{"id":"4361cabb-61f1-470c-8b94-f9c2febee064","arxiv_id":"2608.01926","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":9,"one_line_summary":"ProWorld adds goal-conditioned progress order to a JEPA world model via hyperbolic entailment cones and progress-aware planning, improving average success rate by 9.67% over LeWM on four visual goal-reaching tasks.","lead":"ProWorld is a visual world model that arranges imagined future states in a curved, progress-aware geometry and achieves higher success rates than a strong JEPA baseline on four long-horizon robot tasks. This matters because long-horizon goal reaching is where next-step predictors drift, and the paper offers a cheap geometric fix.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Success metric credits any termination within budget, and ProWorld's C_best planning term directly favors transient goal visits, so the 9.67-point SR gain may be partly an artifact of the evaluation criterion.","rationale":"The reader's weakest_assumption focuses on temporal order as a weak label for progress, which is a training-signal robustness issue acknowledged in the Limitations and supported by ablations. However, the central claim is the empirical SR gain, and the most direct threat to that claim is the success metric's interaction with ProWorld's planning cost. The appendix defines success by any termination within the budget, and ProWorld's C_best term is designed to minimize the best intermediate distance to the goal. The ablation in Table 6 shows this term accounts for 14 points on Cube-S, and Fig. 13 confirms that success can be triggered and then lost. This makes the comparison with terminal-only planners like LeWM structurally unfair. The temporal-order concern, while valid, is about generalization to detour-rich tasks and does not directly undermine the reported numbers on the evaluated benchmarks. A re-evaluation with a final-state success criterion would settle whether the gain is genuine; until then, the verdict remains CONDITIONAL.","tokens_in":27969,"tokens_out":6880,"duration_ms":79686,"concrete_test":"Recompute Table 1 with success redefined as the final state after the full evaluation budget satisfying the environment's goal condition (i.e., the terminal state must be goal-reaching). For Cube-S, also evaluate a version where the episode stops at the first termination and success requires that termination to be the terminal state. Compare the ProWorld-minus-LeWM success-rate gap under these stricter criteria; if the gap shrinks by more than 3 points, the headline gain is partly an artifact of the transient-success metric.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The reported 9.67-point average SR gain over LeWM (Table 1) rests on a success metric that counts an episode as successful if the environment's termination flag fires at any step within the budget (Appendix Eq. 2), regardless of the final state. This interacts directly with ProWorld's planning objective: Eq. 23 includes C_best, the minimum hyperbolic goal distance along the rollout, which selects action sequences that pass near the goal even if they do not terminate there. Table 6's ablation shows that removing C_best drops Cube-S SR from 78% to 64%, and the Fig. 13 caption explicitly acknowledges 'post-success drift' where the cube reaches the target pose and is later released. A terminal-distance-only planner like LeWM is systematically disadvantaged under this transient-success criterion. The paper's own ablation and qualitative note expose a mechanistic link between the method's design and the metric, making the headline gain potentially inflated.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"ProWorld proposes a JEPA-style visual world model that embeds latents in Lorentz hyperbolic space and organizes them by a goal-conditioned progress order. Training combines hyperbolic prediction, contrastive future discrimination, and a goal-anchored adaptive entailment cone; planning uses a CEM over rollouts scored by terminal, best-intermediate, and mean hyperbolic goal distances. The paper reports an average absolute success-rate gain of 9.67 over LeWM on PushT, Cube-S, AntMaze-L, and Scene, supported by Cube-S ablations, sensitivity analyses, and geometric diagnostics.","tokens_in":28294,"tokens_out":5997,"duration_ms":66401,"significance":"If the reported gains survive a standard success criterion, ProWorld would be a solid empirical contribution: it adds goal-progress structure to latent world models with negligible overhead (+0.83% parameters), and the ablations are internally consistent. The explicit formalization of progress order, the adaptive cone design, and the decomposed planning cost are useful ideas. However, the headline claim is currently confounded by the transient-success evaluation criterion and by the absence of significance testing, so the central claim is not yet conclusively established.","major_comments":[{"comment":"The success metric counts an episode as successful if the environment termination flag fires at any step within the evaluation budget, regardless of the final state. The planning cost Eq. (23) includes C_best, the minimum goal distance along the rollout, and Table 6 shows that removing C_best lowers Cube-S SR from 78% to 64%. Fig. 13 also acknowledges post-success drift. A terminal-distance-only planner such as LeWM is systematically disadvantaged by this transient-success criterion, because it is never credited for passing near the goal. Please re-evaluate with a final-state success criterion (or stopping at first termination) and report whether the 9.67-point gain over LeWM and the C_best benefit persist.","section":"Appendix, Eq. (2); Eq. (23); Table 6; Fig. 13"},{"comment":"No significance testing is reported for any comparison. With three seeds, the Scene result (69.33±0.94) versus EB-JEPA (67.33±0.94) has overlapping standard deviations, and the Table 4 MSE/Pearson diagnostics show no meaningful difference from LeWM. The paper's claim of consistent gains needs paired tests or confidence intervals across seeds, at least for the headline LeWM comparison and for the C_best ablation.","section":"Table 1; Table 4"},{"comment":"The central training signal assumes that temporal separation in a hindsight trajectory implies progress toward the goal. The Limitations section itself concedes that this is coarse for backtracking, detours, or sub-goal switching. Since Cube-S and Scene are play-style datasets collected by non-goal-directed policies, it would strengthen the paper to quantify the fraction of training pairs that actually reduce physical or hyperbolic goal distance, or to test sensitivity to pair-level label noise. Without this, the learned progress structure is plausible but not directly evidenced.","section":"Definition 1, Eq. (19), Limitations"}],"minor_comments":[{"comment":"Equation numbers restart in the appendix (e.g., Eq. (1) appears both in the main text and in the appendix). Use distinct numbering or labels to avoid confusion.","section":"Appendix/Notation"},{"comment":"Cube-S ProWorld is reported as 77.33±1.15 in Table 1 but 78% in the ablation tables. Clarify whether ablations use a single seed, a rounded mean, or a different evaluation configuration.","section":"Table 1 vs Table 2"},{"comment":"The caption says the last row is relative improvement over the second-best baseline, but for AntMaze-L the second-best is GCIQL (23.33), not LeWM. A footnote should make this explicit to avoid misreading.","section":"Table 1, 'Rel. Gain' row"},{"comment":"The variable 'Progress' in the correlation matrices is not defined in the main text or appendix. Specify how progress is computed (temporal index, goal-distance reduction, or manual labeling).","section":"Fig. 7 and Fig. 11"},{"comment":"The statement that an episode is counted as successful 'even if the state subsequently deviates from the goal' conflicts with standard termination semantics, where the episode ends at the termination flag. Clarify whether evaluation actually continues after termination and how this interacts with the reported SR.","section":"Appendix, 'Success Criterion'"},{"comment":"'Average absolute success-rate gain' should be phrased as 'average absolute percentage-point gain' to avoid ambiguity with relative gains.","section":"Abstract"}],"recommendation":"major_revision","confidential_remarks":"The transient-success issue is the main risk to the paper: if the re-evaluation with a final-state success criterion shows that the C_best term no longer helps, the headline claim and the method's motivation would be substantially weakened. I would ask the authors to make the success criterion unambiguous and to report both transient and final-state SR before considering acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: ProWorld is a genuine, well-motivated extension of GeoWorld-style hyperbolic world models, and the ablations are the best part of the paper. The goal-conditioned progress order, the adaptive goal-anchored cone, and the progress-aware planning cost are new. The authors show clear thinking, and the experiments include a reasonable baseline set, including a reproduction of GeoWorld under the unified setting. I believe the central claim as stated, however, is softer than the number suggests.\n\nThe stress-test note holds up on reading. Appendix Eq. 2 counts an episode as successful if the environment's termination flag fires at any step within budget, not only at the end. That is a defensible operationalization, but it interacts directly with Eq. 23's C_best term, which scores rollouts by the minimum hyperbolic goal distance along the trajectory. The Cube-S ablation (Table 6) shows that removing C_best drops SR from 78% to 64%, and Fig. 13's caption explicitly acknowledges post-success drift. So a large part of the apparent gain over a terminal-only planner like LeWM could be the metric rewarding a transient goal visit rather than sustained goal reaching. The Scene gain over EB-JEPA is 2 points with overlapping standard deviations, and no significance tests accompany any comparison. The headline 9.67 average is arithmetically correct, but it mixes one big AntMaze-L gain with smaller or statistically indistinguishable gains elsewhere.\n\nOther soft spots are minor. The theoretical motivation is qualitative, not proven, and the paper says so in the appendix. The sensitivity sweeps show large swings for delta, curvature, horizon, and loss weights, which makes me wonder about configuration search, though the main table does report three seeds. PushT is described as part of OGBench in the main text, which is factually wrong (it is from the DINO-WM setting). The limitations section honestly flags backtracking and subgoal-switching as weak spots for the progress-order supervision.\n\nWho this is for: anyone working on JEPA world models, especially for long-horizon planning. The method is cheap (0.83% params, 0.03% rollout FLOPs) and the geometric idea is plausible. It deserves a serious referee; the authors should be asked to re-evaluate with a terminal-state-only success metric, report significance tests, and either drop the transient-success criterion or explicitly justify it and separate out the C_best contribution. I would not desk-reject this.","headline":"A solid method paper whose headline 9.67-point SR gain is probably inflated by a transient-success metric that the planner's own best-intermediate cost directly exploits; worth refereeing.","tokens_in":28736,"tokens_out":1606,"would_cite":true,"duration_ms":18714,"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":"ProWorld claims that organizing visual latent dynamics by goal-conditioned progress order in hyperbolic space lets a JEPA-style world model reach distant visual goals substantially more reliably than local next-step prediction.","keywords":["world models","goal-conditioned planning","hyperbolic geometry","Lorentz model","latent dynamics","long-horizon planning","entailment cone","progress order"],"falsifier":"Run ProWorld on a task whose successful routes require temporarily moving away from the goal before approaching it (e.g., a maze with a detour). If temporal-order supervision mislabels progress, the learned entailment cones should point in the wrong direction on the detour segment, success rate should fall toward the no-cone ablation, and cone-loss violations should concentrate on that segment.","tokens_in":27882,"feed_emoji":"🎯","tokens_out":5515,"duration_ms":56221,"temperature":0.7,"pith_summary":"ProWorld sets out to show that a visual world model can plan long-horizon goal-reaching far better if its latent space is organized by how much each state advances the goal, not just by local next-step accuracy. The paper argues that JEPA-style models—which learn to predict the very next latent state—suffer from two long-horizon failures: multi-step rollouts drift, and locally similar future states can correspond to very different goal progress. To fix this, it introduces goal-conditioned progress order, an asymmetric, coarse-to-fine relation, and embeds states in hyperbolic (Lorentz) space so that early states occupy broad regions and later states concentrate near the goal. A progress-aware planning objective scores candidate action sequences by terminal, best, and mean goal distance. Across four visual goal-reaching tasks, the model reports an average absolute success-rate gain of 9.67 percentage points over a strong local-prediction baseline, with the biggest gains on the longest-horizon task.","feed_headline":"Progress-ordered world model beats local predictors by 9.67 points","feed_subtitle":"A latent space sorted by goal progress keeps long-horizon visual plans from drifting.","key_machinery":"Goal-conditioned progress order, realized through a goal-anchored adaptive entailment cone in the Lorentz model of hyperbolic space. The relation s_i ≺_g s_j says that state j occurs later than i in a goal-conditioned trajectory and has progress potential at least ε higher toward goal g; the paper uses temporally separated pairs as weak supervision. The cone constraint pushes each child state into an adaptive cone pointing from parent to goal, with aperture widening as the parent-goal distance grows, and adds a margin requiring the child to be closer to the goal than the parent. A Lorentz contrastive loss and a planning cost using terminal, best, and mean goal distances complete the mechanis","core_discovery":"The paper claims that long-horizon visual goal planning fails for JEPA-style world models trained only on local next-step consistency, because multi-step rollouts drift and locally similar futures can have different goal progress. It introduces goal-conditioned progress order—a relative ordering that is asymmetric and coarse-to-fine—and shows it can be realized geometrically by embedding states in the Lorentz model of hyperbolic space. Training then imposes three objectives: hyperbolic future prediction, contrastive discrimination of future states, and a goal-anchored entailment cone that keeps later states inside the parent-to-goal direction and closer to the goal. At inference, a progress-","pith_inferences":[],"forward_implications":["On long-horizon visual goal-reaching, models trained only on local transition consistency will drift; the largest reported gains occur on the longest-horizon task, AntMaze-L.","Treating temporal order in hindsight trajectories as weak progress supervision is sufficient to learn useful progress structure without success labels.","The progress-aware planning cost—terminal, best-intermediate, and mean goal distance—is necessary; terminal-only planning drops success from 78% to 62% in the Cube-S ablation.","The entailment cone relies on correct temporal order: reversing or randomizing parent-child pairs drops success from 78% to 44% and 52%, respectively.","Hyperbolic geometry contributes beyond Euclidean latent spaces: replacing the Lorentz model with Euclidean space drops success from 78% to 68%.","A testable extension: use the same progress-order supervision in model-based RL or video prediction, where 'progress' need not be goal-reaching but can be any trajectory-level completion signal.","The ablation order suggests future-state discriminability (contrastive loss) may matter more than hyperbolic geometry per se; a Euclidean contrastive-only variant would test whether the gain is geometric or discriminative.","Because the temporal offset Δ=2 is optimal and Δ=1 is worse, the method assumes a minimum separation to read progress; domains with dense, noisy transitions may need per-domain Δ tuning or learned progress estimation."],"fun_headline_variants":["Hyperbolic world model uses goal progress to beat drift","Goal-progress ordering in hyperbolic space lifts planning by 9.7","ProWorld: hyperbolic latent order for long-horizon goals","Beat local predictors: sort future states by goal progress","Progress-aware hyperbolic model prevents visual goal drift"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise is that temporal order within a goal-conditioned hindsight trajectory is a reliable weak label for goal progress—an assumption that fails for trajectories with detours, backtracking, or subgoal switches, as the paper's limitations section concedes.","fun_headline_variants_meta":{"raw":{"variants":["Hyperbolic world model uses goal progress to beat drift","Goal-progress ordering in hyperbolic space lifts planning by 9.7","ProWorld: hyperbolic latent order for long-horizon goals","Beat local predictors: sort future states by goal progress","Progress-aware hyperbolic model prevents visual goal drift"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000167,"raw_usage":{"total_tokens":1115,"prompt_tokens":789,"completion_tokens":326,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":533,"completion_tokens_details":{"reasoning_tokens":246}},"tokens_in":533,"tokens_out":326,"duration_ms":3674,"temperature":1.0,"reasoning_tokens":246,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T18:05:53.640859+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run ProWorld on a task whose successful routes require temporarily moving away from the goal before approaching it (e.g., a maze with a detour). If temporal-order supervision mislabels progress, the learned entailment cones should point in the wrong direction on the detour segment, success rate should fall toward the no-cone ablation, and cone-loss violations should concentrate on that segment.","supporting_citations":[],"review_version":1}