{"id":"d908c425-6331-411e-a795-44f55a56194b","arxiv_id":"2608.09381","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"JEPA-WAM couples latent transition prediction in a frozen V-JEPA space with action generation through a shared predictor, improving out-of-distribution manipulation success on LIBERO-Plus, RoboTwin 2.0, and a real bimanual robot.","lead":"This paper presents JEPA-WAM, a robot policy that learns to predict visual changes between current and future camera frames without generating images. The predictor that makes these predictions also produces the action decisions, giving the robot better performance under changed scenes and camera conditions.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The core mechanism claim lacks statistical support: the 2.2-point LIBERO-Plus gain over V-JEPA-only is within plausible seed noise and is reversed in four of seven perturbation categories.","rationale":"The reader's CONDITIONAL verdict is appropriate, but the single most load-bearing concern is not primarily the cross-paper baseline protocol. The paper's scientific contribution is the claim that transition supervision through a shared predictor directly shapes the action backbone. That claim is supported only by internally controlled ablations in Table 4, which show a small average gain (2.2 points) over the no-transition baseline and non-uniform per-category effects. No error bars or seed counts are given for any LIBERO-Plus number, and the per-category pattern — improvement on Robot and Background but regression on Language, Lighting, Noise, and Layout — suggests the effect may be specific rather than a general temporal-structure benefit. This is the load-bearing point because if the ablation gain is within noise, the method's core novelty vanishes even after all baseline-protocol questions are resolved. The reader's stated weakest assumption (baseline protocol equivalence) is real but affects only the SOTA headline, not the internal validity of the design. Thus I partially agree with the reader: we both flag missing variance, but I locate the decisive vulnerability in the paper's own controlled comparisons rather than the cross-paper comparisons. The proposed concrete test — multiple seeds with per-category CIs — would settle whether the mechanism claim is reproducible. If it fails, the paper's central contribution is unsupported; if it passes, the conditional acceptance is justified.","tokens_in":21105,"tokens_out":10998,"duration_ms":115340,"concrete_test":"Re-run the full LIBERO-Plus evaluation with at least 5 independent training seeds for JEPA-WAM and the three key ablations (V-JEPA-only, future-only target, joint current-future target), keeping all hyperparameters fixed; report per-category and average success rates with bootstrap 95% CIs. The central mechanism claim stands only if the joint-target variant beats V-JEPA-only by more than the seed noise, and only if the per-category regressions (Language, Lighting, Noise, Layout) are reproducible rather than noise.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central mechanistic claim — that joint current–future transition supervision through the shared predictor improves OOD generalization — rests on Table 4: JEPA-WAM 79.2% vs. the V-JEPA-only variant 77.0%. No confidence intervals, multiple seeds, or per-category episode counts are reported for any LIBERO-Plus result. The gain is not uniform: JEPA-WAM is worse than V-JEPA-only on Language (68.2 vs. 70.9), Lighting (93.3 vs. 96.7), Noise (83.6 vs. 88.3), and Layout (76.1 vs. 79.3), while being much better on Robot (+18.3) and Background (+10.5). This pattern is consistent with the paper's own §6 limitation that language-agnostic transition targets may be less expressive when the same observation leads to different task-conditional transitions, but the paper does not present it as a trade-off. If the 2.2-point average advantage is within run-to-run variance, the paper's central contribution reduces to 'V-JEPA features are useful', which is not the claimed contribution. The cross-paper baseline-protocol issue identified by the reader is secondary: even if every baseline were re-run under a common protocol, the method's own ablation evidence for the shared-predictor mechanism would remain unsupported without variance estimates.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes JEPA-WAM, a latent world action model that couples transition prediction with continuous action generation through a shared predictor. The method operates in a frozen V-JEPA 2.1 representation space, constructs a joint current–future target by stacking observations at times t and t+δ, and supervises patch-level cosine alignment between the predicted and target representations. The same predictor produces action-conditioning representations for a flow-matching action expert. The authors evaluate on LIBERO, LIBERO-Plus, RoboTwin 2.0, and a real bimanual platform, reporting that JEPA-WAM achieves 79.2% on LIBERO-Plus without robot-policy pretraining and that instantiation of the same transition objective in π0.5 improves it to 86.3%, the best overall result. Controlled ablations compare visual encoders, transition targets, spatial supervision, and the placement of transition supervision.","tokens_in":21383,"tokens_out":3721,"duration_ms":35785,"significance":"If the empirical claims hold, the paper makes a meaningful contribution: it shows that a latent world model can provide OOD generalization without video generation, that a frozen V-JEPA space can serve as both representation and target space, and that a shared predictor lets transition supervision directly shape the action-generating backbone. The design is clearly specified, the ablation set is more comprehensive than typical for this area, and the use of a frozen, externally pretrained target avoids the circularity of fitting a self-produced target. The appendices add useful probe analyses and per-rollout real-world records. The main weaknesses are statistical: the central mechanistic claims rest on point estimates without variance or multiple seeds, and the headline comparisons against prior methods rely on published numbers produced under unknown protocols. These issues are fixable within the scope of the manuscript.","major_comments":[{"comment":"The central mechanistic claim that joint current–future transition supervision through the shared predictor improves OOD generalization rests on the comparison between JEPA-WAM (79.2%) and V-JEPA only (77.0%) in Table 4. No confidence intervals, multiple seeds, or per-category episode counts are reported. The advantage is reversed in four of seven categories: Language (68.2 vs. 70.9), Lighting (93.3 vs. 96.7), Noise (83.6 vs. 88.3), and Layout (76.1 vs. 79.3). The paper's own §6 acknowledges that language-agnostic transition targets may be less expressive when the same observation leads to different task-conditional transitions, but the paper does not present this as a trade-off. Please report variance estimates and discuss the category-level pattern; as it stands, the 2.2-point average gain is within plausible run-to-run variance and does not support the mechanistic claim.","section":"§4.3, Table 4; §6"},{"comment":"The headline comparisons ('best result without robot-policy pretraining', 'best overall') rely on baseline success rates taken from prior papers, and the manuscript re-runs none of these baselines under its own protocol. This is especially problematic for Table 3 on RoboTwin 2.0, where methods may differ in task grouping, camera views, training data, and number of evaluation rollouts. Without common-protocol re-evaluation or explicit evidence that the protocols are equivalent (e.g., same episodes, same camera views, same rollout counts), the SOTA claims are not supported.","section":"§4.2, Tables 1–3"},{"comment":"The π0.5+JEPA Obj. variant appends 64 learnable future tokens to the VLM prefix, and the reported gains (84.5→86.3 on LIBERO-Plus; 75.4→84.6 Clean on RoboTwin 2.0) could in principle come from the added parameters rather than from the transition supervision. Please include a control with the same future tokens inserted but the transition loss removed or ablated, to isolate the effect of the supervision. This is needed to support the transfer claim in §3.4.","section":"§3.4, Tables 2 and 3"},{"comment":"The real-world conclusions are based on 10 rollouts per task and setting with normalized task-completion scores. While the per-rollout records in Appendix E are valuable, no uncertainty estimates or statistical tests are reported. The claim of 'strong generalization' would be better supported by bootstrap confidence intervals across rollouts or an explicit statement that the sample size is too small for statistical inference.","section":"§4.4, Table 15"}],"minor_comments":[{"comment":"The caption states that 'the area represents the model backbone's parameter count', but no legend or scale is provided; please clarify how the reader should interpret the area.","section":"Figure 1"},{"comment":"Please clarify whether the visual projector is included in the reported 0.5B parameter count for the main backbone, since the projector (1024→896→896) contributes a non-negligible number of parameters.","section":"Appendix A.1"},{"comment":"For consistency with the other tables, the final row should be labeled 'JEPA-WAM (Ours)' rather than just 'JEPA-WAM'.","section":"Table 4"},{"comment":"The description of the iREPA align. variant says 'per-view 3×3 convolutional transformation and spatial target normalization'; please specify whether the normalization is applied to the prediction, the target, or both.","section":"Appendix C.1"},{"comment":"The abstract states that a project page is available on GitHub, but no URL is provided in the manuscript.","section":"Abstract"},{"comment":"Please report the number of evaluation episodes used for each LIBERO-Plus average so that the reader can assess the reliability of the point estimates.","section":"§4.2"}],"recommendation":"major_revision","confidential_remarks":"The paper is a serious empirical contribution with a clear architecture and a well-designed ablation set. The central mechanism claim, however, needs variance estimates and a control for added parameters in the pretrained-VLA transfer. The cross-paper baseline comparison is a secondary but important concern. I would not reject; the issues are fixable within the manuscript's scope."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The short version: this is a solid, clearly written method paper with two genuinely new pieces—a shared predictor that couples latent transition prediction with action generation, and a joint current–future target in a frozen V-JEPA space. The ablations are well-designed and mostly support the design choices. The soft spot is that the headline LIBERO-Plus gain over the V-JEPA-only variant is 2.2 points, reported without confidence intervals, and it reverses in four of seven perturbation categories. That undercuts the central mechanism claim as stated, though it does not sink the paper.\n\nThe method itself is clean. The target is defined by a frozen, externally pretrained V-JEPA encoder with stop-gradient, so there is no circularity. The patch-level cosine loss preserves spatial structure, and the shared predictor lets transition supervision directly shape the backbone while action placeholders provide a dedicated readout. The comparison against future-only, endpoint-difference, iREPA-style, lower-layer, and full-hidden variants is exactly the right set of ablations to run, and the results give a coherent picture: joint encoding beats endpoint differencing, patch-level beats local mixing, and dedicated action placeholders beat full-hidden conditioning. The appendix probes with bootstrap confidence intervals are a useful addition.\n\nWhere it is soft: the LIBERO-Plus numbers are point estimates without variance. The 2.2-point average advantage over V-JEPA-only could easily be seed noise, and the per-category breakdown shows the method is worse on Language, Lighting, Noise, and Layout while being much better on Robot and Background. The paper's own §6 limitation—language-agnostic transition targets may be less expressive—predicts exactly this trade-off, but the text presents the average as an unqualified win. The SOTA claims in Tables 1–3 also rely on published baseline numbers without re-running under a common protocol; the reader flagged this correctly. That said, the RoboTwin and real-world results are more reassuring: the π0.5+JEPA Obj. improvements are large and consistent across tasks and settings, with per-rollout data in the appendix, so the method is not doing nothing.\n\nWho this is for: anyone working on latent world action models or VLA robustness. The shared-predictor design is worth citing. My recommendation: send it to peer review. The statistical weakness is fixable—add multiple seeds or confidence intervals for the key ablation, re-run at least the main baselines under a common protocol, and frame the per-category trade-off honestly. The core architecture is novel and the experiments are extensive enough to justify referee time.","headline":"A genuinely novel and well-ablated latent WAM architecture whose central LIBERO-Plus claim rests on a 2.2-point gain without variance; still worth sending to review.","tokens_in":21933,"tokens_out":2939,"would_cite":true,"duration_ms":28003,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"JEPA-WAM proposes that a frozen V-JEPA space plus a shared predictor for transition prediction and action generation yields robust out-of-distribution robot control, reaching 79.2% on LIBERO-Plus without robot-policy pretraining and…","keywords":["vision-language-action models","world action models","joint-embedding predictive architecture","latent world models","out-of-distribution generalization","robot manipulation","flow matching","V-JEPA"],"falsifier":"Run JEPA-WAM and its V-JEPA-only variant (transition prediction removed) with identical seeds, data, and rollouts on LIBERO-Plus; if the average success-rate gap between with and without transition supervision does not reproduce the reported 2.2-point difference (79.2 versus 77.0) with non-overlapping confidence intervals, the central claim that transition supervision drives the out-of-distribution gain would be undercut.","tokens_in":20907,"feed_emoji":"🤖","tokens_out":6896,"duration_ms":59391,"temperature":0.7,"pith_summary":"This paper sets out to establish that a robot policy can become more robust to visual and spatial changes by predicting, in latent space, how the scene will change between the current and a slightly later observation, without ever generating an image of the future. The proposed model, JEPA-WAM, builds on a frozen V-JEPA encoder and trains a shared predictor that both predicts a dense, per-image-region joint current-future target and produces representations used to generate continuous actions. The core claim is that this transition supervision directly shapes the same backbone that conditions action generation, so the policy learns task-shared visual temporal structure. On LIBERO-Plus the model reaches 79.2% average success without large-scale robot-policy pretraining, and the same supervision raises a pretrained pi0.5 policy to 86.3%, the best overall result, with similar gains on RoboTwin 2.0 and real-world bimanual tasks.","feed_headline":"World model scores 79.2% on shifted robot tasks without pretraining","feed_subtitle":"Patch-level transition prediction in a frozen V-JEPA space lifts shifted-task success to 79.2%, and to 86.3% on a pretrained policy.","key_machinery":"The load-bearing mechanism is the spatially structured joint current-future target $Y_{t,t+\\delta} = \\mathrm{Concat}_{v}\\,\\mathrm{sg}\\!\\left(E_J(\\mathrm{Stack}_{\\mathrm{time}}(O_t^v, O_{t+\\delta}^v))\\right)$, a frozen V-JEPA encoding of the two-frame stack that keeps the same patch grid as the current representation. A shared predictor $F_\\theta$ (a Qwen2.5-0.5B language model with LoRA) maps the projected current V-JEPA tokens, the task instruction, and dedicated action placeholders to hidden states; a lightweight prediction head maps the visual-token hidden states back into V-JEPA patch space, where they are supervised by mean patch-wise cosine distance to the joint target. The same forward pass yields dedicated action-conditioning tokens that feed a DiT flow-matching action expert. This coupling is what lets the transition loss update the action-generating backbone directly. The paper also transfers the same target to pretrained VLA policies by adding coarse future tokens, reshaping and upsampling their hidden states to match the target grid.","core_discovery":"The central discovery is that a transition target defined by stacking current and future observations and encoding them together with a frozen V-JEPA encoder, rather than encoding the future alone or subtracting endpoint encodings, carries more accessible temporal-relation information, and that supervising a shared Qwen-based predictor to predict it patch-wise improves the representations used for action conditioning. The paper demonstrates this through LIBERO-Plus ablations (joint target 79.2 versus future-only 77.3 versus endpoint-difference 70.9), frozen-probe experiments showing better temporal-gap decoding and residual trajectory prediction, and successful transfer to a pretrained VLA via auxiliary future tokens. On the paper's own terms, the claim is that latent transition prediction in the V-JEPA space, coupled with action generation through a shared predictor, yields strong generalization under visual and spatial shift while preserving in-distribution performance.","pith_inferences":["The paper's stated limitation suggests a testable extension: conditioning the transition target on language or other task context could close the gap when the same observation leads to different transitions under different instructions; a variant with instruction-conditioned targets should be compared on LIBERO-Plus.","Because patch-wise representation changes correlate only moderately with RGB change, the transition signal may encode object relations, contact, and configuration changes beyond low-level motion; an explicit study separating these factors could guide future target design.","The shared-predictor design implies that scaling the transition loss weight or using multiple temporal offsets might continue to improve out-of-distribution action conditioning, a prediction that could be checked by sweeping $\\lambda_{wm}$ and $\\delta$ on LIBERO-Plus.","The same joint-target formulation could be applied to other pretrained visual encoders or to action-chunking choices, potentially removing the need for a separate world-model module in existing VLA stacks."],"forward_implications":["Because the target encoder, prediction head, and transition loss are removed at deployment, the policy keeps the same inference cost as a non-world-model VLA, measured at 85 ms per inference on RoboTwin.","The same transition supervision transfers to pretrained VLAs as an auxiliary loss, improving pi0.5 from 84.5% to 86.3% on LIBERO-Plus and from 77.5% to 90.3% on real-world in-distribution tasks without altering its action pathway.","Patch-level transition supervision outperforms compressed or convolved targets, indicating that future latent world action models should preserve dense spatial structure rather than pool into global tokens.","The frozen V-JEPA representation space itself contributes to robustness: replacing DINOv2+SigLIP with V-JEPA raises LIBERO-Plus success from 73.2% to 77.0% even without transition prediction.","Joint current-future encoding beats both future-only targets and explicit endpoint differencing, suggesting that temporal-relation information, not just the future state, is the useful supervisory signal."],"supporting_citations":[{"why":"Supplies the frozen V-JEPA 2.1 representation space that defines both the current visual state and the joint current-future target.","marker":"Mur-Labadia et al. 2026"},{"why":"Provides the V-JEPA 2 self-supervised video model whose temporally consistent features the paper relies on for transition supervision.","marker":"Assran et al. 2025"},{"why":"Supplies the LIBERO-Plus benchmark and its out-of-distribution evaluation protocol used for the headline generalization results.","marker":"Fei et al. 2025"},{"why":"Supplies the LIBERO benchmark and demonstration suites used for in-distribution training and evaluation.","marker":"Liu et al. 2023"},{"why":"Provides the pi0.5 pretrained VLA policy that the paper uses as a baseline and as the instantiation for its transferred transition supervision.","marker":"Physical Intelligence et al. 2025"},{"why":"Provides the VLA-JEPA baseline with a separate JEPA world model, which the paper contrasts with its shared-predictor design.","marker":"Sun et al. 2026"},{"why":"Supplies the single-stage vision-language finetuning protocol used to initialize the visual projector and Qwen backbone.","marker":"Karamcheti et al. 2024"},{"why":"Supplies the DiT flow-matching action head design that JEPA-WAM adopts for action prediction.","marker":"StarVLA Community 2026"},{"why":"Supplies the RoboTwin 2.0 benchmark and its Clean/Random bimanual manipulation protocol used for generalization evaluation.","marker":"Chen et al. 2025"}],"fun_headline_variants":["Joint-embedding world model beats baselines on shifted robot tasks","JEPA-WAM: latent world model hits 79.2% on shifted tasks","Shared predictor in V-JEPA space boosts robot generalization","Patch-wise future prediction lifts robot task success"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The headline comparisons assume that the published baseline numbers were produced under training and evaluation protocols equivalent enough to the paper's own, including the same benchmark episodes, camera views, and rollout counts, since none of those baselines were re-run in the paper's codebase.","fun_headline_variants_meta":{"raw":{"variants":["Joint-embedding world model beats baselines on shifted robot tasks","JEPA-WAM: latent world model hits 79.2% on shifted tasks","Shared predictor in V-JEPA space boosts robot generalization","Patch-wise future prediction lifts robot task success"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000195,"raw_usage":{"total_tokens":1364,"prompt_tokens":957,"completion_tokens":407,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":573,"completion_tokens_details":{"reasoning_tokens":336}},"tokens_in":573,"tokens_out":407,"duration_ms":4856,"temperature":1.0,"reasoning_tokens":336,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T18:23:01.662740+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run JEPA-WAM and its V-JEPA-only variant (transition prediction removed) with identical seeds, data, and rollouts on LIBERO-Plus; if the average success-rate gap between with and without transition supervision does not reproduce the reported 2.2-point difference (79.2 versus 77.0) with non-overlapping confidence intervals, the central claim that transition supervision drives the out-of-distribution gain would be undercut.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the LIBERO-Plus benchmark and its out-of-distribution evaluation protocol used for the headline generalization results."}],"review_version":1}