{"id":"f379ed18-f842-4c84-949e-6f4ff75550ce","arxiv_id":"2608.12939","paper_version":1,"verdict":"ACCEPT","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"low","formal_verification":"none","parameter_count":6,"one_line_summary":"ACPC, a rollout-consistency diagnostic for JEPA world models, bounds perturbation-induced prediction-error and planning-cost changes and, together with IR and SR, identifies checkpoints that stay robust under visual perturbations.","lead":"This paper introduces a diagnostic, ACPC, that measures how much a visually perturbed view of a history diverges from the clean version after a JEPA world model rolls both forward under the same actions. The measure comes with proved bounds on prediction-error and planner-cost changes, and an IR/SR screen that selects robust checkpoints across tasks.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Unstated task-preservation assumption in Proposition 1 lets small ACPC coexist with large true prediction-error changes when perturbations alter the underlying state.","rationale":"The paper's central theorem is a reverse-triangle inequality; it is unquestionably correct. The load-bearing step is the semantic identification of e_tilde_h as the perturbation-induced error. That identification holds only when the same observed future Y_H^a is the correct target for both the clean and perturbed rollouts, i.e., when the perturbation is task-preserving. The reader flagged exactly this assumption. It is not merely cosmetic: a perturbation that changes the true state (for instance, an object moved, or a different episode's history spliced in) leaves the proposition's inequality true but makes it irrelevant, because the perturbed model's true error should be measured against the future actually generated from the perturbed state. Low ACPC can then coexist with large true error change and with planning failure, so IR/SR could give a false 'robust' label. The paper's experiments avoid this by keeping the goal clean and perturbing only history observations; within that domain the results stand. But the abstract and Proposition 1 do not state the restriction, and Section 3.2's justification ('both predicted rollouts are compared with the same observed future') assumes it silently. The concrete test I propose directly measures whether the unqualified claim fails under a state-changing perturbation. If it fails, the paper's claims should be explicitly scoped to observation-level, task-preserving shifts; if the test unexpectedly holds across the family, the concern is resolved. I therefore recommend making acceptance conditional on adding this qualifier and re-running the pair-level check under at least one state-changing perturbation.","tokens_in":25113,"tokens_out":12852,"duration_ms":133719,"concrete_test":"Run the Section 4.4 pair-level protocol with a perturbation that alters the true state while keeping the observation transformation comparable: in TwoRoom, take each clean history and perturb it by shifting the agent's logged position by a small offset in the observation, or by splicing in the observation from another episode at a nearby state. For each pair, compute ACPC_H under the recorded action sequence and also the true perturbed error e_tilde_true = ||G(E(h_tilde)) - Y_H^a(h_tilde)|| against the future actually observed from the perturbed state under the same actions. Check whether |e_tilde_true - e_h| <= ACPC_H holds for all pairs and whether low-ACPC checkpoints still pass the IR-SR screen while planning success changes. A violation on any pair, or a passed screen with degraded success, confirms that the unqualified bound fails without the task-preservation assumption.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The mathematical inequality in Proposition 1 is correct, but its advertised meaning is not. e_tilde_h is defined as ||G(E(h_tilde)) - Y_H^a|| with Y_H^a the observed future of the clean history. This is the perturbed model's actual prediction error only if the perturbation leaves the task-relevant state unchanged, so that the same action sequence leads to the same future. Section 3.2 asserts this by saying 'the perturbed history is created by applying a visual perturbation to the clean history,' and Section 2 mentions 'task-preserving visual perturbation,' but Proposition 1, the abstract, and the opening of Section 3 state the bound without this qualifier. If a perturbation changes the true state, the correct comparison target for the perturbed rollout is Y_H^a(h_tilde), not Y_H^a, and |e_tilde_true - e_h| can be large even while ACPC is near zero (for example, if the encoder is insensitive to the state change but the true future diverges). The checkpoint screen inherits this gap: low IR and high SR could certify a model as robust to a shift that actually invalidates the plan. This is a scope gap rather than a proof error; the fix is to state the task-preservation condition explicitly before Proposition 1 and to restrict the central claims to observation-level perturbations.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces Action-Conditioned Predictive Consistency (ACPC), a diagnostic that measures the rollout-level divergence between a clean history and a visually perturbed view when both are rolled forward under the same action sequence. It proves two samplewise bounds: Proposition 1 bounds the change in multi-step prediction error by ACPC, and Proposition 2 bounds the change in a model-based planner's predicted cost and gives conditions under which the winner or elite set is preserved. The paper then defines checkpoint-level summaries—Invariance Radius (IR) and Separation Rate (SR)—and empirically evaluates them on four visual control tasks with LeWM and one with PLDM, under Gaussian noise, blur, and resize. The experiments include cross-validated regressions for prediction-error drift and CEM selection regret, destroyed-action controls, and a threshold screen selected on some tasks and applied to held-out tasks.","tokens_in":25349,"tokens_out":6373,"duration_ms":64515,"significance":"If the results hold, the paper provides a useful and inexpensive diagnostic for JEPA world models: pairwise ACPC is a computable quantity that gives samplewise, assumption-light bounds on how a visual perturbation changes prediction error and model-based planning cost, and the IR-SR screen offers a practical way to compare checkpoints across training conditions. The empirical work is careful in several respects: the prediction-error regression uses group-excluded cross-validation and conservative destroyed-action controls; the threshold transfer experiment uses held-out tasks; and the paper explicitly verifies Proposition 1 numerically on all logged pairs. The authors are also unusually candid about the scope of their claims, including the observation that the screen requires an unaugmented reference and that the planner experiment measures predicted cost rather than simulator return. The main gap is that the formal statements of Propositions 1 and 2 do not state the task-preservation assumption on which their interpretation depends.","major_comments":[{"comment":"The inequality in Proposition 1 is mathematically correct, but its advertised meaning as a bound on 'perturbation-induced change in prediction error' requires the perturbation to be task-preserving: the perturbed history must be an observation-level view of the same underlying state, so that the same recorded future Y_H^a is the correct comparison target for both rollouts. The proof uses Y_H^a for both errors, and the text says 'the perturbed history is created by applying a visual perturbation to the clean history,' but Proposition 1, the abstract, and the opening of Section 3 state the bound without this qualifier. If the perturbation changes the task-relevant state, the correct target for the perturbed rollout is the future of the perturbed state, and |e_tilde_true - e_h| can be large even when ACPC is near zero. The experiments apply perturbations to history observations only and appear to satisfy the assumption, so the empirical conclusions are not affected; however, the formal statement and the abstract should be restricted to observation-level, task-preserving perturbations, with the assumption stated before Proposition 1.","section":"Section 3.2, Proposition 1 and Eq. (4)"},{"comment":"The same task-preservation issue affects the planning-cost interpretation. Proposition 2 bounds the change in the model's predicted cost to a fixed goal embedding g; this is a statement about the model's internal cost, not about actual planning performance. If a perturbation changes the true task-relevant state, the fixed goal is no longer the correct target for the perturbed input, and a small ACPC (or small b_j) does not imply that the planner's real-world behavior is preserved. The paper does mention 'task-preserving visual perturbation' once in the Related Work discussion, but this assumption should be made a standing, explicit condition of Propositions 2 and 3 and of the checkpoint screen's interpretation. The adaptive-CEM certificate in Corollary 1 also inherits this scope condition, since equal elite sets in the model guarantee only equal model-based choices, not equal outcomes in the environment.","section":"Section 3.3, Proposition 2 and Eq. (5)-(7)"}],"minor_comments":[{"comment":"The text says 'Code is available here' but no URL is visible; please include the repository link in the final version.","section":"Reproducibility"},{"comment":"The screening score normalizes the IR margin by |tIR| and the SR margin by |tSR|, making the relative weighting of the two conditions depend on the chosen threshold magnitudes. The authors should justify this scaling or state explicitly that it is a heuristic.","section":"Eq. (12)"},{"comment":"Table 1 reports sensitivity of the IR reduction to the horizon and quantile, but the text also claims stability of the diagnostic; please either report the corresponding sensitivity for SR or revise the sentence to refer only to IR.","section":"Section 3.4 and Table 1"},{"comment":"The cost is divided by the embedding dimension d in the experiments ('C_j = ||x_j - g||_2^2 / d'), while Proposition 2 is stated for the summed squared cost. The rescaling is correct, but the appendix should say this explicitly before the claim that the regret bound is unchanged, since a reader may otherwise think the constant factor affects the bound.","section":"Appendix H"}],"recommendation":"major_revision","confidential_remarks":"The central theoretical contribution is a simple, correct inequality, but its scope is narrower than the abstract and Section 3 claim. The missing task-preservation assumption is fixable with a paragraph, but it is load-bearing for the advertised meaning of the bound, so I recommend major revision rather than minor. The reference list is heavily weighted toward 2025-2026 arXiv preprints; the editor may wish to verify that these citations are accurate and that any not-yet-peer-reviewed items are clearly identified as such."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper's core idea is simple and useful: roll a clean history and its visually perturbed view forward under the same recorded actions, measure the rollout distance (ACPC), and show via reverse triangle inequality and Cauchy–Schwarz that this distance bounds changes in prediction error and planning cost. The proofs are elementary and correct, and the bounds are samplewise with no distributional or smoothness assumptions. That is genuine, non-heuristic grounding for a practical diagnostic.\n\nThe empirical work is where the paper really earns its keep. The destroyed-action controls (zeroed, swapped, shuffled) are a conservative way to show that the recorded action sequence carries information beyond rollout length. The nested regressions are cross-validated at the trajectory-group level, threshold transfer is tested on held-out tasks, and the success-rate criterion was fixed a priori. The IR/SR design is also thoughtful: low IR alone would reward collapsed representations, and the SR check with the SIGReg ablation shows the two measures work together. The paper is honest about what is empirical versus derived.\n\nThe main soft spot is the task-preservation assumption in Proposition 1. The inequality itself is correct for the defined e_tilde, which is the perturbed model's error against the clean observed future. But the claim that this is the perturbation-induced change in prediction error only holds if the perturbation leaves the true underlying state unchanged, so the same action sequence leads to the same future. The phrase 'task-preserving visual perturbation' appears in the related work, but Proposition 1 and the abstract state the bound without that qualifier. If a perturbation changes the true state, small ACPC can coexist with large true prediction-error change. This is a scope gap, not a proof error. The fix is straightforward: state the task-preservation condition explicitly before Proposition 1 and restrict the central claims to observation-level perturbations.\n\nOther concerns are minor. The empirical base is four tasks and three training runs, which limits how strongly the checkpoint-level screen can generalize. The code link and commit hash are not visible in the text, and tIR = 0.3 sits at the edge of the tested range, so the useful threshold range is unresolved. These are worth noting in a revision but do not undermine the central result.\n\nThis paper is for researchers working on world-model robustness, model selection, and JEPA-based control. It deserves a serious referee. Send it to review, and ask the authors to tighten the task-preservation language and make the code artifact explicit.","headline":"A correctly proven and carefully evaluated diagnostic for perturbation robustness in JEPA world models; the central bound is real but its advertised meaning needs an explicit task-preservation assumption.","tokens_in":25910,"tokens_out":1470,"would_cite":true,"duration_ms":15050,"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 rollout-level distance between clean and perturbed histories provably bounds how much visual noise can change prediction error and planning cost — and the screen built on it flags JEPA checkpoints that stay reliable.","keywords":["Action-Conditioned Predictive Consistency","JEPA world models","latent prediction","visual perturbations","planning cost bounds","Invariance Radius","Separation Rate","state separation"],"falsifier":"Keep the underlying state fixed and translate the goal image — which the paper deliberately leaves clean — in a TwoRoom or PushT episode, recompute ACPC between the clean and perturbed histories, and record planning success: if success collapses while ACPC stays small, the task-preserving assumption, not the bound, is carrying the diagnostic. A direct numerical check also exists: any logged pair violating $|e_{\\tilde h} - e_h| \\le \\mathrm{ACPC}_H(h, \\tilde h, a)$ would refute the proven claim, and the paper's own sweep found zero violations.","tokens_in":24905,"feed_emoji":"🎯","tokens_out":17656,"duration_ms":145360,"temperature":0.7,"pith_summary":"Joint-embedding predictive world models (JEPAs) forecast in latent space rather than pixels, which frees them from reproducing nuisance appearance but gives no guarantee against visual perturbations that still shift the encoded representation. This paper tries to establish that a single, model-only measurement — rolling a clean history and its visually perturbed copy forward under the same recorded action sequence and measuring how far the two predicted trajectories drift, called Action-Conditioned Predictive Consistency (ACPC) — bounds how much that perturbation can change prediction error and planning cost, samplewise and with no distributional or smoothness assumptions. From pairwise ACPC the paper builds a checkpoint screen: the Invariance Radius (IR) summarizes how much perturbation spreads same-state rollouts, and the Separation Rate (SR) checks that different states remain distinguishable after rollout, so a representation collapse that would fool a low-IR test still fails on SR. Four visual control tasks, three perturbation types, and two model families support the claim that multi-step ACPC carries information about prediction-error drift and CEM selection regret that encoder distance, one-step ACPC, and same-horizon action-destroyed controls lack. A reader should care because this supplies an assumption-light diagnostic for the visual robustness of latent world models that needs no retraining and relies only on loggable data.","feed_headline":"A rollout distance predicts when visual noise breaks world models","feed_subtitle":"How far the two rollouts drift bounds how much prediction error and planner costs can change.","key_machinery":"The central object is the action-conditioned predictive consistency (ACPC) quantity: a weighted rollout vector $\\bar G_a(z)$ stacks the first $H$ predicted representations of a history under a fixed action sequence, and ACPC is the Euclidean distance between the clean and perturbed versions of this vector. It carries the argument in three ways. The prediction-error bound is the reverse triangle inequality applied to the two errors measured against the same observed future. The planning-cost bound factors the difference of two squared distances to the goal and applies Cauchy–Schwarz, giving the per-candidate cost-change bound $b_j$; ACPC enters by bounding the endpoint displacement $r_j$. At checkpoint level, normalized ACPC values are aggregated into the Invariance Radius, the q90 of clean–perturbed rollout distances scaled by each history's typical one-step motion, and the Separation Rate, the fraction of different-label state pairs whose rollout distance exceeds that radius plus a fixed margin, with a relative-IR comparison against an unaugmented reference checkpoint.","core_discovery":"The central claim is that rolling paired observations forward under identical actions exposes downstream perturbation effects that encoder-only and one-step comparisons miss, and that the measured divergence is quantitatively tied to downstream failure. Concretely, for every paired sample, the difference between the two multi-step prediction errors evaluated against the same observed future is bounded by the ACPC value, $|e_{\\tilde h} - e_h| \\le \\mathrm{ACPC}_H(h, \\tilde h, a)$; when a planner scores candidate action sequences by squared distance to a goal embedding, each candidate's cost can shift by at most $b_j = r_j(\\|x_j - g\\|_2 + \\|\\tilde x_j - g\\|_2)$, so if every clean winner-to-competitor cost gap exceeds the combined bounds, the perturbation cannot reverse the selected plan or elite set. Because these bounds hold per pair and require no distributional or smoothness assumptions, a small ACPC together with large clean cost gaps is a certificate that planning decisions are stable under that perturbation, while a large ACPC diagnoses fragility before the environment is consulted. Summarized across histories, the Invariance Radius and Separation Rate jointly characterize a checkpoint, and the paper reports that checkpoints passing the low-IR/high-SR screen recover planning success under Gaussian noise on four tasks, that thresholds chosen on some tasks transfer to held-out tasks, and that the same qualitative pattern appears on a second architecture and under blur and resize.","pith_inferences":["The paper measures ACPC only on frozen models; a natural extension is to minimize clean–perturbed rollout distance during JEPA training as an explicit visual-robustness regularizer, since the quantity needs nothing beyond logged histories and the perturbation itself.","Because the paper shows both bounds can hold with equality, ACPC is not a loose worst-case estimate; it could serve as a vulnerability score ranking different perturbations by how much they can move a planner's costs on a fixed candidate pool, which the paper does not attempt.","The paper always keeps the goal image clean; perturbing the goal embedding instead would move the target that both costs are measured against, so the present bounds would no longer directly apply — an obvious testable extension is a goal-perturbation analogue of ACPC and a check of whether an analogous bound survives.","The bisimulation framing makes ACPC an empirical, rollout-level equivalence induced by the model itself; one could audit whether a world model's internal equivalence classes coincide with task state across a wider perturbation family, including closed-loop replanning rather than only the recorded open-loop action sequence."],"forward_implications":["A trained JEPA world-model checkpoint can be screened for robustness to a chosen visual perturbation using only logged histories, the frozen model, and an unaugmented reference — no retraining, no environment interaction, and no modification of the planner.","On a fixed candidate pool, ACPC plus clean cost gaps yields a per-pair certificate: whenever every clean winner-to-competitor gap exceeds the summed cost-change bounds, the perturbation provably cannot change the planner's chosen action or its elite set, and if the condition holds across CEM iterations the two runs return identical actions.","Multi-step ACPC under the recorded actions predicts perturbation-induced prediction-error drift better than encoder distance, one-step ACPC, or same-horizon controls with destroyed action information across all 12 task-run cells, and planner-horizon ACPC lowers cross-task prediction error for CEM selection regret in all 12 test cases.","Across the augmentation sweep, lower relative IR and higher SR align with recovery of planning success under Gaussian noise; thresholds selected on held-out tasks identify recovery with balanced accuracy 0.900, and the same low-IR/high-SR pattern appears on a second architecture and under blur and resize, with the sign of the IR-SR score change agreeing with the success criterion in 22 of 24 check","SR closes the collapse loophole: a representation whose latent distances shrink to near zero has the lowest possible IR but its SR falls to 0.066 and clean success to 33%, so the two measures are jointly necessary."],"supporting_citations":[{"why":"LeWorldModel, the primary JEPA world-model family whose trained checkpoints the diagnostic screens and the source of the experimental tasks.","marker":"[10]"},{"why":"PLDM, the second joint-embedding dynamics architecture used to show the low-IR/high-SR pattern transfers across architectures.","marker":"[12]"},{"why":"The cross-entropy method planner whose candidate ranking and elite sets the planning-cost bounds protect.","marker":"[11]"},{"why":"Supplies the transition-based bisimulation criterion that motivates measuring state equivalence through action-conditioned consequences.","marker":"[8]"},{"why":"Provides the bisimulation-metric formulation for latent representations that ACPC's clean-perturbed rollout comparison operationalizes.","marker":"[9]"},{"why":"The closest prior method enforcing action-conditioned rollout consistency during training, which ACPC reframes as a frozen-model diagnostic rather than a training objective.","marker":"[39]"}],"fun_headline_variants":["Rollout drift bounds how visual noise breaks world models","Diagnose JEPA fragility from clean-perturbed rollout drift","Perturbation impact on world models predicted by ACPC","Measure rollout divergence to bound prediction error change","One divergence measure ties visual noise to planner cost"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the evaluated perturbation is task-preserving — it changes how the history looks but not the true underlying state, so the same recorded action sequence and the same observed future are the correct comparison target for both the clean and the perturbed rollout; if a perturbation rewrote task-relevant state, a small ACPC would say nothing about whether the planner still works.","fun_headline_variants_meta":{"raw":{"variants":["Rollout drift bounds how visual noise breaks world models","Diagnose JEPA fragility from clean-perturbed rollout drift","Perturbation impact on world models predicted by ACPC","Measure rollout divergence to bound prediction error change","One divergence measure ties visual noise to planner cost"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001328,"raw_usage":{"total_tokens":5458,"prompt_tokens":1056,"completion_tokens":4402,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":672,"completion_tokens_details":{"reasoning_tokens":4324}},"tokens_in":672,"tokens_out":4402,"duration_ms":33159,"temperature":1.0,"reasoning_tokens":4324,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T20:12:47.203182+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Keep the underlying state fixed and translate the goal image — which the paper deliberately leaves clean — in a TwoRoom or PushT episode, recompute ACPC between the clean and perturbed histories, and record planning success: if success collapses while ACPC stays small, the task-preserving assumption, not the bound, is carrying the diagnostic. A direct numerical check also exists: any logged pair violating $|e_{\\tilde h} - e_h| \\le \\mathrm{ACPC}_H(h, \\tilde h, a)$ would refute the proven claim, and the paper's own sweep found zero violations.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"PLDM, the second joint-embedding dynamics architecture used to show the low-IR/high-SR pattern transfers across architectures."},{"cited_title":"Kroese, Sergey Porotsky, and Reuven Y","cited_arxiv_id":null,"evidence_quote":"The cross-entropy method planner whose candidate ranking and elite sets the planning-cost bounds protect."},{"cited_title":"Bellemare","cited_arxiv_id":null,"evidence_quote":"Supplies the transition-based bisimulation criterion that motivates measuring state equivalence through action-conditioned consequences."},{"cited_title":"McAllister, Roberto Calandra, Yarin Gal, and Sergey Levine","cited_arxiv_id":null,"evidence_quote":"Provides the bisimulation-metric formulation for latent representations that ACPC's clean-perturbed rollout comparison operationalizes."}],"review_version":1}