{"id":"f75dbe2c-dc29-475b-acce-1a9d8ae0666e","arxiv_id":"2602.23148","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Goal-conditioned transition models with WL graph embeddings and symbolic successor decoding beat action-sequence transformers on extrapolation in Blocksworld, VisitAll, and Gripper, using far smaller models.","lead":"This paper trains compact models to predict the next world state in a planning problem, then uses symbolic search to turn those predictions into valid plans. It reports better out-of-distribution generalization than large transformer planners on several benchmark domains, using far fewer parameters and training examples.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"OOD gains may stem from Algorithm 1's symbolic successor filtering, not from the learned transition model; action baselines lack this scaffolding.","rationale":"I read the paper in good faith. The strongest claim is that learning explicit transition models yields higher OOD satisficing-plan success than direct action-sequence prediction with fewer data and parameters. For this to hold, the advantage must come from the transition-model objective (predicting successor embeddings) rather than from the neuro-symbolic decoding procedure. The paper's Algorithm 1 intertwines the two: at every step it uses the true symbolic state, enumerates all valid successors from the ground-truth γ, and selects the nearest one to the predicted embedding. This guarantees validity and corrects errors online; no action-centric baseline is given any comparable validity filter or re-ranking. The reader flagged exactly this as the weakest assumption, and I agree it is the load-bearing point.\n\nThe internal inconsistencies (WL dimensions differ between Table 3 and Appendix E.4; beam width 3 in inference but greedy argmin in Algorithm 1) are real but secondary; they affect reproducibility, not the core attribution. The unfair Plansformer cross-domain baseline also weakens the action-centric comparison but does not by itself explain the Blocksworld/VisitAll gains. The decisive issue is the missing control: an action-centric model given the same successor enumeration.\n\nThe concrete test I propose directly addresses this. If an action-ranking baseline with identical scaffolding performs comparably, the paper's central claim is over-stated and the verdict should require re-scoping (e.g., claim a neuro-symbolic ranking advantage, not transition-model learning). If it fails, the state-prediction objective is genuinely contributing, and the conditional accept stands. Either way, the experiment is cheap and decisive.","tokens_in":18196,"tokens_out":4815,"duration_ms":47826,"concrete_test":"Add a controlled action-ranking baseline: at each step, enumerate all applicable actions (identical symbolic scaffolding to Algorithm 1), score each action with an action-centric model (e.g., SymT or PlanGPT logits, or a randomly initialized scorer), execute the highest-scoring valid action, repeat until goal or horizon. Evaluate on the same four domains and splits. If this baseline reaches or approaches WL-XGB delta performance (e.g., Blocksworld extrapolation ≥0.30), then the successor-enumeration scaffolding, not the state-prediction objective, drives the gains, and the central claim needs re-scoping. Additionally, report fθ's raw top-1 accuracy over all states rather than only Succ(st); low raw accuracy with high filtered success would confirm the filter is the source of validity.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—that explicit transition-model learning beats action-sequence prediction for OOD generalization—is not established because the comparison is structurally unfair. In Algorithm 1 (lines 4–6), the method maintains the true symbolic state st, enumerates every valid successor Succ(st) via γ, and only ranks them by embedding distance. The model fθ never has to produce an executable action or even a state outside Succ(st); at worst it must make the desired successor's embedding closer to vt than alternatives. The action-centric baselines (Plansformer, PlanGPT, SymT) receive no such per-step validity enumeration or re-ranking; they must generate action sequences that are valid on their own. Thus the reported gains (e.g., WL-XGB delta 0.50 vs SymT 0.13 in Blocksworld extrapolation) may reflect the neuro-symbolic decoding scaffolding rather than a learned approximation of γ. Supporting evidence: the model achieves 0.00 in Logistics extrapolation, suggesting it is not truly approximating γ under size shift; Appendix G.3 confirms the scaffolding explicitly ('Symbolic State Maintenance... guarantees that every generated action is applicable'). Without a control that gives action-centric models the same scaffolding, the abstract's causal attribution is unsupported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a state-centric formulation of generalized planning: instead of predicting action sequences directly, a goal-conditioned model fθ maps WL graph embeddings of the current state and goal to a residual embedding delta, and plans are decoded by maintaining the true symbolic state, enumerating all valid successors Succ(s_t) via the exact transition function γ, and selecting the successor whose embedding is nearest to the predicted target (Algorithm 1). The authors evaluate LSTM and XGBoost transition models with WL and fixed-size factored representations on Blocksworld, Gripper, VisitAll, and Logistics, comparing against Plansformer, PlanGPT, and Symmetry-Aware Transformers. They report competitive or superior out-of-distribution extrapolation with far smaller models and fewer training instances, while acknowledging failure on Logistics.","tokens_in":18481,"tokens_out":6000,"duration_ms":65286,"significance":"If the central comparison were fair, the result would be practically significant: it would show that a compact, sample-efficient model can learn domain dynamics in embedding space and beat much larger action-sequence Transformers on size extrapolation. The paper is also well organized and unusually transparent about implementation details, with released code and appendices covering data generation, WL features, hyperparameters, and inference. However, the main comparison is confounded by the symbolic decoding scaffold, so the abstract's causal claim is not established. The significance is therefore conditional on additional controls and on evidence that fθ actually approximates γ.","major_comments":[{"comment":"The evaluation is structurally unfair to action-centric baselines. At every step the method maintains the true symbolic state s_t, enumerates all valid successors Succ(s_t) using the exact transition function γ, and only ranks them by embedding distance. It can never emit an inapplicable action or an invalid successor state, whereas Plansformer, PlanGPT, and SymT generate action sequences without such per-step validity filtering. Consequently the abstract's claim that 'learning explicit transition models yields higher out-of-distribution satisficing-plan success than direct action-sequence prediction' is not supported by Table 1: the gains (e.g., WL-XGB delta 0.50 vs. SymT 0.13 in Blocksworld extrapolation) may reflect the symbolic scaffolding rather than fθ approximating γ. A minimal control is to give action-centric baselines the same per-step validity enumeration/reranking, or to run","section":"Algorithm 1, lines 4–6; Appendix G.3"},{"comment":"The paper overstates that fθ is a learned approximation of γ. No experiment measures how often the predicted embedding φ(s_t)+fθ(...) is closer to the true successor embedding than to other valid successors, nor reports prediction accuracy on extrapolation instances. The Logistics result (all learned models 0.00 extrapolation) and the statement in Appendix G.3 that decoding 'guarantees that every generated action is applicable' and 'performs online correction of neural prediction errors' indicate that the symbolic enumeration is doing essential work. Please report transition-prediction metrics (e.g., rank of the true successor among Succ(s_t), top-1 accuracy, or embedding distance to the true successor versus the candidate distribution) on interpolation and extrapolation splits. If the learned model's ranking accuracy is near chance, the method should be described as a learned ranker ove","section":"Section 'State-Centric Transition-Model Learning'; Table 1 (Logistics rows)"},{"comment":"The data- and parameter-efficiency claims are also confounded by task difficulty. Ranking over a handful of γ-valid successors is an easier learning problem than open-ended action-sequence generation, so the 9-instance Blocksworld and 4-instance Gripper training sets do not by themselves demonstrate sample-efficient transition-model learning. The model-size ratios in Table 4 compare architectures solving different tasks; they are informative only after the scaffolding is controlled. At minimum, add (i) a state-centric variant without Succ(s_t) enumeration, or with noisy/partial successor sets, and (ii) an action-centric baseline with the same per-step validity enumeration, and report success rates and learning curves on equal footing.","section":"Appendix C.2; Table 4; sample-efficiency discussion"}],"minor_comments":[{"comment":"The main text defines the training loss as squared error over successor embeddings, but Appendix G.2 states that LSTM state-mode training uses cosine embedding loss and delta-mode uses MSE. Please reconcile.","section":"Equation (1) vs. Appendix G.2"},{"comment":"Algorithm 1 writes the selection as argmin over Euclidean distance, while Appendix G.2 and Table 7 specify cosine distance for state mode. The distance metric should be stated consistently.","section":"Algorithm 1 line 5 vs. Table 7"},{"comment":"The column labels 'SymTE*' and 'SymTED*' are not expanded in the table or its caption. Define them (presumably SymT encoder and SymT encoder-decoder variants).","section":"Table 1 caption"},{"comment":"Plansformer was not trained on Gripper, Logistics, or VisitAll; its 0.00 success in those domains is expected and should not be presented as evidence about action-centric sequence prediction without this caveat being prominent.","section":"Baselines paragraph"},{"comment":"The WL vocabulary is constructed from training graphs only. Out-of-vocabulary colors that appear in extrapolation instances are apparently dropped from the histogram. Discuss the potential impact of this on size extrapolation.","section":"Appendix E.4"},{"comment":"The paper reports Mean±Std over 'successive rollouts (seeds)' but does not state how many seeds. Specify the number of independent decoding runs and, ideally, report confidence intervals for the headline extrapolation comparisons.","section":"Table 1 and repeated rollouts"}],"recommendation":"major_revision","confidential_remarks":"The main issue is fixable with additional experiments: an action-centric baseline with the same validity scaffolding, an oracle/ablation that removes the learned model, and transition-prediction accuracy metrics. If these controls show that the gains persist, the paper would be a solid contribution. If they do not, the central claim must be substantially weakened to 'neuro-symbolic decoding with learned successor ranking.' The paper self-cites Plansformer (same research group) extensively, but that is not the basis for my recommendation."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The one thing to know about this paper: it's a well-run empirical study that overclaims its central result. The authors propose a state-centric approach to generalized planning—predict the next state's embedding rather than the next action—and then decode by enumerating all valid symbolic successors and picking the one closest to the prediction. In a few domains this beats action-sequence transformers on out-of-distribution extrapolation, with far smaller models and less data. If you're working on neural planning, this is worth reading.\n\nWhat's actually new is the specific combination of goal-conditioned successor embedding prediction with WL features and nearest-neighbor symbolic decoding. The related work makes clear that WL features plus lightweight regressors already match GNNs for heuristic learning, and model-based RL is established, but nobody seems to have framed generalized planning this way. The empirical study is careful: they ablate representations (WL vs fixed-size factored), prediction modes (state vs delta), and model classes (LSTM vs XGBoost), and they report splits for validation, interpolation, and extrapolation. Code and data are promised. That's real work and it deserves credit.\n\nThe soft spot is the comparison. In Algorithm 1, the method maintains the true symbolic state at every step, enumerates every valid successor using the domain transition function γ, and only ranks those successors by embedding distance. So the learned model never has to produce an executable action or even a state outside the successor set. The action-centric baselines—Plansformer, PlanGPT, SymT—don't get that scaffolding; they have to generate action sequences from scratch. That means the reported gains could come from the neuro-symbolic decoding shell rather than from the learned transition model. The paper itself says the decoding 'guarantees that every generated action is applicable.' The stress-test note makes exactly this point, and I think it lands.\n\nTo support the attribution, the authors should include a control where the ranking is replaced by something trivial—random choice among valid successors, or a simple heuristic—and show that their transition model beats that. They also need to give an action-centric baseline a similar per-step validity filter or re-ranking to level the playing field. Without that, the abstract's assertion that 'learning explicit transition models yields higher OOD success than direct action-sequence prediction' is not established. The Logistics result (all models 0.00) reinforces the point: the scaffolding isn't sufficient.\n\nMinor issues: the WL dimensions are inconsistent between Table 3 and Appendix E.4, the Plansformer baseline is run with a checkpoint trained only on Blocksworld and then used zero-shot on other domains, which is unfair but not central, and the beam-width implementation isn't described in enough detail.\n\nBottom line: this is a solid empirical paper with an interesting idea, but the headline claim gets ahead of the evidence. It deserves a proper referee, and with additional baselines and a more careful framing it could be a good contribution. I'd send it to peer review, and I'd ask for those controls. I would not cite its abstract claim in my own work until the controls are done.","headline":"The paper shows a promising hybrid pipeline, but the abstract overclaims: the OOD gains are likely driven by symbolic successor enumeration, not by the learned transition model itself.","tokens_in":18928,"tokens_out":4389,"would_cite":false,"duration_ms":39572,"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":"Learning to predict successor states, not action sequences, is the stronger inductive bias for generalized planning, and small size-invariant models achieve it with far fewer parameters and training instances.","keywords":["generalized planning","transition-model learning","neuro-symbolic planning","Weisfeiler-Leman embeddings","out-of-distribution generalization","sample efficiency","plan decoding"],"falsifier":"Replace the learned transition model with a trivial predictor (always predict zero delta, or a fixed random target) while keeping the same WL embeddings and symbolic successor decoding, and measure extrapolation success on the same splits. If the trivial predictor retains most of the 50% and 100% success on Blocksworld and VisitAll, the symbolic scaffolding is carrying the generalization; if success collapses, the learned model is load-bearing. A second check: decode by rounding the predicted embedding to a nearest state without enumerating symbolic successors — collapse would confirm that enu","tokens_in":18104,"feed_emoji":"🧩","tokens_out":10396,"duration_ms":85207,"temperature":0.7,"pith_summary":"This paper tries to establish that generalized planning — finding one strategy that solves a whole family of planning problems — is better learned as a transition-modeling problem than as an action-sequence prediction problem. The authors train a small goal-conditioned model to predict the successor state in a fixed-dimensional embedding space, then decode the plan by selecting the valid symbolic successor closest to that prediction. With Weisfeiler–Leman state embeddings, these compact models extrapolate to planning instances with many more objects than seen in training: in Blocksworld and VisitAll they match or exceed published Transformer-based planners while using orders of magnitude fewer parameters and no data augmentation. The same method does not solve Logistics under strict extrapolation, and the decoding step relies on symbolic successor enumeration at every step. A sympathetic reader would care because the result suggests that learning domain dynamics, not scaling action-sequence models, is the more promising route to out-of-distribution generalization in planning.","feed_headline":"Next-state prediction beats action prediction beyond training sizes","feed_subtitle":"Compact models that forecast successor states match or beat far larger action-sequence models on much larger planning instances","key_machinery":"The load-bearing object is the residual transition model in WL embedding space plus the neuro-symbolic decoding rule. WL embeddings map any state–goal pair to a fixed-dimensional histogram of color-refinement classes, making input size independent of object count; residual prediction encodes STRIPS frame axioms; and the nearest-valid-successor rule converts the embedding-space prediction into a guaranteed-valid symbolic action. LSTM and XGBoost serve as the function approximators, demonstrating that neither architectural scale nor sequential memory is the source of the gains.","core_discovery":"The paper's discovery is a state-centric formulation of generalized planning: instead of modeling p(action sequence | problem), learn a goal-conditioned transition model fθ that predicts a residual Δ_t in a Weisfeiler–Leman embedding space, so that the predicted next-state embedding is φ(s_t) + Δ_t. At each step the planner computes all valid symbolic successors Succ(s_t) from the domain operators and selects the one whose embedding is closest to the prediction; this guarantees every executed action is applicable and corrects drift. The empirical claim is that this explicit transition-model bias, combined with permutation- and size-invariant WL embeddings, yields higher out-of-distribution s","pith_inferences":["Because decoding always has the true symbolic state and the full successor set, the reported gains may be due to symbolic scaffolding rather than the learned model; a zero-residual control would separate these.","Using the transition model as a soft heuristic inside a classical search, rather than greedy nearest-neighbor rollouts, is a natural extension that might address the myopia behind the Logistics failure.","Training on multi-step or abstract transitions, while keeping the same decoding interface, could extend the approach to hierarchical and long-range dependency domains.","If a trivial predictor retains most of the extrapolation success, the practical lesson shifts to size-invariant representations plus symbolic decoding, with the model quality bar reduced to ranking near-miss successors."],"forward_implications":["Small state-centric models can extrapolate to object counts far beyond training, so data and parameter scaling are not the only route to out-of-distribution planning.","Permutation- and size-invariant state encodings are a necessary ingredient: fixed-slot factored encodings obtain near-zero extrapolation success across all domains.","Residual (delta) transition prediction improves non-parametric models in sparse STRIPS domains, consistent with the add/delete semantics of transitions.","Symbolic successor decoding guarantees plan validity at every step, eliminating the invalid-action problem that action-sequence models face.","Hierarchical domains with multi-layer causal coupling remain out of reach for one-step learned transitions under strict extrapolation."],"fun_headline_variants":["Predicting next states beats action prediction in planning","State-focused models outperform action-seq for generalization","Learn transitions, not actions, for sample-efficient planning","Size-invariant planning via learned successor-state prediction","Transition learning yields generalizable plans with less data"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise is that the extrapolation success is attributable to the learned transition model: at every step the method knows the true symbolic state, enumerates every valid successor, and only needs to rank them, so if the symbolic scaffolding — not the learned prediction — drives generalization, the paper's central claim about transition-model learning is not established.","fun_headline_variants_meta":{"raw":{"variants":["Predicting next states beats action prediction in planning","State-focused models outperform action-seq for generalization","Learn transitions, not actions, for sample-efficient planning","Size-invariant planning via learned successor-state prediction","Transition learning yields generalizable plans with less data"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000216,"raw_usage":{"total_tokens":1286,"prompt_tokens":778,"completion_tokens":508,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":522,"completion_tokens_details":{"reasoning_tokens":437}},"tokens_in":522,"tokens_out":508,"duration_ms":5771,"temperature":1.0,"reasoning_tokens":437,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T20:26:32.156961+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Replace the learned transition model with a trivial predictor (always predict zero delta, or a fixed random target) while keeping the same WL embeddings and symbolic successor decoding, and measure extrapolation success on the same splits. If the trivial predictor retains most of the 50% and 100% success on Blocksworld and VisitAll, the symbolic scaffolding is carrying the generalization; if success collapses, the learned model is load-bearing. A second check: decode by rounding the predicted embedding to a nearest state without enumerating symbolic successors — collapse would confirm that enu","supporting_citations":[],"review_version":1}