{"id":"88f3c5ae-1235-4e97-b93f-42e8c725adb8","arxiv_id":"2608.11605","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"ForeWAM conditions a direct action-diffusion policy on hidden video-transformer key-value states over stochastic future slots, without decoding future video, achieving 96.7% LIBERO and 61.6% LIBERO-Plus success.","lead":"This paper introduces ForeWAM, a robot-control model that lets an action policy draw on learned future scene dynamics without generating future video frames at run time. The method reports high success on LIBERO benchmarks while cutting action-generation latency by up to 67% compared with a recent world-action model.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Deployment-time noise-filled future slots are asserted, not shown, to carry predictive dynamics; until Eq. 3's substrate is tested with matched interventions, 'foresight without seeing' is an unsupported mechanism.","rationale":"The reader and I locate the load-bearing risk in the deployment-time substrate rather than in the benchmark or latency figures. The paper is unusually honest about its limitations, repeatedly disclaiming causal attribution in Sec. 3.3 and Sec. 4.4, but the architecture's conceptual identity depends on Eq. 3: noise-filled future slots processed by a video prefill must expose action-relevant predictive dynamics. No intervention varies the substrate at inference, and no probe of the cached K/V states' content is provided; the paper's own statements concede that the aggregate comparison does not establish the causal contribution of either pathway. The proposed test settles this without retraining or new benchmarks, since only the deployment-time construction of the future slots changes. If performance is invariant to the substrate, the \"foresight\" claim reduces to current-observation conditioning plus training-time future supervision, and the central novelty is not established. If the substrate matters, the concern is answered. This is therefore a condition on acceptance rather than a rejection, so the reader's CONDITIONAL verdict stands unchanged.","tokens_in":12413,"tokens_out":5719,"duration_ms":64119,"concrete_test":"Run a matched deployment intervention on the exact Table 4 evaluation protocol (same 1,482 LIBERO-Plus rollouts and same trained ForeWAM): replace epsilon_F in Eq. 3 with (A) zero latents and (B) the current latent zcur repeated across future positions, holding all other inputs and attention masks fixed; compute per-task Wilson 95% confidence intervals against the 61.6% baseline. Also run (C) a mask intervention that removes future-slot keys and values from action cross-attention under the same protocol. If both A and B stay within the baseline confidence interval, the stochastic future substrate is not load-bearing and the foresight claim fails; if either drops materially, the substrate transfers learned predictive content.","verdict_should_be":"UNCHANGED","load_bearing_attack":"ForeWAM's central claim is that at deployment the Action DiT receives predictive dynamics through Future-KV (Sec. 3.1). Eq. 3 constructs the deployment substrate as concat(zcur, epsilon_F) with epsilon_F drawn fresh per query, and Eqs. 4-5 condition the policy on the resulting H_KV and D_theta. Nothing in the objective guarantees that K/V states computed from noise-filled future slots carry task-relevant future information. During training the video branch is supervised on ground-truth future latents z1:T (Eq. 8), but at inference those positions contain pure noise; the statement that the slots' \"usefulness is learned from the joint video-action objective rather than from future observations at deployment\" is an assertion. The only component evidence, Table 4, reports no error bars and explicitly states that the aggregate comparison \"does not by itself establish the causal contribution of either pathway,\" with the Base row evaluated under a different coverage profile. The Future-KV path could be acting as a learned, observation-conditioned context that adds capacity without encoding futures; if so, the claimed conceptual contribution, exposing predictive dynamics without generating futures, is not demonstrated.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces ForeWAM, a direct-policy World Action Model that aims to expose predictive visual dynamics to an action-generation transformer without generating or decoding future videos at inference. The proposed architecture performs a single Video DiT prefill over the current observation latent and stochastic future slots filled with noise, caches the resulting layer-wise key-value states, and reuses them as conditioning context during action denoising. In addition, a set of dynamics registers is supervised during training by a frozen latent-action teacher to encourage the hidden states to encode interaction-relevant transitions. The manuscript reports average success rates of 96.7% and 96.9% on the standard LIBERO benchmark, 61.6% on an 'observed subset' of LIBERO-Plus, and reduced action-generation latency (568 ms; 220 ms for an accelerated variant) relative to a reported Fast-WAM baseline, using roughly one-third of the baseline's policy parameters.","tokens_in":12664,"tokens_out":8042,"duration_ms":77609,"significance":"If the central mechanism is genuine, ForeWAM would be a valuable step toward making world-action models efficient at inference while retaining a notion of predictive dynamics. The paper is admirably explicit about several limitations, including the configuration-level nature of the component comparisons and the descriptive rather than causal status of cross-method benchmarks. However, the significance is conditional on demonstrating that the noise-substrate future slots actually carry task-relevant predictive information; as written, the evidence is consistent with alternative explanations. The benchmark numbers are useful but lack uncertainty quantification and matched comparisons, so the reported gains should be interpreted cautiously.","major_comments":[{"comment":"The central claim that Future-KV exposes predictive dynamics to the Action DiT at deployment is not supported by the presented evidence. At training time the video branch receives ground-truth future latents z1:T (Eq. 8), but at inference the future-slot positions contain fresh noise epsilon_F (Eq. 3). The paper asserts that the usefulness of this noise substrate is learned from the joint video-action objective, yet it provides no analysis of the resulting train/inference distribution shift and no matched intervention showing that the cached K/V states carry task-relevant future information. The component ablation in Table 4 explicitly states that the Base configuration was evaluated under a different coverage profile (10,027 evaluations vs. 1,482 for the other rows) and that the aggregate comparison does not establish the causal contribution of either pathway. Consequently, the 4.9-point gap between Base and Future-KV-only could be attributable to coverage, added capacity, or other confounds. The phrase 'foresight without seeing' and the claim of exposing predictive dynamics therefore overstate what the experiments demonstrate; at minimum, the paper should report a matched ablation with the same evaluation set and provide a diagnostic (e.g., probing the cached states for future information or comparing against a zero-filled future-slot baseline) to support the mechanistic interpretation.","section":"Sec. 3.1, Eqs. (3)-(5); Sec. 4.4, Table 4"},{"comment":"The LIBERO-Plus evaluation is performed on an unspecified 'observed subset,' and no details are given about which tasks, perturbation intensities, or number of rollouts are included. The headline comparison against Fast-WAM (61.6% vs. 51.5%) uses a number taken from another paper that may have been computed under different evaluation conditions; the authors themselves describe these cross-method differences as 'descriptive rather than coverage-matched causal estimates.' Without a precise definition of the subset and matched evaluation of the baseline under the same protocol, the reported 10.1-point improvement cannot be verified or given an error bar. Please specify the subset composition, per-category evaluation counts, and ideally re-run Fast-WAM under identical conditions.","section":"Sec. 4.1 and Table 2"},{"comment":"The success-rate tables report no error bars, confidence intervals, or significance tests. For a benchmark with 50 rollouts per task (LIBERO) and unknown per-category counts (LIBERO-Plus), the binomial standard error is substantial; for example, a 0.9-point overall difference on LIBERO corresponds to roughly 1.8 successes across a suite of 200 rollouts, which is within one standard deviation of sampling noise. This is especially concerning for per-category gains such as the +46.1-point camera-viewpoint difference, which may rest on a small number of tasks. The paper should report standard errors or confidence intervals for every success-rate cell, or at minimum per-cell rollout counts.","section":"Tables 1, 2, and 4"},{"comment":"The paper claims that dynamics registers 'encourage the implicit future states to capture interaction-induced transitions such as object motion, contact changes, and task progress,' but no evidence directly measures these semantic properties. The LA loss (Eq. 9) only encourages the mean-pooled register vector to match a frozen teacher embedding; the paper does not probe or decode the registers to verify that they encode transitions, nor does it isolate the contribution of LA supervision while holding Future-KV fixed (the closest ablation, 'LA supervision only', also omits Future-KV). The semantic claims about what the registers represent are therefore unsupported and should be rephrased as a training objective, not an observed property, or supplemented with a representational analysis.","section":"Sec. 3.2, Eq. (9), and the Abstract"},{"comment":"The inference-latency comparison against Fast-WAM does not state whether the 667 ms figure was measured in the same environment (single A800 GPU), with the same batch size, action horizon, and denoising-step count. If it is a reported number from another paper, the claimed 14.8% and 67.0% reductions are not controlled comparisons. Please clarify the measurement protocol and, if possible, measure the baseline on the same hardware to support the efficiency claims.","section":"Sec. 4.3, Table 3"}],"minor_comments":[{"comment":"Rows labeled 'Ours' and 'Ours-Flash' should use the model names 'ForeWAM' and 'ForeWAM-Flash' to avoid ambiguity and match the text.","section":"Tables 1 and 2"},{"comment":"The caption and the text in Sec. 1 use inconsistent labels: panel (c) is described as 'Direct-policy WAM' in the text but the caption reads 'Direct-policy WAMs'; also, the reference to Fast-WAM in the text should align with the figure's terminology.","section":"Figure 1"},{"comment":"The term 'readability registers' is introduced but never explained; a brief definition or reference is needed for readers unfamiliar with the Wan architecture.","section":"Sec. 3.2"},{"comment":"The action expert is 'initialized from a linearly interpolated Wan2.1 ActionDiT checkpoint'; the term 'ActionDiT' and the interpolation procedure are not defined, so please clarify what this component is and how the interpolation is performed.","section":"Sec. A"},{"comment":"The temporal ratio of 4 between the action and video streams is stated without justification; a short explanation of its effect on the learned dynamics would improve reproducibility.","section":"Sec. A"},{"comment":"The abstract and conclusion say the results 'demonstrate' that direct-policy WAMs can retain efficient action prediction while exposing predictive dynamics; given the explicit caveats in Sec. 4.4 and Sec. 5, a more measured wording such as 'suggest' or 'indicate' would better match the evidence.","section":"Abstract and Conclusion"},{"comment":"There are several typographical issues, including the extra space in 'W AMs' throughout the paper and the reference entry 'Shuang Li Yihuai Gao Dorsa Sadigh and Shuran Song. Unified video action model,' which appears incomplete and should be checked.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The authors are commendably transparent about the limitations of their component analysis, but the abstract and conclusion assert a mechanistic interpretation that the experiments do not yet establish. The main risk is that the reported success of ForeWAM may be due to added conditioning capacity rather than a learned predictive-dynamics pathway. I believe this is fixable with matched ablations, uncertainty reporting, and a specification of the LIBERO-Plus subset, hence my recommendation of major revision rather than reject. The paper would also benefit from a more careful separation of measured results from architectural design intent."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"There's something here, but it's a proposal more than a demonstrated mechanism. The paper identifies a real gap in direct-policy WAMs—after removing future-video generation, there's no explicit route for predictive dynamics to reach the action pathway—and offers a concrete interface: Future-KV, where a single video prefill over the clean current latent and noise-filled future slots produces a K/V cache reused throughout action denoising, plus dynamics registers supervised by a frozen LaWM teacher. That combination is genuinely new relative to Fast-WAM and LaWAM, and the efficiency numbers are concrete: 568 ms action latency (220 ms for the distilled variant) at about one-third of Fast-WAM's parameters, with roughly matched LIBERO success.\n\nThe paper is also honest about what it can't yet claim. It explicitly says the LIBERO-Plus comparison to Fast-WAM is 'descriptive rather than coverage-matched,' that the base-policy row in the ablation is a contextual reference under a different coverage profile, and that the aggregate ablation 'does not by itself establish the causal contribution of either pathway.' Those admissions are in the right places.\n\nThe soft spots are real, though. There are no error bars or significance tests anywhere, so the 3–4 point gaps in the ablation could be noise. The 'observed subset' of LIBERO-Plus is never defined, so the headline 61.6% number is hard to situate. Most importantly, the stress-test concern holds up: at deployment the future slots are pure noise (Eq. 3), while training supervises the video branch on ground-truth future latents. The paper asserts that the slots' usefulness is learned from the joint video–action objective, but nothing in the objective guarantees the K/V states computed from noise carry task-relevant future information. The Future-KV path could be acting as an observation-conditioned context that adds capacity without encoding futures. The paper's own design rationale—that the two pathways impose different inductive biases—is plausible, but it isn't evidence.\n\nThat said, this isn't a desk reject. The architecture is worth testing, and a serious referee can push for the missing controls: matched evaluation with confidence intervals, a defined LIBERO-Plus subset, and an intervention that swaps the noise substrate for something else to see whether the predictive content actually matters. The paper is for robot-learning researchers thinking about efficient world-conditioned policies; they'll find a clear statement of the design space and a sensible candidate. I'd send it to review.","headline":"A plausible new interface for direct-policy WAMs, honest about its limits, but the central mechanism—that noise-filled future slots carry predictive dynamics at deployment—is asserted rather than demonstrated.","tokens_in":13203,"tokens_out":2780,"would_cite":true,"duration_ms":26511,"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":"A robot policy can exploit predictive world dynamics without ever generating future video, via a latent key-value interface that keeps action generation fast.","keywords":["world action model","latent futures","key-value caching","video diffusion transformer","robot manipulation","LIBERO","flow matching","dynamics registers"],"falsifier":"Take the trained ForeWAM and replace the stochastic future slots with a constant value (all zeros) before the prefill, or skip the future-slot tokens entirely while keeping the current-latent keys and values. If LIBERO-Plus success remains near 61.6%, the latent-future interface is not doing the causal work claimed; if success drops substantially, the noise-substrate prefill is genuinely carrying task-relevant context into the action decoder.","tokens_in":12195,"feed_emoji":"🤖","tokens_out":10189,"duration_ms":101025,"temperature":0.7,"pith_summary":"World Action Models that generate future video before acting pay a large inference cost, while direct-policy variants that skip generation lose an explicit route through which predictive dynamics reach the action decoder. ForeWAM offers a resolution: it keeps the fast, direct-policy structure but exposes predictive context through a single video-model prefill over the current visual latent and noise-filled future slots, caching the resulting key-value states for the action decoder to read. A set of dynamics registers, trained against a frozen latent-action teacher, steers those hidden future states toward interaction-induced transitions such as object motion and contact changes. The central claim is that a policy can be future-aware at inference without ever materializing future observations, and the reported numbers -- 96.7% success on LIBERO, 61.6% on LIBERO-Plus, 568 ms action latency (220 ms after distillation) at about one-third of the baseline's parameters -- are offered as evidence. The reason to care: predictive world modeling becomes a conditioning signal rather than a simulated rollout, so foresight no longer has to cost control speed.","feed_headline":"Robot policy hits 96.7% without generating future video","feed_subtitle":"ForeWAM reuses cached video key-values as predictive context, cutting action latency to 568 ms with one-third the parameters.","key_machinery":"The central object is Future-KV, a latent key-value interface between the video backbone and the action backbone. It performs one Video DiT prefill over the clean current visual latent concatenated with noise-filled future slots, then caches the layer-wise keys and values so that action queries can attend to both current and latent-future context at every denoising step without iterative future-video decoding. The second mechanism is a set of $N_D=16$ dynamics registers, mean-pooled and projected through a trainable head to match a frozen latent-action teacher's target; this training-only signal biases the registers toward interaction-induced transitions. A structured attention mask routes current tokens, registers, and future-slot tokens so action tokens can read the full video sequence, while future-slot tokens integrate the current frame and the registers. These two pathways, one distributed and one compact, are what the paper argues carry predictive dynamics into the action pathway without explicit future generation.","core_discovery":"ForeWAM claims to decouple predictive foresight from future-video generation. The deployment-time policy is still a direct policy $p(a_{1:H}|o,l,p)$ with no ground-truth future, no teacher, and no decoded video; the predictive interface is a latent substrate $\\tilde{z}^F_{1:T} = \\mathrm{concat}(z_{cur}(o), \\epsilon_F)$ prefilled once through the Video DiT at $\\sigma=1.0$, yielding per-layer key-value states $H_{KV}$ and a dynamics-register slice $D_{\\theta}$ that the Action DiT reads throughout denoising. Training combines a flow-matching video loss on demonstrated future latents, a flow-matching action loss, and a latent-action distillation loss that pulls the mean-pooled registers toward a frozen teacher's representation of the demonstrated visual transition. These ingredients are meant to make the noise-filled future slots useful as task-relevant predictive context even though they are stochastic at deployment. The paper reports 96.7% average success on LIBERO and 61.6% on its observed LIBERO-Plus subset (96.9% and 58.2% for the two-step distilled Flash variant), with 568 ms (220 ms Flash) standalone action-generation latency using roughly one-third of Fast-WAM's parameters.","pith_inferences":["The paper compares Future-KV-only and LA-supervision-only configurations, but it does not ablate the future slots themselves while keeping the current-latent prefill, so the causal role of the stochastic substrate is not fully isolated.","Because the future slots are stochastic, an immediate extension is to make the substrate action-conditioned or learned, which could strengthen the predictive content of the cache and likely improve performance on contact-heavy tasks.","The same latent-future interface could be transferred to settings without paired before/after observations by distilling a latent-action teacher into the registers during pretraining, then fine-tuning on tasks where only current observations are available.","The reported design does not require the video prefill to be refreshed after every action chunk; testing whether the cached key-value states can persist across closed-loop steps would clarify how much of the foresight is truly long-horizon."],"forward_implications":["Predictive world modeling can be used as a conditioning interface rather than a generative step, so future-aware robot policies need not pay iterative video-denoising costs at inference.","Training with ground-truth future latents and a teacher, then deploying with neither, means the same policy can benefit from privileged simulation signals without requiring them at execution time.","The Future-KV interface is parameter-light: the full policy uses about one-third of Fast-WAM's parameters while staying within roughly a point of its LIBERO success, widening the design space for resource-constrained robot systems.","The distilled two-step variant, ForeWAM-Flash, retains LIBERO performance and reaches 220 ms action-generation latency, indicating that the latent-future context survives aggressive action-denoising acceleration.","The larger gains on camera-viewpoint and sensor-noise LIBERO-Plus perturbations suggest that latent predictive context can improve robustness to observation distribution shift, though the paper presents this evidence as descriptive rather than coverage-matched."],"supporting_citations":[{"why":"Defines Fast-WAM, the direct-policy baseline that this paper benchmarks against and improves on in latency, parameters, and LIBERO-Plus success.","marker":"Yuan et al., 2026b"},{"why":"Supplies the frozen LaWM latent-action teacher that supervises the dynamics registers during training.","marker":"Chen et al., 2026a"},{"why":"Provides the OneDP diffusion-distillation method used to produce the accelerated ForeWAM-Flash variant.","marker":"Wang et al., 2024"},{"why":"Defines the four LIBERO suites on which the in-distribution success rates are measured.","marker":"Liu et al., 2023"},{"why":"Defines LIBERO-Plus, the robustness benchmark whose perturbations drive the out-of-distribution comparison.","marker":"Fei et al., 2025"},{"why":"Supplies the Wan2.1 video DiT, text encoder, and video VAE that initialize the visual branch.","marker":"Wan Team, 2025"},{"why":"Grounds the continuous flow-matching objective used for both video and action training.","marker":"Lipman et al., 2022"}],"fun_headline_variants":["Foresight without video: 96.7% robot success","Predictive future, no video: 96.7% success","Skip future video, keep foresight: 96.7%","Latent futures, no video: robot 96.7%","ForeWAM: 96.7% success, no future video"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the noise-filled future slots at deployment, after one prefill, actually provide the Action DiT with task-relevant predictive information rather than serving as a fixed stochastic input whose usefulness is only an artifact of training-time supervision.","fun_headline_variants_meta":{"raw":{"variants":["Foresight without video: 96.7% robot success","Predictive future, no video: 96.7% success","Skip future video, keep foresight: 96.7%","Latent futures, no video: robot 96.7%","ForeWAM: 96.7% success, no future video"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000525,"raw_usage":{"total_tokens":2619,"prompt_tokens":1110,"completion_tokens":1509,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":726,"completion_tokens_details":{"reasoning_tokens":1417}},"tokens_in":726,"tokens_out":1509,"duration_ms":10827,"temperature":1.0,"reasoning_tokens":1417,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T00:33:41.819401+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the trained ForeWAM and replace the stochastic future slots with a constant value (all zeros) before the prefill, or skip the future-slot tokens entirely while keeping the current-latent keys and values. If LIBERO-Plus success remains near 61.6%, the latent-future interface is not doing the causal work claimed; if success drops substantially, the noise-substrate prefill is genuinely carrying task-relevant context into the action decoder.","supporting_citations":[],"review_version":1}