{"id":"536b7cc9-4283-40dd-9c90-a1b480bc5b55","arxiv_id":"2608.12959","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"low","formal_verification":"none","parameter_count":2,"one_line_summary":"Squared latent distance, the objective a latent-world-model planner minimizes, saturates and then inverts with true distance; replacing it with a reachability cost fixes long-horizon planning without retraining.","lead":"This paper shows that long-horizon planning failures in a latent world model can come from the planner's scoring function, not from a bad predictor. Swapping the objective alone lifts goal-reaching from 26% to 98% on a hundred-step task and reveals that reachability, not accuracy, is what a planner needs.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Objective-inversion claim uses pooled pairwise real-frame distances, not the fixed-goal imagined-latent surface CEM actually scores; the mechanism may not hold.","rationale":"The central claim is that the planning objective, not the predictor, is the binding constraint, and the paper's most direct causal evidence is the objective-replacement experiment: with frozen weights, changing only the cost lifts offset-100 success from 26.0% to 98.0%. That intervention is strong and survives the concern below. The mechanism offered for why the baseline objective fails is the saturation and inversion of squared latent distance. This mechanism is load-bearing because it appears in the abstract, the title, and the explanation of the overshooting failures. The current support for it is a pooled pairwise measurement over encoded real frames, which does not directly describe the goal-conditioned cost surface over imagined latents that CEM actually scores. The paper itself demonstrates in §7.4 that imagined latents can differ substantially from real encodings, so the real-frame landscape may not transfer. A concrete goal-conditioned test on imagined endpoints would settle whether the inversion is a genuine property of the planner's objective or an artifact of pooling. The reader's identified weakness (rollout test under recorded rather than planner-chosen actions) is related but less central: the objective-replacement result already shows the predictor can support planning under a different objective, whereas the metric concern targets the explanation of the original failure. Because the concern is addressable and the main empirical claim is otherwise well-supported, the verdict should remain conditional.","tokens_in":7022,"tokens_out":11813,"duration_ms":125862,"concrete_test":"Re-run the §4 measurement in a goal-conditioned, imagined-latent way. Use the 50 offset-100 episodes from §8. For each episode, roll the frozen predictor under (i) the actual CEM-selected action sequence, (ii) straight-line actions toward the goal, and (iii) straight-line actions away from it, recording imagined endpoints at 5, 10, 15, 20, and 25 environment steps. For each endpoint compute squared latent distance to the goal embedding and true distance to the goal. Check, per episode and pooled, whether cost decreases with true distance beyond 120 units in the relevant range. If the cost is monotone increasing for most episodes, the inversion claim fails; if it saturates or inverts, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The diagnosis in §4 is that squared latent distance saturates and then inverts, so CEM can be led away from the goal. But the evidence is a marginal correlation over 7,140 pairs of encoded real frames, pooled over all goal positions. CEM optimizes ||z_hat_T - z_goal||^2 for a fixed goal, scoring imagined latents from the predictor, not encoded real frames. A pooled mean can saturate even when every fixed-goal cost curve is monotone, if embedding scale varies by location; and §7.4 shows imagined embeddings can drift far from real encodings (authors' checkpoint: +74% degradation for a learned head). Thus the specific claim 'moving away from the goal can lower the cost' is not established for the distribution the planner actually evaluates. The objective-replacement experiment still shows the objective can be repaired, but the paper's mechanism for why the baseline fails would be unsupported if the goal-conditioned imagined cost is monotone.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper investigates why a latent world model planner fails at long horizons on the TwoRoom environment. Using a reproduced checkpoint of LeWorldModel and the original authors' released weights, the author reports that the predictor's imagined states remain informative for at least 75 environment steps while the planner only looks 25 steps ahead; that the squared latent distance used as the CEM planning objective saturates and then inverts with true distance; that long-horizon planning success across four checkpoints rank-orders exactly with this metric's quality and inversely with one-step prediction error; and that replacing only the planning objective with a decoded-position cost or a learned temporal-distance cost raises offset-100 goal success from 26% to 88--98% on the reproduced checkpoint and from 14% to 70% on the authors' checkpoint. The paper concludes that the binding constraint is the planner's objective, not the predictor, and that the objective must measure reachability rather than Euclidean proximity.","tokens_in":7218,"tokens_out":4288,"duration_ms":46895,"significance":"If the central claim holds, this is a valuable result for the latent world model literature: it provides a concrete dissociation between prediction quality and planning success, a measurable property of the embedding geometry that predicts that dissociation, and a cheap, retraining-free repair. Strengths include paired planning comparisons that reproduce all published baselines exactly, exact McNemar tests on discordant pairs, multiple independent measurements (rollout errors, 7,140 distance pairs, probe recovery, wall-cost ratios), and unusually honest limitation statements covering the single environment, single-seed checkpoints, and confounded training conditions. The main open question is whether the saturation/inversion mechanism is established for the distribution the planner actually scores; this is a load-bearing point that needs additional direct measurements.","major_comments":[{"comment":"The evidence that squared latent distance 'saturates and then inverts' is a pooled marginal correlation over 7,140 pairs of encoded real frames. The planner, however, minimizes ||z_hat_T - z_goal||^2 for a fixed goal image, scoring imagined latents produced by the predictor. A pooled mean can saturate and even invert when every fixed-goal cost curve is monotone, if the scale of the embedding varies by location. Moreover, §7.4 shows that on at least one checkpoint imagined embeddings drift substantially from encoded real frames (+74% MAE degradation for a learned head), so the real-frame distribution is not automatically the distribution the planner scores. The paper's central mechanism — that 'moving away from the goal can lower the planner's cost' — therefore needs direct evidence: for a fixed goal, the cost should be plotted as a function of true distance for that goal, and the same should be repeated on imagined latents from the predictor under the planner's action distribution. Without this, the objective-replacement experiment remains valid, but the proposed explanation of why the baseline planner fails is not yet established.","section":"§4 and Table 1"},{"comment":"The rollout in §3 evaluates prediction error under recorded actions from validation clips, whereas CEM evaluates imagined action sequences that differ from recorded ones. The paper's claim that 'the predictor is not the bottleneck' is partially supported indirectly by the repair experiment, which uses the same predictor and succeeds at high rates, but that experiment changes the objective and therefore does not isolate prediction error under planner-chosen actions. To make the bottleneck claim airtight, the author should measure imagined-latent error for the action sequences CEM actually samples, for example by rolling the predictor out on CEM's elite candidates and comparing the imagined latents with the encoder outputs on the states actually reached. This would close the gap between the rollout statistic and the planner's operating distribution.","section":"§3 and §8"}],"minor_comments":[{"comment":"The abstract contains a typo ('decreasesbeyond' should be 'decreases beyond'), and the inline file references such as 'Fromfollowup/...' are missing spaces; these should be cleaned up.","section":"Abstract and §4"},{"comment":"The table columns 'monotone 80--300 spread' and 'spread' should define the exact criterion for monotonicity and how the percentage spread is computed, since these terms carry the weight of the inversion claim.","section":"Table 1 and §5.2"},{"comment":"The two protocols ('offset 25, budget 50' and 'offset 100, budget 150') should be defined explicitly in terms of environment steps and frameskip, because the relationship between planning horizon, offset, and budget is central to interpreting the results.","section":"§2"},{"comment":"The wall-reachability comparison is reported only for real-frame pairs; a sentence clarifying that the same ratio was not measured on imagined pairs would help readers calibrate how much of the planner's actual scoring distribution this analysis covers.","section":"§7.3"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is a careful, reproducible empirical study with an unusually transparent limitations section. The main risk is that the headline mechanism (objective saturation/inversion) is inferred from pooled real-frame distances rather than from the fixed-goal imagined-latent cost that CEM actually minimizes. The author should be encouraged to supply the fixed-goal and imagined-latent measurements; the paper's scope and framing are otherwise well suited to the journal."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here is my honest read. The paper does one concrete, useful thing: on the released LeWorldModel checkpoint, it swaps the planning objective with every weight frozen and lifts long-horizon success from 26% to 98%, with exact McNemar tests and baselines reproduced to the digit. The temporal-distance cost that uses no state supervision and charges 24% more for crossing the wall is a nice result, and the failure on the authors' checkpoint—a head fit on encoded frames evaluated on imagined ones—is a real, transferable lesson. This is a measurement paper, not a position paper, and most of the measurements are solid.\n\nThe soft spot is the mechanism in §4. The claim that \"moving away from the goal can lower the cost\" is inferred from a pooled correlation over 7,140 pairs of encoded real frames across all goal positions. But CEM evaluates a fixed-goal cost over imagined latents. The stress-test note is right: a pooled mean can saturate even if every fixed-goal curve is monotone, and the paper's own §7.4 shows imagined embeddings can drift far from real encodings. So the specific non-monotonicity offered as the explanation of baseline failure is not directly measured on the distribution the planner actually scores. The repair experiment still stands on its own, but the mechanism needs a direct plot of cost versus true distance for fixed goals under imagined latents.\n\nOther concerns are minor and mostly self-acknowledged: one seed per checkpoint, confounded training conditions, correlations without confidence intervals. These do not dent the central swap experiment.\n\nWho is this for? Anyone building or evaluating latent world models, and people working on planner-aware metrics. It deserves a serious referee, but the referee should push for the fixed-goal imagined-latent measurement. As it stands, I would not cite the inversion claim as established; I would cite the objective-swap result.","headline":"The objective-swap experiment is real and valuable, but the saturation/inversion mechanism that explains the baseline failure is inferred from pooled real-frame pairs, not from the fixed-goal imagined-latent cost CEM actually scores.","tokens_in":7694,"tokens_out":2103,"would_cite":true,"duration_ms":23505,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The planner's objective, not the world model, is what breaks long-horizon planning","keywords":["latent world models","model-based planning","cross-entropy method","planning objective","reachability metric","squared latent distance","TwoRoom environment","reproduction study"],"falsifier":"Run the same cross-entropy planner with the same squared-latent-distance objective but a perfect predictor — for example, feed the encoder the true future frames instead of imagined latents — at goal offset 100. If success stays near baseline rather than rising, the objective is not the binding constraint; if it rises, the predictor's imagined rollouts were the limiting factor.","tokens_in":1437,"feed_emoji":"🎯","tokens_out":1913,"duration_ms":47942,"temperature":0.7,"pith_summary":"The paper argues that long-horizon planning failure in a reproduced latent world model is caused by the planner's objective, not by the predictor. The squared Euclidean distance between imagined and goal latents, which cross-entropy planning minimises, stops rising around eighty arena units and falls beyond about a hundred and twenty, so moving away from the goal can look like progress. The predictor is not the bottleneck: its imagined state seventy-five steps ahead is still only about a fifth as wrong as assuming the world froze, and a linear probe recovers position from the frozen embedding at $R^2 = 0.9922$. Replacing only the objective, with no retraining, lifts goals reached at offset 100 from 26.0% to 98.0%, making planning horizon-independent. The sharper claim is that what a planning cost must capture is reachability, not spatial proximity.","feed_headline":"Swap the planner's cost; long-horizon goals jump 26% to 98%","feed_subtitle":"The predictor still sees 75 steps ahead; the squared-distance cost inverts past 120 units. Fixing only the objective restores success.","key_machinery":"The central object is the planning cost: squared Euclidean distance between the imagined final latent and the goal latent, $\\| \\hat{z}_T - z_{\\mathrm{goal}} \\|_2^2$, which cross-entropy-method planning minimises. The paper shows this metric saturates and then inverts with true distance, and contrasts it with a learned temporal-distance head — a small MLP trained on pairs of real frames supervised only by their separation in time — which orders distance monotonically and charges more for crossing the environment's wall. The probe that decodes positions from the frozen embedding is the instrument that shows the information is present; the temporal head is the instrument that shows the right objective is reachability, not proximity.","core_discovery":"The central discovery is that the objective the planner minimises — squared latent distance — is not monotone in true distance: it saturates by roughly eighty units and inverts past roughly a hundred and twenty, so a cross-entropy planner can be led away from a distant goal while scoring itself as improving. This is shown to be a property of the method as released, present in the original authors' weights and across four checkpoints, where long-horizon planning success rank-orders exactly with metric quality and inversely with one-step prediction accuracy. The information needed for planning is present in the representation — a ridge probe decodes position at $R^2 = 0.9922$ — but the $\\ell^2$ cost discards it. Replacing the objective with a head trained on frame separation (reachability) raises offset-100 goals from 26.0% to 98.0% without retraining the model, and the three objectives rank by how well they charge for the wall between rooms, not by how well they predict spatial distance.","pith_inferences":["If this generalises beyond TwoRoom, the practical prescription is a cheap pre-deployment test: measure whether the planner's cost is monotone in true reachability before attributing long-horizon failure to the predictor.","The saturation-and-inversion pattern suggests a possible selection pressure in representation learning: objectives trained purely on next-step prediction may compress far-apart states onto similar latents, which would make many latent-planning methods fail for the same reason.","The temporal-distance head is a prototype of a self-supervised reachability cost; one could extend it to continuous control by training it on action-conditioned rollouts rather than recorded frames, which would also address the imagined-versus-real distribution shift the paper exposes."],"forward_implications":["If the objective is the bottleneck, long-horizon planning can be repaired without retraining the world model, by changing only the cost the planner optimises.","Selection of world models by one-step prediction loss can select against the embedding geometry that planning depends on; metric quality under the planner's own objective should be measured instead.","A learned planning cost must be trained on the distribution the planner scores — imagined embeddings — or it can fail on checkpoints where the predictor drifts off the encoding manifold.","The horizon dependence of planning success (94.0% at 25 steps versus 26.0% at 100) dissolves once the objective is fixed: under the temporal cost success is 98.0% at both offsets.","Reachability, not Euclidean proximity, is what a planning objective must encode; a cost that predicts spatial distance worse can plan better if it charges for barriers."],"supporting_citations":[{"why":"Supplies the reproduction, the released checkpoints, baseline planning figures (94.0%, 26.0%, 14.0%), and the accuracy/planning dissociation whose mechanism this paper identifies.","marker":"arXiv:2608.10145"},{"why":"Defines the LeWorldModel joint-embedding predictive architecture and provides the authors' released weights used to show the pathology belongs to the method, not to one reimplementation.","marker":"arXiv:2603.19312"},{"why":"States the general argument that Euclidean proximity in latent space is a poor proxy for reachability, which this paper measures directly on a published model.","marker":"arXiv:2605.22164"}],"fun_headline_variants":["The bottleneck is the planner's cost, not the prediction","Swap the lost objective: 26% to 98% without retraining","Squared latent distance misleads planners—use reachability instead","Long-horizon planning fixed by changing the cost, not the model"],"cache_read_input_tokens":9984,"weakest_assumption_plain":"The conclusion that the predictor is not the bottleneck assumes prediction error under the actions the planner actually chooses is no worse than under recorded actions from validation clips, and the paper does not measure error on planner-chosen rollouts directly.","fun_headline_variants_meta":{"raw":{"variants":["The bottleneck is the planner's cost, not the prediction","Swap the lost objective: 26% to 98% without retraining","Squared latent distance misleads planners—use reachability instead","Long-horizon planning fixed by changing the cost, not the model"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000295,"raw_usage":{"total_tokens":1795,"prompt_tokens":1105,"completion_tokens":690,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":721,"completion_tokens_details":{"reasoning_tokens":614}},"tokens_in":721,"tokens_out":690,"duration_ms":6941,"temperature":1.0,"reasoning_tokens":614,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T19:42:25.499538+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same cross-entropy planner with the same squared-latent-distance objective but a perfect predictor — for example, feed the encoder the true future frames instead of imagined latents — at goal offset 100. If success stays near baseline rather than rising, the objective is not the binding constraint; if it rises, the predictor's imagined rollouts were the limiting factor.","supporting_citations":[],"review_version":1}