{"id":"344878ef-9116-431d-8279-8c5b337088c0","arxiv_id":"2412.08565","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A discrete-flow sequence model with energy and entropy guidance enables adaptive planning that beats LEAP and Decision Transformer in BabyAI adaptive benchmarks.","lead":"GenPlan is a new AI planner that generates sequences of actions, goals, and states by iteratively denoising random trajectories with discrete flow models. It reports higher success rates than prior planners on adaptive multi-task benchmarks in simulated maze and robot environments, including zero-shot transfer from simple training tasks to harder missions.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. 5's loss is identically zero for the uniform interpolant (Eq. 3b), so GenPlan-U cannot be trained by the published objective; the reported uniform-variant results are unverifiable as described.","rationale":"The reader's weakest assumption concerns OOD generalization of the learned energy function. That is a legitimate concern, but it is secondary: if the described training objective cannot train one of the paper's two main variants, the empirical claim is unverifiable before any discussion of generalization. The masking-based variant (GenPlan-M) remains well-defined, so the headline result might survive if restricted to GenPlan-M, but the paper explicitly presents both interpolants as interchangeable and reports results for both. The internal inconsistency in Eq. 3b vs Eq. 5 is concrete, checkable, and directly affects the reported numbers. The code release is claimed but unverified, so a code check is the natural first step. The verdict should move from CONDITIONAL to UNVERDICTED because the current text does not permit an independent reader to reproduce GenPlan-U, and the central claim depends partly on that variant. If the code reveals a different loss for uniform corruption, the paper can be corrected; until then, the method as written is incomplete.","tokens_in":19587,"tokens_out":7923,"duration_ms":89203,"concrete_test":"Inspect the released code's training loop for GenPlan-U, or run Algorithm 1 verbatim on a single BabyAI task (e.g., GoToLocalS10N10G2) using Eq. 3b and Eq. 5. If the training loss is exactly zero and gradients with respect to θ vanish for all corrupted tokens, the published objective cannot train the uniform variant, so the reported GenPlan-U results must have come from an unreported objective. Alternatively, re-run GenPlan-U with the corrected all-positions loss and compare success rates; if performance drops materially, the manuscript's error bar and ablation analysis change.","verdict_should_be":"UNVERDICTED","load_bearing_attack":"The training objective and the uniform corruption schedule are mutually inconsistent. Eq. 3b defines the uniform interpolant as p_unif_{t|1}(x_t|x_1) = C(t δ{x_1,x_t} + (1-t)/|X|), whose support is the ordinary token set {1,...,|X|}; the mask state [M] never appears. The denoising loss in Eq. 5 is L_x = -Σ_{k=1}^H δ{x_k^t, [M]} log pθ_1|t(x_1|x_t,o), so under the uniform interpolant the Kronecker delta is identically zero for every position and L_x ≡ 0. Algorithm 1 line 4 also only calls pmask, and line 5 sets LNLL ← La + Ls + Lg using Eq. 5; for GenPlan-U the total loss in line 7 would therefore contain no likelihood term, leaving only the entropy regularizer, which can be trivially satisfied by a uniform output. Yet Tables 1, 2, 4, and 9 report GenPlan-U success rates close to GenPlan-M. This is not an OOD-generalization concern; it is an internal contradiction in the specification of a reported variant. Either the implementation actually uses a different objective (e.g., Eq. 4a without the mask delta, or a loss over positions where x_t ≠ x_1) and the paper misdescribes it, or the uniform variant cannot learn at all. The central claim 'GenPlan outperforms SOTA by over 10%' partly relies on GenPlan-U (AP mean 46.8% vs LEAP 38.4%), so this gap in the method description is load-bearing for the empirical support.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes GenPlan, a generative planner built on discrete flow models (CTMC-based denoising) that jointly denoises sequences of actions, states, and goals. Training uses a negative log-likelihood objective with an entropy lower bound, and planning is done by iterative reverse-CTMC sampling from a learned denoiser. The authors evaluate GenPlan on BabyAI trajectory planning, instruction completion, and adaptive planning tasks, plus continuous PushT and Franka Kitchen tasks, and report success-rate improvements, including AP results showing GenPlan outperforming LEAP by roughly 13.8 percentage points.","tokens_in":19965,"tokens_out":4486,"duration_ms":52373,"significance":"If the reported results are reproducible, GenPlan is a useful contribution: it demonstrates that a discrete flow model with joint goal/state/action denoising and entropy regularization can transfer from simple single-goal demonstrations to novel multi-goal, door-opening, and obstacle-unblocking tasks without oracle goal conditioning. The manuscript is also commendable for releasing code, comparing against energy-based and diffusion baselines, and including ablations of the joint prediction, entropy, and noise-schedule components. However, the uniform-noise variant appears to be untrainable under the published objective, the central empirical claims lack error bars, and the advertised 'energy-guided sampling' is not what Algorithm 2 actually does; these issues currently prevent the results from being verified as described.","major_comments":[{"comment":"The uniform interpolant and the training loss are mutually inconsistent. Under Eq. (3b), x_t takes values only in the ordinary token set, so the mask state [M] never appears; consequently the Kronecker delta in Eq. (5) is identically zero for every position, making L_x identically zero, and Algorithm 1 would train only the entropy regularizer. Since Tables 1, 2, and 9 report GenPlan-U success rates (and the AP mean in Table 2 is a load-bearing part of the abstract's 'over 10%' claim), the paper must either specify the actual loss minimized for GenPlan-U (e.g., a loss over positions where x_t differs from x_1) or state explicitly that the uniform variant is trained with a masking schedule and only sampled with a uniform rate matrix. As written, the reported GenPlan-U results cannot be reproduced.","section":"§3–§4, Eqs. (3b), (5), Algorithm 1"},{"comment":"The headline comparisons are single numbers with no estimates of variance. Success rates are averaged over 250 environments, but no standard errors, confidence intervals, or number of seeds are reported for the main tables; Figure 10A reports three seeds for only one environment. Without variance information, the observed differences (e.g., GP-U 46.8% vs. LEAP 38.4% on adaptive planning) cannot be assessed for statistical significance, and some ablation differences in Table 3 are as small as 0.8 percentage points. Please report mean ± standard deviation (or confidence intervals) across seeds, or at minimum across environment resamplings.","section":"§5, Tables 1, 2, and 9"},{"comment":"The text repeatedly states that sampling is energy-guided and that the optimization 'stays at an energy minimum at each step,' but Algorithm 2 contains no energy evaluation and no energy-gradient or energy-based accept/reject step; it only samples from the rate matrix R^θ_t that is built from the learned denoiser p^θ_{1|t}. If the energy function E(a) is only an implicit construction of the training objective, the paper should say so and should not claim energy-guided sampling without either placing an energy term in Algorithm 2 or explaining precisely how the energy landscape shapes the CTMC sampling. This is not just a wording issue, because Section 1 and the limitations discussion attribute the generalization gains to the energy landscape.","section":"§4, Energy-Guided Denoising Model vs. Algorithm 2"}],"minor_comments":[{"comment":"The expectation in Eq. (4a) is written over a0 and o but the loss involves a1; the intended distribution over a1 (presumably the data distribution) should be stated explicitly.","section":"§4, Eq. (4a)"},{"comment":"The update rule for the Lagrangian multiplier λ is not fully specified: the step size for the λ update is omitted, and the entropy term in line 6 is evaluated at τ0,o while Eq. (4b) is stated over a0,o; please clarify.","section":"Algorithm 1, line 9"},{"comment":"The column 'GenPlan-M Reduction' is confusing: it reports the reduced success rate and the decrease relative to a baseline, but it is not clear whether the baseline is the full GenPlan-M from Tables 1–2 or a different run; please state the reference value explicitly.","section":"Table 3"},{"comment":"The parenthetical deltas such as 'TP (7.6 ↑)' and 'AP (13.84 ↑)' are not defined in the table captions or text; specify which baseline(s) these deltas are computed against.","section":"§5, Tables 1–2"},{"comment":"The uniform-interpolant example in Table 20 shows a token '0' that is not in the action set {0,...,5} defined elsewhere; please clarify how the uniform noise state space relates to the actual token vocabulary.","section":"Appendix A, Tables 19–20"},{"comment":"The AP discussion says GenPlan 'can eventually find a plan given enough timesteps as long as the energy function generalizes well,' which is a conditional statement and not a demonstrated property; please mark this clearly as an assumption rather than a conclusion.","section":"§5, Results section"}],"recommendation":"major_revision","confidential_remarks":"The uniform-variant training inconsistency is the most serious issue: it directly affects a reported variant that contributes to the abstract's central performance claim. If the released code resolves the inconsistency (for instance, by using a different loss for GenPlan-U), the authors should be asked to align the paper with the code and to add the missing statistical reporting. I would not reject outright because the core masked variant is clearly specified and the code is public, but the current write-up is not verifiable as it stands."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague, you should know two things about this paper. The core idea is genuinely interesting: a sequence-level planner built on discrete flow models, trained with an entropy-regularized negative-log-likelihood energy objective, and sampled via CTMC. Unlike LEAP or DT, it jointly denoises actions, states, and goals, letting the agent propose sub-goals during rollout instead of relying on an oracle. That is a real step for adaptive multi-task planning, and the BabyAI results for the masked variant (GenPlan-M) are respectable: 52.3% vs LEAP's 38.4% on adaptive planning, with similar gains in instruction completion.\n\nThe soft spots are more serious than the paper admits. First, the uniform-noise variant GenPlan-U is internally inconsistent as described. The training loss in Eq. (5) is only nonzero for positions corrupted to the mask token [M], but the uniform interpolant in Eq. (3b) never produces [M]. So Lx is identically zero for GenPlan-U, and Algorithm 1's loss reduces to the entropy regularizer alone, which can be satisfied trivially. Yet Tables 1, 2, 4, and 9 report GenPlan-U success rates close to GenPlan-M. Either the implementation uses a different objective (and the paper misdescribes it) or those numbers are not reproducible. This is load-bearing because the headline 10% improvement partly relies on GenPlan-U's AP mean.\n\nSecond, the paper describes energy-guided sampling, but Algorithm 2 is just vanilla CTMC sampling with the denoiser; no energy gradient or energy value enters the sampling loop. The energy is used only in training, so the description overstates the mechanism. Third, the main BabyAI tables have no error bars, and the 'novel skill' claim is explicitly a hypothesis, not established evidence.\n\nThe core method and masked-variant results are credible enough to deserve a careful referee, but the paper needs major revisions: fix or drop the uniform variant, clarify the energy guidance, and add variance reporting. If the code checks out, this could be a useful contribution.","headline":"Promising DFM-based planner with a clean idea, but a load-bearing inconsistency in the uniform-noise variant and loose energy-guidance description make the experimental claims hard to verify as written.","tokens_in":20459,"tokens_out":5821,"would_cite":false,"duration_ms":58082,"reading_group":"yes","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"GenPlan is a discrete-flow planner that jointly denoises states, actions, and goals; the paper reports it outperforms prior planners by over 10 percentage points on adaptive planning tasks where an agent trained on simple demos must solve…","keywords":["offline reinforcement learning","sequence modeling","discrete flow models","continuous-time Markov chains","energy-based planning","adaptive planning","zero-shot generalization","BabyAI"],"falsifier":"Train GenPlan only on single-goal demonstrations in a planar environment, then evaluate it in a novel maze where the mission requires an unblocking or door-opening skill that never appeared in training, with no simulator-provided goal oracle. If its success rate does not substantially exceed LEAP without goal conditioning and Decision Transformer under identical conditions, the claimed generalization of the energy function fails. A more direct check is to compute, in a held-out maze, whether the learned energy assigns lower values to expert trajectories than to stalling or repeatedly-forward trajectories; if it does not, the energy objective is not the mechanism carrying the adaptation.","tokens_in":19400,"feed_emoji":"🤖","tokens_out":8761,"duration_ms":89776,"temperature":0.7,"pith_summary":"GenPlan turns planning into a generative sequence-modeling problem: a discrete flow model iteratively denoises a corrupted trajectory of states, actions, and goals until it forms a plan. The paper trains an energy-based denoiser with a lower bound on entropy so the planner keeps exploring instead of collapsing onto repeated actions from the demonstrations. Its central claim is that this recipe lets an agent trained only on simple single-goal demonstrations adapt zero-shot to harder missions involving multiple goals, locked doors, and obstacles. A sympathetic reader should care because the method removes the need for a simulator-given goal oracle at test time: the goal sequence is generated together with the actions, which is what prior planners required.","feed_headline":"Discrete-flow planner beats prior planners by 13.8 points","feed_subtitle":"Trained on single-goal demos, GenPlan zero-shot handles locked doors and obstacles.","key_machinery":"The load-bearing object is the joint denoising model $p^\\theta_{1|t}$ together with the reverse CTMC sampler of the discrete flow model. The denoiser is trained to minimize the negative log-likelihood of clean tokens at masked positions, subject to $\\mathbb{E}[H(p^\\theta_{1|t})] \\geq \\beta$; this entropy constraint is what lets the planner propose actions and goals that were not in the training demonstrations. The reverse process uses a rate matrix $R_t$ with masking or uniform interpolants to evolve a corrupted trajectory $\\tau^t$ back to a clean plan $\\tau^1$. All three trajectory components, states, actions, and goals, are denoised jointly by a bidirectional transformer conditioned on FiLM-encoded image, instruction, and position observations, so goal discovery is part of planning rather than an input.","core_discovery":"On its own terms, the paper's central discovery is that a maximum-likelihood objective is the wrong training target for adaptive planning and that an entropy-regularized energy objective over whole trajectories is the right one. GenPlan models a trajectory as discrete tokens, corrupts it with either masking or uniform noise, and trains a bidirectional transformer denoiser $p^\\theta_{1|t}$ to predict the clean states, actions, and goals. The energy of a sequence is the accumulated negative log-likelihood the denoiser assigns to it, and the entropy constraint keeps the action distribution from collapsing. At inference, the reverse continuous-time Markov chain turns noise into a plan, jointly producing goals and actions rather than taking goal positions from a simulator. In the BabyAI adaptive-planning suite, GenPlan-M reaches 52.3% mean success versus 38.4% for LEAP and 18.2% for Decision Transformer, and it is the only method to make progress on key-collection and obstacle-unblocking missions; in continuous PushT and Kitchen benchmarks it is competitive with, though not ahead of, diffusion-policy and VQ-BeT baselines.","pith_inferences":["A direct but untested extension is to treat GenPlan's entropy-regularized discrete-flow objective as a general sequence-generation prior: any domain where low-energy sequences are rare and mode collapse is fatal, such as instruction following or program synthesis, could inherit the same zero-shot refinement behavior.","The reported adaptive-planning margin is established in discrete grid-world planning; in the paper's continuous manipulation benchmarks GenPlan is competitive but not dominant, which suggests the advantage is tied to discrete, order-structured planning rather than low-level control.","One could test the generalization mechanism directly by measuring whether the learned energy ranks expert trajectories below stalling or repeatedly-forward trajectories in novel mazes; if it does not, the energy objective would need a different explanation for the observed success."],"forward_implications":["A planner can be trained on one simple task distribution and evaluated on substantially harder tasks with zero fine-tuning, since the goal-generation module proposes sub-goals during denoising.","Simulator-based goal oracles, which LEAP and Decision Transformer rely on, become unnecessary; this matters for real deployments where goal positions are not known in advance.","Order-critical long-horizon missions, such as collecting a key before opening a door, become feasible for demonstration-trained planners, where behavior-cloning baselines stall or loop.","The entropy regularizer provides a mechanism for discovering skills such as door opening and obstacle unblocking that are absent from the training data, not just for stitching known skills.","Masking and uniform noise interpolants both work, with masking training faster, so the planner can trade exploration for convergence by choosing the corruption schedule."],"supporting_citations":[{"why":"Supplies the discrete flow model objective, conditional rate matrices, and joint denoising formulation that GenPlan is built on.","marker":"Campbell et al. 2024"},{"why":"Provides the continuous-time Markov chain framework for discrete denoising models used in sampling.","marker":"Campbell et al. 2022"},{"why":"LEAP is the main energy-minimization baseline and the source of the iterative energy minimization idea GenPlan adapts.","marker":"Chen et al. 2023"},{"why":"Decision Transformer is the sequence-modeling baseline and supplies the transformer-style architecture that GenPlan modifies with bidirectional masks.","marker":"Chen et al. 2021"},{"why":"Contributes the BabyAI simulation suite and instruction encoder used for all discrete planning evaluations.","marker":"Chevalier-Boisvert et al. 2019"},{"why":"Provides the interpretation of masked-language-model scores as an implicit energy network that GenPlan's energy objective is adapted from.","marker":"Goyal, Dyer, and Berg-Kirkpatrick 2021"},{"why":"Supplies VQ-BeT, the categorical latent-action baseline and action tokenization used in continuous manipulation comparisons.","marker":"Lee et al. 2024"},{"why":"Supplies Diffusion Policy baselines and the transformer and FiLM architecture details used in the implementation.","marker":"Chi et al. 2023"}],"fun_headline_variants":["GenPlan's discrete-flow planner beats prior planner by 13.9 points","Flow-based adaptive planner zero-shot solves locked-door tasks","Adaptive planner with discrete flow outdoes baselines on unseen missions","GenPlan: discrete-flow generative model for adaptive planning","13.9-point win: GenPlan's flows master adaptive missions"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the energy function learned on simple single-goal training environments keeps assigning low energy to successful trajectories in harder, unseen test environments; the paper states this as a hypothesis (Section 1: 'as long as the objective function remains consistent'; Section 4: 'provided the energy function is generalizable') rather than proving it.","fun_headline_variants_meta":{"raw":{"variants":["GenPlan's discrete-flow planner beats prior planner by 13.9 points","Flow-based adaptive planner zero-shot solves locked-door tasks","Adaptive planner with discrete flow outdoes baselines on unseen missions","GenPlan: discrete-flow generative model for adaptive planning","13.9-point win: GenPlan's flows master adaptive missions"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000946,"raw_usage":{"total_tokens":4086,"prompt_tokens":1041,"completion_tokens":3045,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":657,"completion_tokens_details":{"reasoning_tokens":2958}},"tokens_in":657,"tokens_out":3045,"duration_ms":24564,"temperature":1.0,"reasoning_tokens":2958,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T17:44:01.995541+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train GenPlan only on single-goal demonstrations in a planar environment, then evaluate it in a novel maze where the mission requires an unblocking or door-opening skill that never appeared in training, with no simulator-provided goal oracle. If its success rate does not substantially exceed LEAP without goal conditioning and Decision Transformer under identical conditions, the claimed generalization of the energy function fails. A more direct check is to compute, in a held-out maze, whether the learned energy assigns lower values to expert trajectories than to stalling or repeatedly-forward trajectories; if it does not, the energy objective is not the mechanism carrying the adaptation.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the continuous-time Markov chain framework for discrete denoising models used in sampling."},{"cited_title":"B.; and Vela, P","cited_arxiv_id":null,"evidence_quote":"LEAP is the main energy-minimization baseline and the source of the iterative energy minimization idea GenPlan adapts."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Decision Transformer is the sequence-modeling baseline and supplies the transformer-style architecture that GenPlan modifies with bidirectional masks."}],"review_version":1}