchmark where WAVE-0 reports 88.0%, the best among compared methods."},{"cited_title":"LIBERO-PRO: Towards Robust and Fair Evaluation of Vision-Language-Action Models Beyond Memorization","cited_arxiv_id":"2510.03827","evidence_quote":"LIBERO-Pro, the object, spatial, and language generalization benchmark where WAVE-0 reports 61.4%."}],"lane_accuracy":{"status":"not yet measured","lane":{"model":"deepseek/deepseek-v4-flash","input_kind":null},"url":"https://pith.science/accuracy"},"review_version":1}