{"id":"30a3cc3b-dc33-4088-81e3-85d102f3e2e6","arxiv_id":"2607.17973","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"SAGE makes long-horizon latent world-model planning much more effective by generating latent subgoals and conditioning action proposals on them, raising 150-step success to 64.7% on PushT and 67.3% on OGBench Cube.","lead":"Peking University researchers built two small networks: one predicts a reachable intermediate goal inside a world model's latent space, the other proposes actions aimed at that goal. Guiding a frozen world-model planner this way raised 150-step success from 12.7% to 64.7% on PushT and from 26.7% to 67.3% on OGBench Cube.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The paper asserts but never measures that Eq. (4) subgoals are reachable under the frozen LeWM dynamics; if they are not, the Eq. (3) objective has no feasible minimum and the H=150 gains may not transfer.","rationale":"The reader's weakest_assumption identifies exactly the point I find most load-bearing: the paper never measures whether the subgoals generated by Eq. (4) are reachable under the frozen LeWM dynamics. This matters because the planning cost Eq. (3) evaluates candidate action sequences by their distance to the generated subgoal; if that subgoal is outside the set of latents LeWM can actually reach, then CEM is minimizing a cost with no good feasible minimum, and the action prior must be doing the real work through imitation rather than through the stated subgoal-conditional proposal mechanism. The empirical gains on PushT and Cube are still consistent with the paper's broad claim, and the matched ablations, per-seed tables, and schedule study provide useful support. But the central mechanism is asserted rather than demonstrated: the subgoal generator is trained to regress expert future latents, which are assumed to be reachable by the frozen dynamics, and the action generator is trained on expert action segments, so the whole pipeline could be learning a privileged imitation policy that happens to work on these two benchmarks. A direct reachability measurement would settle this: compare the CEM-achieved cost for generated subgoals against that for expert future latents under the same planning budget. If they match, the concern is resolved; if not, the central claim must be weakened to 'SAGE improves planning on these benchmarks' without the 'reachable subgoal' explanation. This is a limitation, not a falsification, so I do not recommend changing the reader's CONDITIONAL verdict.","tokens_in":10037,"tokens_out":5641,"duration_ms":52663,"concrete_test":"On the released checkpoints (or a faithful re-implementation), sample 50 start-goal pairs from the three fixed held-out manifests at H=150. For each SAGE stage, run the same 300-candidate, 30-round CEM procedure and record the minimum cost d(F_{\\theta}(z_{t-k:t}, a), \\hat{z}_{t+\\tau}) achieved over the final elite set. Repeat the same computation with the expert future latent z_{t+\\tau} in place of \\hat{z}_{t+\\tau}. If the generated-subgoal minima are systematically and significantly worse than the expert-target minima (e.g., normalized cost more than 2x higher), the 'reachable subgoal' assumption fails and the central mechanism is not established. If they are comparable, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central mechanism relies on \\hat{z}_{t+\\tau} being reachable by the frozen LeWM dynamics, because CEM's cost in Eq. (3) directly minimizes d(F_{\\theta}(z_{t-k:t}, a), \\hat{z}_{t+\\tau}). Eq. (4) is trained with Eq. (6) to regress expert future latents, and Eq. (7) trains the action generator on expert segments. No experiment measures whether the predicted subgoals lie on the manifold that LeWM can actually roll out to. If generated subgoals are not reachable under the frozen dynamics, then the learned action prior may still imitate expert-like actions in the training distribution, but the planner's cost function has no feasible minimum and the reported long-horizon gains would depend on task-specific regularity rather than on the proposed subgoal decomposition. The absence of a reachability metric is especially load-bearing because the phrase 'reachable latent subgoal' appears in the abstract and Section 3.2, and because the claimed contribution is improving proposals to a frozen world model: if the targets are not reachable, the improvement is not explained by the stated mechanism.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes SAGE, a planning framework that augments a frozen latent world model (LeWM) with two learned modules: a multi-duration latent subgoal generator and a subgoal-conditioned action generator. At planning time, the subgoal generator predicts a local latent target at a requested horizon, and the action generator proposes action candidates conditioned on that target; the frozen LeWM then scores and refines these candidates via CEM. Experiments on PushT and OGBench Cube report large gains at long horizons under matched candidate budgets, CEM rounds, and the same frozen checkpoint, with ablations isolating the contributions of the subgoal target, the learned action prior, and CEM refinement.","tokens_in":10380,"tokens_out":4908,"duration_ms":47831,"significance":"If the empirical findings hold, the contribution is notable: it shows that the planning capability of a frozen latent world model can be substantially extended by lightweight, task-specific proposal shaping, without modifying the world model or its cost function. The evaluation is carefully controlled — identical queries, frozen LeWM, equal candidate budgets and CEM rounds, three fixed manifests, and per-seed results — and the ablations are informative. The main strength is the clean decomposition of the gain into target generation and action-prior conditioning. The central limitation is that the paper asserts but never measures the 'reachability' of its latent subgoals, which is load-bearing for the proposed mechanism.","major_comments":[{"comment":"The term 'reachable latent subgoal' is central to the claimed mechanism, yet no experiment measures whether the subgoals produced by Eq. (4) actually lie in the set of futures that the frozen LeWM dynamics F_θ can roll out to. The subgoal generator is trained to regress the expert future latent z_{t+τ} (Eq. (6)), and the planner's cost in Eq. (3) directly minimizes d(F_θ(z, a), \\hat{z}_{t+τ}). If \\hat{z}_{t+τ} is not reachable under F_θ, then (i) the Eq. (3) objective may have no feasible minimum, and (ii) the observed gains could stem from the action generator imitating expert action segments rather than from the stated subgoal decomposition. I request a quantitative reachability analysis: for held-out windows, report d(F_θ(z, a^*_{t:t+τ-1}), \\hat{z}_{t+τ}) versus d(F_θ(z, a^*), z_{t+τ}) and versus d(z_{t+τ}, \\hat{z}_{t+τ}), or an analogous measure of whether the predicted subgoal is in","section":"§3.2–3.3, Eq. (4)–(6)"}],"minor_comments":[{"comment":"The 'ordering effect' at H=75 and H=100 is based on a single manifest of n=50 and two schedules, with no error bars or statistical test. Since this finding is used to motivate future work on duration ordering, it should be framed as a diagnostic result or supported with additional seeds.","section":"Table 3"},{"comment":"The low-dimensional state x_t is used as an input to both modules but is never precisely defined for PushT and Cube (e.g., gripper pose, object pose). Please specify its content and dimensions.","section":"§4.1 / Appendix A.3"},{"comment":"The trajectory-level Gaussian mixture is described as having M=8 modes, but the covariance structure (diagonal vs full, shared vs per-mode) is not stated. This matters for how the prior explores the action space; please add a sentence in the architecture details.","section":"§3.3, Eq. (5)"},{"comment":"The PRISM baseline is said to use 'its released state-conditioned proposal head.' It should be stated explicitly whether this head was used off-the-shelf or retrained on the same demonstrations; a mismatch would weaken the comparison.","section":"§4.1"}],"recommendation":"major_revision","confidential_remarks":"The reachability concern is the only load-bearing issue. The empirical protocol is otherwise strong, and the per-seed results in Table 7 support the headline numbers. I would be willing to accept after the authors add a reachability measurement or explicitly soften the 'reachable' claim to 'predicted local target' and discuss the implications. The schedule study should also be downgraded in strength."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a solid empirical methods paper. SAGE adds two lightweight modules on top of a frozen LeWM — a subgoal generator that predicts a latent future at a chosen duration, and an action generator that produces action candidates conditioned on that subgoal. The headline result is real: on PushT and OGBench Cube, with the same frozen checkpoint, candidate budget, and CEM refinement as baselines, long-horizon success at H=150 jumps from 12.7% to 64.7% and 26.7% to 67.3%. The short-horizon performance doesn't collapse.\n\nWhat I like: the experimental discipline. Matched comparisons, three fixed seeds per horizon, per-seed tables (Table 7) showing consistent gains, and ablations that separate the contribution of the subgoal target (Generator-only) from the learned action prior (SAGE vs generator-only) and from the CEM refinement (prior-top vs full SAGE). The scheduling study is also thoughtful; it shows duration ordering matters, which is a nice diagnostic.\n\nThe main soft spot is the same one the stress-test flags: the paper calls the subgoals 'reachable' but never measures whether they lie on the manifold the frozen dynamics can actually roll out to. That is a legitimate gap, and a reviewer should ask for it. But I don't think it is fatal. The Generator-only ablation uses the generated subgoal as the CEM cost target with a generic Gaussian proposal, and it already lifts H=150 PushT from 12.7% to 58.7%. If the targets were wildly unreachable, you would not expect that. So the targets are at least informative, even if not exactly reachable. The paper would be stronger with an explicit reachability metric — e.g., comparing the distance from predicted subgoals to the best actual rollout under the option prior — but the absence of that metric is a limitation, not a reason to reject.\n\nOther softer spots: only two benchmarks, no code or weights released, and the duration schedules are hand-picked, which means a bit of tuning went into the main numbers. But the schedules are shared across benchmarks and the paper reports schedule sensitivity.\n\nBottom line: this is for anyone working on latent planning who wants a clean demonstration of how to get more out of a frozen world model without retraining it. It deserves a serious referee. I'd send it out, with a request for a reachability experiment and code release.","headline":"A well-controlled empirical paper showing learned latent subgoals and subgoal-conditioned action proposals can push frozen-world-model planning much further, with subgoal reachability as the main unmeasured piece.","tokens_in":10827,"tokens_out":6129,"would_cite":true,"duration_ms":48705,"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":"Coupling predicted latent subgoals with conditioned action proposals sharply improves long-horizon planning in frozen latent world models.","keywords":["latent world models","subgoal generation","action proposal priors","long-horizon planning","model-based planning","cross-entropy refinement","manipulation benchmarks","offline goal-conditioned control"],"falsifier":"Evaluate the method on a held-out task where the frozen world model's dynamics are systematically miscalibrated for the expert target latents (e.g., by shifting goal semantics slightly) and compare SAGE against random proposals: if both collapse together, the bottleneck was not proposal quality but reachability of the learned subgoals. Concretely, measure the distance between SAGE's predicted subgoals and the nearest latent that the frozen dynamics can generate within the action budget.","tokens_in":9948,"feed_emoji":"🎯","tokens_out":3354,"duration_ms":26888,"temperature":0.7,"pith_summary":"This paper argues that long-horizon planning in latent world models fails mainly because random action proposals rarely expose the model to goal-directed futures, and that this gap can be closed without retraining the world model. It introduces two lightweight modules: a goal-conditioned generator that predicts a reachable latent subgoal a short duration ahead, and an action generator that samples candidate action sequences conditioned on that subgoal. The frozen world model then evaluates and refines these candidates. Across two manipulation benchmarks, the combined planner lifts success at a 150-step horizon from 12.7% to 64.7% on a planar pushing task and from 26.7% to 67.3% on a cube placement task.","feed_headline":"Subgoal priors lift long-horizon planning success from 12.7% to 64.7%","feed_subtitle":"Two small learned modules give a frozen latent world model better action candidates, sharpening far-horizon control without retraining its d","key_machinery":"The latent subgoal generator is a four-layer Transformer decoder that maps history, low-dimensional state, far-goal latent, remaining offset, and requested duration to a residual added to the far-goal latent, producing a local target at horizon τ. The subgoal-conditioned action generator is a three-layer Transformer decoder with a trajectory-level Gaussian mixture output (eight modes) over action options of length τ. The frozen world model scores imagined futures against the generated subgoal, and elite refinement search selects among the scored candidates.","core_discovery":"The core claim is that proposal quality—not world-model accuracy—is the bottleneck for latent planning at a distance. SAGE inserts a latent subgoal generator that produces a duration-matched local target (a residual added to the goal latent), and a subgoal-conditioned trajectory mixture from which candidate action sequences are drawn. With the same frozen encoder, dynamics, and search budget, these two modules raise success substantially as the goal recedes, while preserving performance on short-horizon queries.","pith_inferences":["If the subgoal generator's predictions are not reachable under the frozen dynamics, the cost function has no feasible optimum; measuring the distribution of generated subgoals against the set the world model can actually roll out would test this directly.","The method suggests that other bottlenecked planners could be improved by learned proposal-conditioning modules targeting intermediate states, provided those states are defined in the space the planner optimizes.","A schedule-ordering result hints at a learnable duration policy: at fixed total computation, shorter early commitments outperformed longer ones, and reversed schedules changed outcomes, so choosing durations per state may yield further gains.","The two-module design opens a testable extension: applying the subgoal-conditioned proposal to other frozen planners should transfer gains if the mechanism is generic."],"forward_implications":["Long-horizon planning gains can come from the proposal distribution alone, leaving the predictive world model untouched.","A single trained subgoal/option pair supports multiple planning durations, so temporal abstraction can be chosen online without retraining.","Generated local targets improve search even when the action distribution is standard Gaussian, showing the target itself carries much of the benefit.","The world model's scoring and refinement remains essential: executing the best learned option directly without refinement performs much worse, so learned proposals complement, not replace, model-based search.","Across two benchmarks the pattern holds consistently: gains grow with goal offset, with short-horizon performance maintained or improved."],"fun_headline_variants":["Subgoal priors lift long-horizon planning to 64.7% success","Latent subgoals sharpen action search for distant goals","SAGE: subgoal-conditioned actions beat random proposal planning","Frozen world model plans farther with subgoal-guided actions"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The subgoals predicted by the generator are assumed to lie on the manifold of latent states the frozen world model can actually reach, but reachability is never measured; if that fails on a new task, planning has no feasible target and the gains would disappear.","fun_headline_variants_meta":{"raw":{"variants":["Subgoal priors lift long-horizon planning to 64.7% success","Latent subgoals sharpen action search for distant goals","SAGE: subgoal-conditioned actions beat random proposal planning","Frozen world model plans farther with subgoal-guided actions"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000226,"raw_usage":{"total_tokens":1307,"prompt_tokens":747,"completion_tokens":560,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":491,"completion_tokens_details":{"reasoning_tokens":486}},"tokens_in":491,"tokens_out":560,"duration_ms":4826,"temperature":1.0,"reasoning_tokens":486,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T16:29:27.964356+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Evaluate the method on a held-out task where the frozen world model's dynamics are systematically miscalibrated for the expert target latents (e.g., by shifting goal semantics slightly) and compare SAGE against random proposals: if both collapse together, the bottleneck was not proposal quality but reachability of the learned subgoals. Concretely, measure the distance between SAGE's predicted subgoals and the nearest latent that the frozen dynamics can generate within the action budget.","supporting_citations":[],"review_version":1}