{"id":"495ef1e6-40c8-4865-9a3c-dea2988272a9","arxiv_id":"2607.25337","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Mining a directed temporal distance from demonstration logs, and choosing the planner cost per task, improves JEPA latent MPC over LeWM on four benchmarks.","lead":"Robots can learn to plan by watching demonstrations, but choosing which imagined future is \"closer\" is hard when no rewards exist. This paper trains a new time-aware score from demonstration logs—how many steps apart two situations are—and shows slightly better planning on four standard control tasks when the right score is used for each task.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Per-task plan-time protocol (d_psi vs L2, and w=0.3 aggregation) is selected from locked results; with one fixed cost TD-JEPA loses on Reacher or Push-T, so 'every environment' is conditional on oracle protocol choice.","rationale":"The reader's weakest assumption identifies cost selection from locked results as the critical protocol issue. My reading agrees and sharpens it: the same concern extends beyond d_psi-vs-L2 to the Reacher planner aggregation w=0.3, where the terminal-only sensitivity check (Table 10) reverses the result. The paper's top-line claim is explicitly 'every environment', and that claim is only true when each environment is assigned its best-performing plan-time configuration. The authors do provide a task-structure rationale—topology-dominated vs contact-rich—which is a reasonable scientific hypothesis, but it is not pre-registered or independently validated; it is consistent with the same Table 11 numbers used for the headline. This is a correctness/validity risk rather than an accusation of selective reporting, and it is addressable. The qualitative contribution—temporal-distance mining can improve JEPA planning—probably survives, but the quantitative 'matches or exceeds on every environment' should be reframed as conditional on the two-role plan-time protocol. Since the reader already recommends CONDITIONAL, this stress-test does not move the verdict; it reinforces the condition.","tokens_in":15333,"tokens_out":6435,"duration_ms":67461,"concrete_test":"Run a fixed-protocol comparison on the four environments using (a) terminal-only latent L2 for all methods (LeWM's default plan cost) and (b) a cost rule chosen a priori from task structure on a validation split, then applied to the locked test manifest. Record Table 2 rows; if TD-JEPA fails to match/exceed LeWM and RC-aux on every environment under either fixed protocol, the 'every environment' claim is conditional on protocol selection. Also report Reacher with w=1.0 and Push-T/OGB-Cube with pure d_psi vs the same baselines.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—'matches or exceeds both methods on every environment'—holds only under a per-environment plan-time protocol chosen after inspecting the locked results. Table 2 uses d_psi on Two-Room/Reacher and latent L2 on Push-T/OGB-Cube; Table 11 shows no single cost achieves the claim: pure d_psi gives Push-T 69.0 (below LeWM 83.6 and RC-aux 81.4) and pure L2 gives Reacher 95.0 (below LeWM 96.0). Reacher also requires a non-default trajectory-mean/terminal aggregation w=0.3; Table 10 shows the terminal-only sensitivity check reverses the comparison (TD-JEPA 74 vs LeWM 86). The stated topology-vs-contact taxonomy is plausible, but it is applied to the same four locked outcomes, so it functions as test-set selection unless an independent validation split or pre-registration establishes it. This is load-bearing because the abstract's 'every environment' claim is exactly what distinguishes TD-JEPA from a method that helps only when the right cost is chosen.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Temporal-Distance-JEPA, an extension of the LeWM JEPA world-model backbone that mines a directed, quasimetric-style temporal cost d_psi from reward-free demonstration logs. Same-trajectory step counts provide regression targets, cross-trajectory pairs provide hinge negatives, and a horizon-matched rollout-consistency loss couples the learned cost to the open-loop planner rollout. At plan time, the paper deploys d_psi on Two-Room and Reacher and latent Euclidean distance on Push-T and OGB-Cube, claiming that under this locked protocol Temporal-Distance-JEPA matches or exceeds both LeWM and the RC-aux baseline on every environment. The paper also presents ablations (directed head, cross-trajectory hinge, rollout consistency), a plan-cost comparison matrix, and diagnostics explaining why pure d_psi underperforms on contact-rich manipulation.","tokens_in":15681,"tokens_out":4185,"duration_ms":47023,"significance":"The core idea is timely and potentially useful: explicitly mining temporal progress from reward-free logs, rather than relying on embedding geometry, is a sensible way to close the train–plan gap in JEPA planners. The paper is also commendable for reporting matched plan-cost matrices, conducting ablations, and releasing code. However, the headline claim—'matches or exceeds both methods on every environment'—is only true under a per-environment plan-time cost selection that is made after inspecting the locked evaluation outcomes. Because no single cost achieves the claim, and because the navigation/contact taxonomy is applied to the same four test environments, the central comparison is not a fair test of the method's superiority. The contribution remains plausible, but the evaluation protocol must be strengthened before the claim is supported.","major_comments":[{"comment":"The claim that Temporal-Distance-JEPA 'matches or exceeds both methods on every environment' is not supported by a single protocol. Table 11 shows pure d_psi yields 69.0% on Push-T (below LeWM 83.6 and RC-aux 81.4) and 77.0% on OGB-Cube (below RC-aux 81.6), while pure latent L2 yields 95.0% on Reacher (below LeWM 96.0). The Table 2 result is therefore obtained by selecting d_psi for Two-Room/Reacher and L2 for Push-T/OGB-Cube. This per-environment selection is made on the same locked outcomes that are then reported as the headline result. Unless the selection rule was fixed before seeing the results, or an independent validation split is used to choose the cost per environment, this is test-set selection and the 'every environment' claim is inflated.","section":"§4.1, Table 2 and Appendix Table 11"},{"comment":"The Reacher result depends on a non-default aggregation rule. Table 10 shows that with w=0.3 (primary) Temporal-Distance-JEPA scores 96% vs LeWM 94%, but with terminal-only scoring (w=1) the comparison reverses: 74% vs 86%. This is a second protocol choice that is selected to make the method look favorable. Similarly, the two-role taxonomy (topology -> d_psi, contact -> L2) is plausible but is justified only by the same four locked outcomes in Fig. 3/Table 11. No independent validation, pre-registration, or leave-one-out procedure is provided. This is load-bearing because the central claim is precisely that the method helps 'on every environment'; under fixed costs or terminal-only aggregation the claim fails on at least one environment.","section":"§4.5, Table 10; §4.3, Fig. 3"},{"comment":"The contact-gate experiment tests hand-designed gates (hard and soft) and finds they do not recover pure L2. This is informative, but it does not address the protocol-selection problem. A learned or validated rule that chooses the plan-time cost from training/validation episodes would be needed to establish that the per-environment choice is not an artifact of peeking at the test manifest. As written, the gate results only reinforce that pure d_psi is suboptimal on Push-T and that the choice of planning cost is consequential.","section":"§4.4, Table 9"}],"minor_comments":[{"comment":"The literature-context table lists LeWM as 87/86/96/74 on the four environments, while the matched Table 2 reports 97.4/96.0/83.6/68.0. The text does state these rows are not matched, but the discrepancy is large and may confuse readers; consider moving the table to an appendix or adding a stronger visual warning.","section":"Table 3"},{"comment":"The Spearman correlations for d_psi are partly by construction, since d_psi is regressed directly to temporal separation tau in Eq. (6). The paper already frames this as calibration rather than transfer, but it would help to state explicitly that Table 6 does not provide independent evidence of temporal ranking quality.","section":"Table 6"},{"comment":"The loss weights lambda_roll=0.5, lambda_td=1.0, lambda_sig=0.09 and margin scale eta are fixed constants. No sensitivity analysis is reported for these hyperparameters, even though the claimed gains are relatively small on Reacher and Push-T. A small sweep or at least a statement that the results are robust to reasonable variations would strengthen the paper.","section":"§3.6, Eq. (9)"},{"comment":"The appendix says RC-aux uses the same locked manifests and latent L2 planning, but it is unclear whether this is the authors' re-implementation or the original code. For reproducibility, specify the source and version of the RC-aux baseline, and similarly for LeWM checkpoints.","section":"Appendix C"}],"recommendation":"major_revision","confidential_remarks":"The manuscript has a solid core idea and careful reporting of cost matrices, but the headline claim is not supported by a single coherent evaluation protocol. The per-environment plan-time cost selection and the w=0.3 aggregation rule are both chosen from the locked results, and removing either choice reverses or weakens the comparison on at least one environment. I recommend asking the authors to either (a) pre-register or independently validate the cost-selection rule, or (b) substantially temper the 'every environment' claim to reflect the protocol-dependent nature of the gains. If a clean validation-based selection cannot be provided, the paper may be better reframed as a diagnostic study showing when temporal costs help and when they hurt, rather than as a method that 'matches or exceeds' baselines on every benchmark."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is a coherent, well-written extension of LeWM. The new bit is the combination: mining a directed temporal cost from reward-free logs and then deciding, per task, whether to use that cost directly or use it only to shape representations for Euclidean planning. That two-role idea is genuinely useful, and the ablations back it up—especially the symmetric-head ablation, which shows the MRN form matters, and the Push-T phase logging, which explains why pure d_psi fails on contact-rich tasks. The matched cost matrix in Table 11 is honestly reported; a reader can see exactly what happens with each cost.\n\nThe soft spot is the central claim. The abstract says \"matches or exceeds both methods on every environment,\" but that is only true if you choose d_psi for Two-Room/Reacher and L2 for Push-T/OGB-Cube after looking at the same locked results. Fix one cost and the claim breaks: pure d_psi gives Push-T 69.0 (below LeWM and RC-aux), and pure L2 gives Reacher 95.0 (below LeWM). The paper's topology-vs-contact taxonomy is plausible, but it is derived from these four environments, not validated on a holdout or pre-registered. Reacher also needs w=0.3 aggregation; with terminal-only scoring the comparison reverses. No significance testing is reported, and most differences are within one standard deviation. Separately, the LeWM baseline on OGB-Cube is 68.0 in Table 2 while the published LeWM is 74; the paper notes these are not matched, but the 14.2-point gain is partly against a weaker baseline.\n\nThese are addressable problems, not fatal ones. The core mechanism—temporal supervision improves JEPA planners—survives. But the quantitative claims need to be revised or the protocol tightened. A serious referee should ask for a pre-registered cost-selection rule, significance or confidence intervals, and a discussion of why the in-house LeWM baselines differ from published numbers.\n\nFor a JEPA planning researcher, this is worth reading. I'd take it to a reading group. I'd probably cite it for the cost-matrix findings and the two-role idea. Send it to peer review—it deserves referee time—but expect heavy revision.","headline":"Solid incremental extension of LeWM, but the 'every environment' claim only holds under a per-task cost selection made after seeing the results; worth peer review with revision.","tokens_in":16164,"tokens_out":3365,"would_cite":true,"duration_ms":35320,"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":"Temporal-Distance-JEPA mines a directed, asymmetric temporal cost from reward-free demonstration logs and shows that using this cost as the plan-time objective or as representation shaping lets a latent JEPA world-model planner match or exc","keywords":["temporal-distance learning","JEPA world models","latent model predictive control","quasimetric cost","cross-entropy method","offline reward-free learning","goal-conditioned planning","representation learning"],"falsifier":"Fix a single plan cost in advance for all four tasks—either always latent L2 or always d_psi—and re-run the locked evaluation; the paper's own Table 11 shows pure d_psi falling to 69.0% on Push-T and 77.0% on OGB-Cube, so any pre-registered single-cost protocol that reproduces those numbers would refute the 'matches or exceeds on every environment' claim. Alternatively, an independent check of Fig. 1's Spearman correlations (0.95 for d_psi versus 0.65 for L2 on held-out Push-T pairs) would test whether the mined cost truly ranks temporal separation as claimed.","tokens_in":15177,"feed_emoji":"🤖","tokens_out":5663,"duration_ms":54401,"temperature":0.7,"pith_summary":"The paper argues that JEPA world models trained for latent prediction do not automatically provide a plan-time ranking of imagined futures; planning usually falls back on Euclidean distance in embedding space, which correlates only weakly with actual progress. It introduces Temporal-Distance-JEPA, which keeps the LeWM backbone and adds a directed temporal cost mined from reward-free demonstration logs: same-trajectory step order gives positive targets, cross-trajectory pairs give negative hinges, and a rollout-consistency loss matches the planner horizon. This mined signal plays two roles: it is deployed directly as the planning cost where progress is topological, and used as representation shaping that improves geometric planning where contact dominates. Under locked evaluation with the cost chosen per task, the method matches or exceeds both LeWM and RC-aux on all four environments, including 100.0% Two-Room success and a 14.2-point gain over LeWM on OGB-Cube. A sympathetic reader would care because it offers an offline, reward-free route to a genuine progress signal for latent MPC without learning a value function.","feed_headline":"Mined temporal cost beats Euclidean planning in all four tasks","feed_subtitle":"Mined from reward-free logs, the directed cost improves planning on all four tasks.","key_machinery":"The central object is d_psi, a directed, non-negative temporal cost parameterized by a metric-residual network: d_psi(z_s, z_g) = ||phi_sym(z_s) - phi_sym(z_g)||^2 + max_k ReLU(phi_asym,k(z_s) - phi_asym,k(z_g)). The symmetric term captures configuration similarity; the asymmetric residual encodes directed reachability, so the cost from A to B can differ from the cost from B to A. This cost is calibrated by regression to same-trajectory step counts tau(i,j) = j - i, pushed above a margin on cross-trajectory pairs, and coupled to the planner through an H-step rollout-consistency loss applied to the same open-loop operator used at plan time. At deployment, d_psi is used directly on topology-do","core_discovery":"Temporal-Distance-JEPA discovers temporal progress structure directly from offline demonstration trajectories and uses it to supervise the same cost that the planner optimizes, rather than letting plan-time ranking be an untrained byproduct of embedding geometry. The central claim is that this mined, directed temporal distance—a quasimetric-style energy with a symmetric component plus an asymmetric residual—outperforms plain latent Euclidean distance on navigation and reaching when deployed directly, and also improves Euclidean planning on contact-rich tasks when used only as representation shaping. Under locked evaluation, the resulting planner matches or exceeds LeWM and RC-aux on every en","pith_inferences":["If the per-task cost-selection rule (d_psi for navigation, L2 for contact-rich manipulation) is not fixed in advance, the headline 'matches or exceeds on every environment' rests on an oracle-like choice; the paper's own Table 11 shows that pure d_psi drops to 69.0% on Push-T and 77.0% on OGB-Cube, so the transferable claim is likely the representation-shaping effect, not the per-task cost deploym","The same mined step-count signal could plausibly serve as a dense reward or goal-reaching value surrogate for offline RL, since it calibrates to steps-to-goal along the demonstration manifold; the paper does not test this, but the calibration makes it a natural next probe.","A learned gating network that switches between d_psi and L2 based on predicted contact phase would directly test the two-role thesis; the paper's hand-gated variants failing to beat fixed L2 suggests the geometric signal must act throughout CEM search, not only near contact.","The 14.2-point OGB-Cube gain under shared L2 planning deserves scrutiny: if it survives a pre-registered single-cost protocol, it isolates a representation-level effect of temporal supervision independent of any plan-cost change."],"forward_implications":["Deploying the mined temporal cost on topology-dominated tasks raises Two-Room success to 100.0% and Reacher to 97.0%, exceeding LeWM's geometric planner on both.","On contact-rich tasks, temporal supervision improves geometric planning on the same checkpoint: OGB-Cube rises 14.2 points over LeWM and Push-T improves, even though the planner still uses latent L2.","The method requires no reward model, value function, or policy: progress supervision comes entirely from reward-free demonstration logs.","The directed head, cross-trajectory negative hinge, and rollout-consistency loss each contribute; removing any one degrades all planner settings on Push-T.","Better temporal ranking does not by itself mean a better contact-rich plan cost: d_psi tracks step order with Spearman 0.91 versus 0.79 for L2, yet pure d_psi underperforms L2 on Push-T and OGB-Cube.","The paper's ablations show that a symmetric Euclidean head, even with the same temporal labels, sharply underperforms the directed MRN head, establishing that cost form and supervision are jointly load-bearing."],"fun_headline_variants":["Mined temporal cost tops Euclidean planning on all tasks","TD-JEPA: plan with temporal distance, not embedding geometry","Offline logs teach JEPA to plan with temporal progress","Temporal cost mined from demos improves latent planning"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The evaluation lets the authors choose per task whether to plan with the mined temporal cost or with latent L2, and that choice is made from the same locked results used to claim superiority; if the cost had to be fixed in advance, the advantage on Push-T and OGB-Cube would disappear, since pure d_psi yields 69.0% and 77.0% there.","fun_headline_variants_meta":{"raw":{"variants":["Mined temporal cost tops Euclidean planning on all tasks","TD-JEPA: plan with temporal distance, not embedding geometry","Offline logs teach JEPA to plan with temporal progress","Temporal cost mined from demos improves latent planning"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001222,"raw_usage":{"total_tokens":4911,"prompt_tokens":841,"completion_tokens":4070,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":585,"completion_tokens_details":{"reasoning_tokens":4004}},"tokens_in":585,"tokens_out":4070,"duration_ms":32798,"temperature":1.0,"reasoning_tokens":4004,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T02:43:03.218712+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Fix a single plan cost in advance for all four tasks—either always latent L2 or always d_psi—and re-run the locked evaluation; the paper's own Table 11 shows pure d_psi falling to 69.0% on Push-T and 77.0% on OGB-Cube, so any pre-registered single-cost protocol that reproduces those numbers would refute the 'matches or exceeds on every environment' claim. Alternatively, an independent check of Fig. 1's Spearman correlations (0.95 for d_psi versus 0.65 for L2 on held-out Push-T pairs) would test whether the mined cost truly ranks temporal separation as claimed.","supporting_citations":[],"review_version":1}