{"id":"93b45708-2f56-4b7a-84ea-3f112833a856","arxiv_id":"2608.08558","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Vid2WAM trains a compact robot world-action model by distilling video-generation priors and inverse-dynamics pseudo-actions, improving novel-task success without target-task expert demonstrations.","lead":"This paper distills a large video-generation model into a compact robot policy so robots can learn new tasks from generated future frames and predicted actions instead of human demonstrations. It reports consistent gains on simulated and real bimanual manipulation benchmarks while keeping inference fast.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Novel-task gains rest on the video teacher's pretraining corpus excluding held-out tasks, but Table S5 audits only fine-tuning exposure, so the generalization claim is not yet verified.","rationale":"The reader identified the same load-bearing weakness, and I agree it is the most severe because the entire novel-task claim depends on the teacher's ability to synthesize correct futures for tasks never seen anywhere in its training. The paper's own audit (Table S5) covers only fine-tuning exposure and explicitly distinguishes that from using a trajectory's first frame to seed a rollout; it does not address the LVP checkpoint's pretraining corpus, which is described as robot manipulation data. Independent support exists: the ablations show future-latent supervision and source-aware adapter contributions, the Teacher Policy comparison rules out test-time teacher inference as the source of gains, and the multi-split sensitivity analyses show the effect is not tied to one task selection. However, all of these experiments use the same potentially contaminated teacher, so they cannot resolve the leakage question. Error bars and missing code are secondary concerns about precision and reproducibility; they do not bear on whether the novel-regime numbers mean what the paper claims. The correct status remains CONDITIONAL pending verification that the teacher's pretraining corpus excludes the held-out tasks, or a leakage-free re-run of the novel regime.","tokens_in":22816,"tokens_out":4093,"duration_ms":46702,"concrete_test":"Obtain or reconstruct the LVP pretraining data manifest and check the 15 held-out RoboTwin tasks and the LIBERO/LIBERO-Plus novel-task splits for overlapping clips. If the corpus is not accessible, re-run the RoboTwin novel regime with a teacher whose pretraining data is verifiably disjoint from the evaluation tasks (e.g., a video model trained only on non-robot web video, fine-tuned identically) and compare the novel-subset gain over Fast-WAM. As an intermediate check, compute perceptual retrieval similarity between teacher-generated rollouts for held-out tasks and the nearest LVP pretraining trajectories; high similarity concentrated on successful novel tasks would indicate memorized replay rather than generalization.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that Vid2WAM's novel-task improvements (RoboTwin novel subset 54.7/55.3 vs. Fast-WAM 45.0/42.8; real-world novel tasks) come from distilling a video teacher's generalizable world knowledge. This requires that teacher rollouts for held-out tasks are genuine synthesis, not replay of memorized trajectories. The paper's data-exposure audit (Table S5) only records whether the teacher and IDM were fine-tuned on held-out-task trajectories, and marks this as 'No'. But the teacher is initialized from LVP (Chen et al. 2025a), a Wan2.1-14B checkpoint explicitly described as pretrained on robot manipulation data (Implementation Details). If RoboTwin 2.0 or LIBERO task videos appear in LVP's pretraining corpus, then the held-out tasks are not novel to the teacher: its future predictions, and the IDM pseudo-actions distilled from them, can encode memorized task solutions. This would not merely weaken the numbers; it would invalidate the interpretation of the novel-regime experiments as evidence for generalization, because Vid2WAM would have hidden access to target-task visual and action information that baselines (which do not query this teacher) do not. The mechanism evidence in the ablations is real, but all runs share the same teacher, so it cannot distinguish distillation of a generalizable prior from distillation of memorized content.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper presents Vid2WAM, an offline distillation framework that transfers the predictive priors of a large video diffusion model into a compact World Action Model (WAM) student. Given an observation and language instruction, the video teacher generates future rollouts, which are decoded and re-encoded into student latent space to supervise the future-prediction branch, while an inverse dynamics model recovers pseudo-actions to supervise action learning. Source-aware residual adapters separate real and pseudo domains during training, and at inference only the compact student is deployed. Experiments on RoboTwin 2.0, LIBERO, LIBERO-Plus, and a real dual-arm platform report improvements over Fast-WAM, Motus, and π0.5 in low-data and novel-task regimes, with particular gains on the RoboTwin novel subset (54.7–55.3% vs. 45.0–42.8%).","tokens_in":23146,"tokens_out":7508,"duration_ms":74494,"significance":"The paper is a carefully executed empirical system contribution. Its main idea—decoupling future supervision for WAMs from target-task expert trajectories by distilling a pretrained video teacher into a compact student—is timely and practically attractive, since it preserves low-latency deployment while borrowing broad visual priors. The paper includes a thorough set of ablations (dual action DiT, no adapter, pseudo-action only, future-latent only, teacher policy), sensitivity analyses over different novel-task splits, and real-robot validation, all of which strengthen the causal interpretation of the components. If the results hold, the framework demonstrates a scalable route for injecting video-foundation-model knowledge into efficient robot policies. However, the headline novel-task generalization claim rests on the unverified assumption that the teacher's large-scale pretraining corpus does not already contain the held-out evaluation tasks; the paper audits only fine-tuning exposure, not pretraining exposure.","major_comments":[{"comment":"The novel-task claims (e.g., RoboTwin novel subset 54.7–55.3% vs. Fast-WAM 45.0–42.8% in Table 1) are interpreted as evidence that the video teacher provides generalizable world knowledge. This interpretation requires that the teacher's initialization, the LVP checkpoint (Chen et al. 2025a), which is described as pretrained on robot manipulation data, does not already contain the held-out RoboTwin 2.0 or LIBERO tasks. Table S5 only audits whether the teacher and IDM were fine-tuned on held-out-task trajectories ('No'), which is not the same as absence from the pretraining corpus. If any held-out task appears in LVP's pretraining data, the teacher's future predictions (and hence the IDM pseudo-actions distilled from them) could be near-memorized solutions, and the novel-regime comparisons would no longer measure generalization. Please provide verifiable evidence about the LVP pretraining corpus, or add an experiment with a teacher checkpoint known to be task-exclusive, or at minimum (c) explicitly re-frame the novel-regime results as conditional on this assumption and discuss the contamination risk.","section":"Implementation Details / Table S5"},{"comment":"In the novel-task regime, baselines are directly transferred from the remaining tasks with no exposure to target-task data, whereas Vid2WAM is allowed to use target-task instructions and initial observations to generate pseudo supervision. This is an asymmetric comparison: the observed gains could stem in part from Vid2WAM having seen the target-task language instruction and initial visual context during training, rather than from distillation of a generalizable prior. The headline wording 'improves novel-task generalization' should make this information asymmetry explicit (e.g., 'zero-shot transfer vs. task-conditioned offline distillation'). A useful control would be a baseline that is also given the target-task instructions and initial observations (without the video teacher) to isolate the teacher's contribution.","section":"Experimental Setup / Evaluation Regimes"},{"comment":"All success rates are point estimates from a single training run, with no confidence intervals, seed variance, or binomial intervals. Several headline improvements are modest (e.g., LIBERO novel average 78.3 vs. 76.7 in Table 3; LIBERO-Plus overall novel 48.1 vs. 45.8 in Table 2), and the real-world table reports only 20 trials per task, so differences of 10–15 points could be within sampling noise. Please report at least three seeds (mean ± std) for the simulation benchmarks, or binomial confidence intervals for task-level rates, and state explicitly whether the 20 real-world trials per task are the entire evaluation. This is essential for assessing the reliability of the central empirical claims.","section":"Tables 1–5 and S10–S15"}],"minor_comments":[{"comment":"The manuscript does not mention code or data release; please add a statement on availability to support reproducibility.","section":"General"},{"comment":"The caption says 'Fast-WAM (top) and Vid2WAM (bottom)', but the sub-figures are not labeled. Please label each row inside the figure for clarity.","section":"Figure 3"},{"comment":"The adapter residual scale α is set to 1.0 in Table S6 but its effect is not ablated; please clarify whether α was tuned and, if not, state why a fixed value was used.","section":"Eq. (3) / Table S6"},{"comment":"The 'Teacher exposure' and 'IDM exposure' columns use '/' for low-data regimes; consider using 'N/A' to avoid implying a binary value.","section":"Table S5"}],"recommendation":"major_revision","confidential_remarks":"The paper is a solid systems contribution with strong ablation coverage, but the reader's concern about pretraining exposure is legitimate and load-bearing for the novel-task claims. The authors should be pushed to either audit the teacher's pretraining corpus, provide a task-exclusive teacher control, or clearly temper the generalization claim. The lack of seed variance further weakens the headline numbers. I would not reject outright, as the core idea is promising and the evidence is extensive, but the revision needs to address these points."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a genuinely new transfer recipe, not a repackaging. The dual channel—teacher future latents directly supervising the student's video branch plus IDM pseudo-actions through source-aware residual adapters—is a real departure from DreamGen-style action-only pseudo-labeling. The paper does the distillation job carefully: teacher and IDM frozen, student-only inference, zero-init adapters, and ablations that separate future-latent, pseudo-action, and adapter contributions. The gains are consistent across RoboTwin, LIBERO, LIBERO-Plus, and a real bimanual platform, and the novel-task numbers are believable as a directional result. Appendix B's sensitivity to novel-split selection is a good-practice addition.\n\nThe main soft spot is the one the stress-test note flags. The video teacher is initialized from LVP, a Wan2.1-14B checkpoint pretrained on robot manipulation data. Table S5 audits whether the teacher or IDM was fine-tuned on held-out-task trajectories, which is fine as far as fine-tuning goes, but it does not tell us whether the pretraining corpus contained those tasks. RoboTwin 2.0 and LIBERO are standard and could plausibly be in a robot-manipulation pretraining set. If they are, the 'novel' tasks are not novel to the teacher, and the 9.7–12.5 point gains over Fast-WAM partly reflect hidden access to target-task visual knowledge rather than transfer of a generalizable video prior. I don't think this is deliberate; it is a provenance reporting gap. But it is load-bearing, because the entire novel-regime story flows through the teacher. The authors need to either document LVP's data composition, rerun with a teacher whose pretraining exclusion is known, or show that held-out task rollouts require genuine composition rather than retrieval.\n\nSmaller, addressable issues: no confidence intervals or seed variance anywhere in Tables 1–5, so several headline gaps (e.g. LIBERO average 78.3 vs 76.7) are statistically thin. No code or data release is promised. And the four free parameters (λ_pseudo, β, α, rollout cutoff) get no sensitivity analysis. These are minor next to the leakage question.\n\nWho this is for: anyone working on WAMs, video-diffusion policy learning, or synthetic-data distillation for manipulation. It deserves a serious referee. My recommendation: send it to review, but with the leakage audit and variance reporting as conditions, not optional polish.","headline":"Vid2WAM is a genuinely new distillation recipe with consistent gains, but the novel-task claim rests on an unverified pretraining-leakage assumption the authors need to close.","tokens_in":23663,"tokens_out":2523,"would_cite":true,"duration_ms":27642,"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":"Vid2WAM claims that future supervision for World Action Models can come from generated video rollouts rather than target-task expert demonstrations, and that this improves novel-task success and data efficiency without adding…","keywords":["world action model","video diffusion model","knowledge distillation","inverse dynamics model","future prediction","novel-task generalization","data efficiency","robot manipulation"],"falsifier":"Search the teacher's pretraining data (the LVP / Wan2.1-14B corpus) for overlap with the 15 RoboTwin novel tasks, the 8 LIBERO novel tasks, and the three real-world novel tasks; if any held-out task is present, re-run the novel regime with a teacher pretrained only on seen-task data. The claim collapses if the gains concentrate in the overlapped tasks and vanish after retraining on a disjoint corpus.","tokens_in":22662,"feed_emoji":"🤖","tokens_out":7312,"duration_ms":64949,"temperature":0.7,"pith_summary":"The paper challenges the assumption that World Action Models must be trained on target-task expert demonstrations that provide both actions and future video. It proposes Vid2WAM, which uses a large video diffusion model to generate task-conditioned future rollouts from just an initial observation and a language instruction, then distills those rollouts into a compact WAM student through two channels: future-video latents and inverse-dynamics-model-recovered pseudo-actions. A source-aware residual adaptation mechanism lets the student absorb noisy synthetic actions without corrupting the shared action representation, and at deployment only the student runs. On RoboTwin 2.0 novel tasks, Vid2WAM raises success from Fast-WAM's 45.0%/42.8% to 54.7%/55.3%, and it is the best method on LIBERO in both low-data and novel regimes. If correct, this means robot policies can be adapted to new tasks using only observations and language, not costly target-task demonstrations.","feed_headline":"Video-prior distillation lifts novel robot task success to 55%","feed_subtitle":"A compact world action model learns new tasks from generated futures plus pseudo-actions, then drops the teacher at runtime.","key_machinery":"The central mechanism is dual-channel distillation from a frozen video teacher into a compact student, with the student's VAE re-encoding teacher-generated frames into the same latent space used for its future-prediction targets. Action supervision comes from an inverse dynamics model trained on embodiment-compatible, but not target-task, trajectories, and the two supervision sources are combined through source-aware residual adapters: zero-initialized bottleneck projections around a shared action backbone, one for real actions and one for pseudo-actions, with only the real adapter retained at inference. Flow matching supervises both future-video latents and action chunks. This machinery lets generated futures serve as direct world-model supervision rather than being compressed through an imperfect action bottleneck, while residual adapters contain the noise from synthetic actions.","core_discovery":"Vid2WAM establishes that the future supervision used to train a World Action Model need not be recorded target-task trajectories. An embodiment-adapted video diffusion teacher generates task-conditioned rollouts, and the teacher's predictive prior is transferred into a compact student WAM by supervising both the future-prediction branch, through re-encoded teacher latents, and the action branch, through IDM-recovered pseudo-actions. The student trains jointly on real demonstrations and teacher rollouts with source-aware residual adapters, so noisy pseudo-actions get source-specific corrections instead of corrupting a shared action backbone. After training, the video teacher and the inverse dynamics model are discarded, leaving only the compact student for deployment. The paper reports consistent gains over the strongest WAM baseline on novel-task and low-data regimes in simulation and on real bimanual tasks, with ablation evidence that future-latent supervision and pseudo-action supervision are complementary.","pith_inferences":["The same dual-channel distillation could be driven by other generative sources, such as VLMs, subgoal proposers, or action-conditioned video models, so the specific 14B video teacher may be replaceable with smaller or cheaper generators.","The 8-second rollout cutoff used for simulation pseudo-data suggests long-horizon hallucination is the current bottleneck, so improving teacher temporal consistency may yield further gains without changing the student architecture.","If confirmed, novel-task evaluation protocols should audit the teacher's pretraining corpus for held-out tasks, because otherwise gains attributed to generalization could be memorized replay.","Source-aware residual adaptation is a general recipe for mixing real labels with noisy synthetic labels, potentially applicable beyond robotics to any distillation setting with unreliable teacher outputs."],"forward_implications":["Robot policies can be adapted to novel tasks with no target-task expert trajectories: only an initial observation, a language instruction, and offline teacher rollouts are needed.","The video teacher and IDM add zero latency at deployment; the student's chunk-level inference time is comparable to Fast-WAM and far faster than online video-generation pipelines.","Data efficiency improves: with only 10 expert trajectories per task, Vid2WAM outperforms VLA and WAM baselines on RoboTwin, LIBERO, and LIBERO-Plus.","Future-latent supervision is not redundant with action supervision: the ablation shows Future Latent Only beats Pseudo Action Only, and the full dual-channel model is best.","Offline distillation beats an online teacher-plus-IDM controller, so the gains come from training-time transfer rather than from a stronger runtime policy.","The benefits persist across different held-out task splits, indicating they are not an artifact of one task selection."],"supporting_citations":[{"why":"It supplies Fast-WAM, the closest architectural baseline and the starting architecture for the student, and establishes that future imagination can be removed at inference.","marker":"(Yuan et al. 2026)"},{"why":"DreamGen supplies the closest prior pipeline of video-generated rollouts plus IDM pseudo-actions, which Vid2WAM extends by supervising future latents directly.","marker":"(Jang et al. 2025)"},{"why":"It supplies the Large Video Planner checkpoint used to initialize the video teacher's diffusion prior.","marker":"(Chen et al. 2025a)"},{"why":"It supplies the Wan2.1-14B architecture of the video teacher.","marker":"(Wan et al. 2025)"},{"why":"It supplies RoboTwin 2.0, the primary benchmark and simulator where the novel-task and low-data results are measured.","marker":"(Chen et al. 2025b)"},{"why":"It supplies LIBERO, the benchmark used for low-data and novel-regime comparisons.","marker":"(Liu et al. 2023)"},{"why":"It supplies LIBERO-Plus, the perturbation benchmark used to test robustness across seven factors.","marker":"(Fei et al. 2025)"},{"why":"It supplies the flow-matching objective used to supervise both action chunks and future-video latents.","marker":"(Lipman et al. 2022)"}],"fun_headline_variants":["Video diffusion priors distill into robot action models for new tasks","Generated futures and pseudo-actions train WAMs without expert demos","Compact student WAM learns novel skills from video teacher rollouts","Vid2WAM: distilling video priors into world action models beats demos"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The novel-task gains assume that the large video teacher's pretraining corpus does not already contain the held-out evaluation tasks, so the teacher's rollouts for those tasks are genuine generalization rather than memorized replay.","fun_headline_variants_meta":{"raw":{"variants":["Video diffusion priors distill into robot action models for new tasks","Generated futures and pseudo-actions train WAMs without expert demos","Compact student WAM learns novel skills from video teacher rollouts","Vid2WAM: distilling video priors into world action models beats demos"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000259,"raw_usage":{"total_tokens":1579,"prompt_tokens":933,"completion_tokens":646,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":549,"completion_tokens_details":{"reasoning_tokens":567}},"tokens_in":549,"tokens_out":646,"duration_ms":7313,"temperature":1.0,"reasoning_tokens":567,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T04:31:15.518952+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Search the teacher's pretraining data (the LVP / Wan2.1-14B corpus) for overlap with the 15 RoboTwin novel tasks, the 8 LIBERO novel tasks, and the three real-world novel tasks; if any held-out task is present, re-run the novel regime with a teacher pretrained only on seen-task data. The claim collapses if the gains concentrate in the overlapped tasks and vanish after retraining on a disjoint corpus.","supporting_citations":[],"review_version":1}