{"id":"c5a09548-af5a-435e-86c0-b0ea64885aa3","arxiv_id":"2607.29235","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A training-free masked pseudoinverse-guidance mechanism injects newly observed states and committed actions into the active flow-matching chunk of world-action models, improving long-horizon execution on selected manipulation benchmarks.","lead":"This paper adds an inference-time correction loop to flow-matching 'world-action' models that predict robot futures: while one chunk of actions is being executed, newly observed images are fed back into the still-running generation to keep the predicted future on track. It is training-free and improves success rates on selected manipulation benchmarks, though the gains are modest and partly tuned on the benchmark tasks themselves.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reported gains may come from the action-prefix constraint, not FBFM's novel state feedback: the RTC-only mask mode is defined but never reported for task success.","rationale":"I read the paper in good faith. The FBFM construction is coherent: Eq. (19) is a standard pseudoinverse-guided field, the stage-wise and joint instantiations are specified carefully, and the authors honestly disclose the aligned-coordinate approximation, the finite-prefix artifacts in Appendix F, and the limited DreamZero gains. Those disclosures strengthen the manuscript's credibility. However, the central empirical claim is that the newly observed latent-state feedback improves task success, not merely that it changes the velocity field. The paper defines an action-only RTC mode and says the implementation exposes a common mask-controlled path, but it never reports task-success numbers for that mode. This is the single most load-bearing gap because the only reported benchmark comparisons are Base vs full FBFM. Since the previous-action prefix constraint alone is already known to improve action-chunk continuity (RTC), the +2.98 pp and +0.625 pp results may not validate the new state-feedback component. The mechanism diagnostic in §5.2 is suggestive but only demonstrates that the cache switch alters the action path; it does not establish that the alteration improves task success. I therefore identify this attribution gap as the key concern, rather than the aligned-coordinate approximation, which the paper already qualifies as model-specific and which is supported by the authors' own Appendix F artifact analysis. My recommended verdict remains conditional: the FBFM claim should not be accepted as stated until the RTC-only ablation is reported and shows FBFM is better than RTC-only at the task-success level.","tokens_in":23922,"tokens_out":11550,"duration_ms":130268,"concrete_test":"Run the existing three-mask protocol on the same RoboTwin2.0 42-task and LIBERO 40-task sets, with identical seeds, solver budget, and pseudo-clock, and report equal-weight macro success for NONE, RTC (action mask only), and FBFM (action + state mask). If RTC-only success is statistically indistinguishable from FBFM (e.g., paired bootstrap 95% CI of the difference includes 0), then the state-feedback mechanism is not responsible for the gains and the central claim fails. If FBFM beats RTC-only beyond noise, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim credits newly observed latent states, injected through Eq. (19), with improving task success. But the paper never reports task-success results for the action-only mask mode, even though §C.1 defines NONE, RTC (action mask only), and FBFM (action + state mask) as sharing checkpoint, noise, solver budget, and pseudo-clock. The main RoboTwin and LIBERO tables compare only Base vs FBFM. The +2.98 pp RoboTwin gain and the +0.625 pp LIBERO pooled gain could therefore be produced entirely by the fixed previous-action prefix constraint, which is the existing RTC mechanism [47], not by FBFM's state feedback. This is not merely a missing ablation: the LingBot-VA schedule in §C.2 releases the first complete state target only after the 16th action of the 16-action suffix, when the 25-step video flow has only two evaluations left. The state mask can correct at most the final numerical video update; the rest of the state chunk is generated open-loop. The mechanism diagnostic in §5.2 shows that state feedback changes action velocity, but a computational effect is not evidence of task-success gain. Without an RTC-only comparison, the paper's proposed state-feedback mechanism is not established as the cause of the reported improvements.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Feedback Flow Matching (FBFM), a training-free inference-time mechanism that applies masked pseudoinverse guidance to the conditional velocity field of a frozen Flow-Matching world-action model (WAM). FBFM injects two kinds of time-aligned constraints into the active generation chunk: dynamically observed latent states (state feedback) and previously committed actions from the preceding chunk (action-prefix consistency). The method is instantiated on a stage-wise WAM (LingBot-VA on RoboTwin2.0) and a joint-generation WAM (DreamZero on LIBERO), reporting equal-weight task-success gains of +2.98 percentage points on RoboTwin2.0 and +0.625 percentage points pooled on LIBERO, plus a real-world video-prediction diagnostic showing improved state tracking. The central claim is that state feedback, injected through Eq. (19), improves task success and visual re-grounding without modifying or retraining the WAM.","tokens_in":24208,"tokens_out":5955,"duration_ms":68381,"significance":"If established, FBFM would be a useful, simple, and architecture-agnostic interface for closed-loop correction in flow-matching WAMs, complementing action-only chunking methods like RTC. The paper is unusually transparent about implementation details: it provides tensor layouts, pseudo-clock schedules, commit hashes, a full parameter-search ledger, and an explicit statement and bound for the aligned-coordinate approximation in Appendix B.3. The real-world observation-prediction diagnostic is also a valuable step beyond pure simulation. However, the manuscript's headline empirical claims are not yet conclusive: the task-success tables never isolate the state-feedback component from the fixed action-prefix constraint, and the DreamZero gains are obtained with hyperparameters selected on the same benchmark tasks used in the final evaluation.","major_comments":[{"comment":"The central attribution of task-success gains to state feedback is not supported, because the task-success tables report only Base vs. FBFM. Appendix C.1 defines three mask settings—NONE, RTC (action mask only), and FBFM (action + state mask)—and states that they share checkpoint, noise, solver budget, and pseudo-clock. Since FBFM always includes the RTC-style fixed previous-action prefix, the reported +2.98 pp (RoboTwin) and +0.625 pp (LIBERO pooled) could in principle be produced entirely by the action-prefix constraint alone, which is not the paper's novel contribution. The §5.2 mechanism diagnostic shows only that state feedback changes computed action velocities, not that it affects task success. Please add task-success results for the RTC-only mode on the same tasks and configurations; this is a necessary control for the paper's central claim.","section":"§C.1, Tables 1 and 2"},{"comment":"The DreamZero configuration is not independently evaluated. The state preconditioner PZ and proportional gain kp were selected by screening on LIBERO-Spatial, LIBERO-Object, LIBERO-Goal, and LIBERO-10 subsets (E.2–E.4), and the same suites appear in the final pooled estimate in Table 2. The selected gain kp=0.0486968 is the best point estimate on the screened tasks ('neither paired comparison establishes statistical superiority'), and the final table therefore reflects a configuration chosen on the evaluation data. Please provide a held-out evaluation on tasks or suites not used in tuning, or report a cross-validated estimate with appropriate uncertainty. At minimum, quantify how much of the Table 2 gain is specific to the screening tasks.","section":"Appendix E, Table 2"},{"comment":"The RoboTwin timing schedule undermines the interpretation of the LingBot-VA gains as evidence for fine-grained state feedback. The first complete dynamic state target is formed only after action 16 of the 16-action suffix, when only 2 of the 25 numerical video updates remain. Thus state feedback can correct at most the final numerical video update; the remainder of the state chunk is generated open-loop. The task-success gain could stem from the fixed action-prefix constraint, from the corrected-context refresh of the action stage, or from a combination—but not from direct visual correction of the active chunk, as the paper's phrasing suggests. The RTC-only control requested above would disentangle this, but the timing itself should also be discussed explicitly in the interpretation of Table 1.","section":"§C.2, Table 1"},{"comment":"The RoboTwin2.0 evaluation uses a 'selected' set of 42 tasks, with longer tasks excluded. The abstract and introduction motivate FBFM by 'long-horizon reliability' and 'drift in long-horizon tasks,' but the main success-rate experiment omits the longer tasks where drift would be most severe. Please justify the selection criterion or include at least a sample of longer tasks; as presented, the evidence for the long-horizon motivation is indirect.","section":"§D.1, Abstract"}],"minor_comments":[{"comment":"The 'State preconditioner' row renders as '1P Z = 56/9600', which is ambiguous. Please separate the LingBot-VA value (presumably 1) and the DreamZero value (PZ = 56/9600) into distinct columns or cells.","section":"Table 5"},{"comment":"The mechanism diagnostic reports MSE and RMS changes on four paired task–trial units. No error bars or test statistics are given; given the small sample, please state whether the reported difference (e.g., 0.6828 vs. 0.6751) is within run-to-run variability.","section":"§5.2/Figure 5"},{"comment":"The sentence 'This formulation follows the Flow-Matching inpainting construction of Black et al. [47]' conflates pseudoinverse guidance (Song et al. [9]) with the RTC action-chunking paper. Please cite both appropriately and distinguish the general inverse-problem guidance from the action-prefix specialization.","section":"§3.2"},{"comment":"The real-world diagnostic uses all 30 measured latent slots (full-oracle coverage), while the paper's own Appendix F shows that finite-prefix coverage produces large artifacts. Please clarify the deployment relevance of the full-coverage condition and discuss how the finite-coverage failure mode would be handled in actual closed-loop execution.","section":"§5.4/Appendix F"}],"recommendation":"major_revision","confidential_remarks":"The paper has a genuinely useful mechanism and unusually thorough implementation appendices, but the empirical case for the state-feedback contribution is not yet established. The missing RTC-only task-success ablation is the main blocker; without it, the reported gains are compatible with the known action-prefix effect. The DreamZero hyperparameter selection on the evaluation tasks is a second, equally important concern. Both are addressable within the manuscript's scope, so I recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know two things before spending time on this one. First, the paper's central causal claim—that the new state-feedback path improves task success—is not actually demonstrated. The action-only mask mode (RTC) is defined in §C.1 but never reported for task success. The main tables compare only Base vs FBFM, so the +2.98pp RoboTwin gain and +0.625pp LIBERO pooled gain could come entirely from the existing action-prefix constraint, which is prior work (Black et al. RTC). Second, the DreamZero LIBERO numbers are not an independent evaluation: kp and PZ were screened on the same LIBERO tasks used in the final table, and the abstract's \"over 5%\" overclaims the full-text numbers (3.0pp on RoboTwin, 0.6pp pooled on LIBERO).\n\nThat said, the paper has real substance. The idea—treating the active flow-matching chunk as a dynamically partially observed generation problem and using masked pseudoinverse guidance to inject both fresh latent states and committed actions—is cleanly formulated. The unified interface for stage-wise and joint WAMs, including the cross-modal Jacobian path in joint models, is a genuine extension of RTC. The derivation is sound conditional on the aligned-coordinate approximation, which the authors state plainly in Appendix B.3. The implementation detail is unusually careful: audited commits, exact tensor layouts, schedules. The mechanism diagnostic in §5.2 is a legitimate attempt to trace the state-feedback-to-action path, and the real-world video tracking experiment is an honest diagnostic even if it is not a closed-loop task.\n\nThe soft spots are proportionate to the claims. The missing RTC-only ablation is the load-bearing one. The schedule in §C.2 makes it worse: for LingBot-VA, the first complete state target arrives only after the 16th action, when only two video evaluations remain, so the state mask can correct at most the final numerical video update. That is a very thin channel for re-grounding. The real-world demo conditions on recorded future frames, not on an asynchronous execution loop, and Appendix F documents large artifacts when feedback covers only a finite prefix. The paper is honest about several of these limitations, which earns some credit, but honesty about a missing ablation does not substitute for running it.\n\nWho should read this: anyone working on inference-time control for flow-matching policies or world models. The formulation and the implementation are useful even if the empirical claims are not yet trustworthy. For peer review, this deserves a serious referee, but it needs major revision before acceptance: report RTC-only task-success numbers, do held-out hyperparameter selection, and fix the abstract. If the RTC-only numbers show no added gain from state feedback, the central novelty collapses; if they show a gain, the paper becomes a solid contribution.","headline":"A clean, well-formulated inference-time feedback mechanism for world-action models whose headline task-success gains are not yet attributable to the new state-feedback path, because the action-only ablation is never reported.","tokens_in":24714,"tokens_out":5241,"would_cite":true,"duration_ms":51816,"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":"This paper claims that a frozen flow-matching world-action model can be re-grounded at individual time steps inside its active chunk by imposing newly observed latent states and committed actions as masked pseudoinverse constraints on the v","keywords":["world-action models","flow matching","training-free guidance","pseudoinverse guidance","closed-loop robot control","latent state feedback","asynchronous execution","manipulation benchmarks"],"falsifier":"On the auxiliary ball-collision sequence, compare the onset of large-area image corruption under finite 10-, 20-, and 30-slot feedback coverage. If the corruption onset does not track the end of the measured prefix, or if the same artifact appears when the exact decoded-and-reencoded residual h†(h(ˆX^1)) is substituted for the aligned-coordinate residual, then the paper's mechanism-level explanation—codec/backbone mismatch as distinct from the aligned-coordinate approximation—would be falsified.","tokens_in":23809,"feed_emoji":"🤖","tokens_out":4430,"duration_ms":46705,"temperature":0.7,"pith_summary":"This paper tries to show that a world-action model—a system that jointly predicts future images and actions—can be corrected during generation, not just between chunks. The authors propose FBFM, a training-free inference mechanism that treats the flow-matching solver as a dynamically partially observed generation problem: newly encoded real observations and actions committed by the previous chunk become time-aligned constraints on the active chunk. A masked pseudoinverse guidance term steers the velocity field toward those measurements while the pretrained model stays frozen. On a stage-wise model the correction raises equal-weight task-configuration success by about three percentage points; on a joint-generation model the pooled gain is modest and suite-dependent. The paper also records an unresolved limitation: when state feedback covers only a finite prefix of a generated video, large visual artifacts appear after the measured prefix ends, which it attributes tentatively to the aligned-coordinate approximation or a codec/backbone mismatch.","feed_headline":"Training-free feedback lifts robot success by 3 points","feed_subtitle":"Masked pseudoinverse guidance pulls live observations into the active chunk, suppressing drift without retraining.","key_machinery":"The central object is the masked pseudoinverse guidance update (Eq. 19): v^Q_FBFM = v̄^Q + λ (J^Q)^T P^Q W^Q (Y^Q − Q̂^1), where Q̂^1 is a clean-endpoint estimate from the current noisy sample, J^Q is its Jacobian, W^Q is the time-aligned support or confidence mask, P^Q is a modality scale preconditioner, Y^Q is the lifted target (committed previous action or encoded real latent state), and λ is a flow-time-dependent guidance strength. The mask converts chunked inference into a dynamically partially observed generation problem; the transpose Jacobian propagates residuals across state and action coordinates in joint models; the guidance schedule controls how strongly the frozen prior is pulle","core_discovery":"The central claim is that re-grounding can be pushed inside the actively generated chunk. While a world-action model is still solving its flow, the newly encoded latent state z_{t+i} and the preceding committed action a^{prev}_{t+i} can be imposed as masked targets on the corresponding temporal slots of the chunk. At every solver evaluation, the clean-endpoint estimate is compared with the lifted target on the masked subspace, and the discrepancy is propagated through the endpoint Jacobian as a vector–Jacobian product that corrects the velocity field. No parameters are updated. The paper argues that this time-aligned masked pseudoinverse correction is a common interface for both stage-wise W","pith_inferences":["The same masked pseudoinverse interface could be lifted out of the WAM setting and applied to other latent-conditional flow or diffusion generators whenever real measurements arrive during sampling, turning this paper's recipe into a general online-guidance primitive.","The paper's aligned-coordinate approximation suggests a concrete testable refinement: decoding predicted endpoints and re-encoding them before forming the residual might reduce the post-prefix artifacts that the paper observes in Appendix F, at the cost of extra compute.","For joint models, the gain search shows that reciprocal scaling inside one solver linearization does not guarantee closed-loop stability after actions are executed; a natural next step, which the paper explicitly leaves open, is to close the loop with a learned or actual environment-transition model and use PID-style feedback instead of a fixed proportional gain.","Because FBFM only requires feedback to be visible before a solver evaluation, it could be combined with a learned predictor that anticipates future observations, allowing corrections to be applied before the physical measurement arrives."],"forward_implications":["FBFM applies to any frozen flow-matching WAM without retraining, preserving the native chunking, solver schedule, KV cache, and classifier-free guidance settings.","On LingBot-VA over 42 RoboTwin2.0 tasks, the equal-weight task-configuration success rate rises from 80.1% to 83.1% overall, with +2.86 points in the clean setting and +3.10 points in the randomized setting.","On DreamZero over 800 LIBERO episodes, the pooled success rate rises from 70.125% to 70.75%, with a 2.5-point gain on LIBERO-Goal and LIBERO-10 and small decreases on LIBERO-Spatial and LIBERO-Object.","On real-world robot observation prediction, causally injecting 30 measured latent slots lowers MAE from 9.63 to 9.27 and raises PSNR from 20.06 to 23.10 dB, improving visual tracking of the task-relevant ball position.","In joint-generation WAMs, the cross-modal Jacobian block J^T_ZA lets a state residual directly correct action coordinates in the same solver step, without waiting for a separate context handoff."],"fun_headline_variants":["Training-free feedback boosts robot success over 5% in long-horizon tasks","No-retrain feedback loop fixes robot predictions mid-chunk","Asynchronous flow feedback suppresses drift in robot control","Flow-matching robots get live corrections without training","Pushing feedback into the active chunk lifts robot success"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise is the aligned-coordinate approximation h†(h(ˆX^1)) ≈ ˆX^1 on the masked subspace: if the frozen encoder's latent coordinates do not match the model's predicted clean-latent coordinates, injected measurements can push generation off the data manifold and produce the large visual artifacts the paper observes after a finite measured prefix ends.","fun_headline_variants_meta":{"raw":{"variants":["Training-free feedback boosts robot success over 5% in long-horizon tasks","No-retrain feedback loop fixes robot predictions mid-chunk","Asynchronous flow feedback suppresses drift in robot control","Flow-matching robots get live corrections without training","Pushing feedback into the active chunk lifts robot success"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000803,"raw_usage":{"total_tokens":3404,"prompt_tokens":821,"completion_tokens":2583,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":565,"completion_tokens_details":{"reasoning_tokens":2516}},"tokens_in":565,"tokens_out":2583,"duration_ms":18223,"temperature":1.0,"reasoning_tokens":2516,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T11:02:19.298552+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On the auxiliary ball-collision sequence, compare the onset of large-area image corruption under finite 10-, 20-, and 30-slot feedback coverage. If the corruption onset does not track the end of the measured prefix, or if the same artifact appears when the exact decoded-and-reencoded residual h†(h(ˆX^1)) is substituted for the aligned-coordinate residual, then the paper's mechanism-level explanation—codec/backbone mismatch as distinct from the aligned-coordinate approximation—would be falsified.","supporting_citations":[],"review_version":1}