{"id":"3e7726c3-7afc-4a66-8858-4ec7e01620d3","arxiv_id":"2505.21750","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"HIDI trains a GP-regularized conditional diffusion model to generate high-level subgoals and mixes GP mean selection with diffusion sampling, outperforming HRL baselines on continuous control benchmarks.","lead":"Researchers combined diffusion models with Gaussian process uncertainty to generate subgoals for hierarchical reinforcement learning, and tested the method on MuJoCo continuous control tasks. The method, called HIDI, reports better sample efficiency and success rates than several prior HRL baselines.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"GP regularization objective in Eq. 11 is not the predictive NLL used in the stated gradient; as written, the paper's core mechanism is unspecified and the theoretical support (Thm 3.1, Prop 3.2) does not apply.","rationale":"The paper's headline claim is that HIDI outperforms prior HRL methods; the empirical curves and ablations are consistent and 10-seed averages are reported. The reader's weakest assumption, the Section 2 conjecture that HIRO-relabeled subgoals asymptotically approximate an optimal high-level policy, is a real limitation, but it is upstream justification rather than a direct threat to the benchmark comparisons: even a biased relabeled distribution can yield a practically useful subgoal generator, and prior HIRO-based methods share the same assumption. The more load-bearing issue is internal: the GP prior is the paper's central novelty ('regularized by a Gaussian Process prior' in the abstract), yet the written objective in Eq. 9 and Eq. 11 is a joint prior over training outputs, not the conditional predictive density used in Theorem 3.1/Proposition 3.2 (Eq. 31) and Appendix A.1.3. The gradient formula in Eq. 31 follows only for the sparse GP predictive NLL, whose gradient w.r.t. g is (g - μ*)/σ*², not for the prior NLL, whose gradient is (KN + σ²I)^{-1}g. There is also a dimensional mismatch: KN is N×N over replay states, so g in Eq. 9 would need to be in R^N, while generated subgoals are in R^d. Because no code is released, the actual implemented loss cannot be recovered. This means the regularization mechanism—and hence the ablation gains of roughly 15% attributed to it—cannot be verified from the paper as written. I also note that Theorem 3.3 and Proposition 3.4 are weak: the regret bound is a decomposition of the mixture policy, and the policy-improvement proof assumes high-Q subgoals cluster around the GP mean, which is close to the conclusion. These concerns reinforce the reader's CONDITIONAL verdict rather than changing it; they are corrigible in revision. Hence verdict_should_be = UNCHANGED.","tokens_in":23628,"tokens_out":9594,"duration_ms":93852,"concrete_test":"Re-derive ∇θh Lgp from Eq. 11 using the chain rule and compare it to Eq. 31. For Eq. 9's density, ∇_g[-log N(g; 0, KN + σ²I)] = (KN + σ²I)^{-1}g, so the gradient is not (g - μ*(s))/σ²*(s); additionally, dimensional analysis shows g in Eq. 9 must lie in R^N while generated subgoals lie in R^d, making the expression undefined for N ≠ d. If the intended loss is instead the sparse GP predictive NLL of Appendix A.1.3, then Eq. 9 and Eq. 11 must be corrected, and the proof of Theorem 3.1 must be re-derived to connect the prior to the predictive distribution. Without this correction, the 15% ablation gain credited to GP regularization is not attributable to a well-defined mechanism.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central novelty is the GP prior that 'regularizes' the diffusion subgoal generator, yet the written objective is internally inconsistent. Eq. 9 defines p(g|s; θgp) = N(g; 0, KN + σ²I), where KN is an N×N kernel matrix over states in the replay buffer. This is a joint prior over training outputs, not a conditional density for a single subgoal given a query state: it does not depend on s as required, and g would need dimension N, not the subgoal dimension d. Eq. 11 then defines Lgp = E_{s∼Dh, g∼πθh}[-log p(g|s)] using this prior. However, Theorem 3.1, Proposition 3.2, and Appendix A.1.3 analyze a different loss: the negative log density of the sparse GP predictive distribution N(g | μ*(s), σ²*(s)I). The gradient formula in Eq. 31, ∇θh Lgp = E[((g - μ*)/σ*²)ᵀ ∇θh g], follows only for the predictive NLL; the gradient of the Eq. 11 prior NLL with respect to g is (KN + σ²I)^{-1} g, which is not equal to (g - μ*)/σ*² in general. Thus the stated theoretical justification does not apply to the stated objective. Because no code is released, the actual implemented loss is not recoverable, so the roughly 15% ablation improvement attributed to GP regularization cannot be checked. This is a load-bearing specification failure: the paper's contribution is precisely the GP-guided diffusion mechanism, and without a consistent loss the described method is not reproducible.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes HIDI, a hierarchical reinforcement learning method in which the high-level subgoal policy is a conditional diffusion model. The diffusion policy is trained with a denoising objective on HIRO-relabeled subgoals, an auxiliary Gaussian Process (GP) regularization term intended to quantify uncertainty, and a TD3-style deterministic policy gradient objective. At decision time, subgoals are sampled from the diffusion policy with probability 1−ε and from the sparse GP predictive mean with probability ε. The paper reports experiments on continuous control benchmarks (Reacher, Pusher, Point Maze, Ant Maze variants, Ant Fall, Ant FourRooms, including stochastic and image-based variants) showing that HIDI outperforms HIRO, HRAC, HIGL, SAGA, and HLPS on all twelve reported task/reward settings, with ablations isolating the diffusion model, GP regularization, subgoal selection, diffusion steps, and the weights ψ and η. Appendix A provides proofs for diffusion-distribution validity, a GP gradient formula, and regret and policy-improvement statements for subgoal selection.","tokens_in":24107,"tokens_out":12277,"duration_ms":130466,"significance":"If the empirical results are reproducible, HIDI is a valuable contribution: it demonstrates that a generative high-level policy can outperform strong off-policy HRL baselines across a broad and consistent suite of tasks, with 10 seeds, systematic ablations, and qualitative subgoal-reachability visualization. The sparse-GP machinery is a principled way to add uncertainty information to a diffusion policy, and the proposed subgoal selection strategy is simple and effective. However, the paper's theoretical support is not reliable in its current form, and the main-text specification of the GP loss is inconsistent with the appendix analysis. The empirical story is externally grounded in environment success rates and does not depend on the theory, so the central claim may survive revision, but the method as written cannot be reimplemented without guessing which of two different GP losses was actually used. No code is released, which amplifies the reproducibility risk.","major_comments":[{"comment":"The GP regularization objective is specified inconsistently. Equation (9) defines p(g|s; θ_gp) as an N-dimensional joint prior N(0, K_N + σ²I) over replay-buffer outputs, so the density in Eq. (11) does not condition on the query state s and cannot be evaluated for a d-dimensional generated subgoal g. The gradient claimed in Proposition 3.2 and derived in Eq. (31), (g − μ_*(s))/σ_*², is the gradient of the sparse GP predictive NLL in Eq. (30), not of Eq. (11); for Eq. (11), the gradient through g is (K_N + σ²I)^{-1} g, which pulls g toward zero rather than toward the GP predictive mean. The appendix analyzes a different loss from the one defined in the main text, and since no code is released, the actually implemented objective is not recoverable. This is a load-bearing specification failure: the ablation gain of roughly 15–16% attributed to GP regularization and the theorems in §3.2 are attached to an internally inconsistent description. Please unify the objective, either by using the sparse predictive NLL throughout and removing or replacing Eqs. (9) and (11), or by specifying the implemented loss and proving the gradient formula for that loss.","section":"§3.2, Eqs. (9)–(11); Appendix A.1.3, Eqs. (30)–(31)"},{"comment":"The theoretical guarantees for subgoal selection are assumption-laden to the point of near-circularity. Theorem 3.3's regret bound is ε(R*(s) − R_min) + (1−ε)δ, where R_min is only assumed to be a lower bound on R(s, μ(s)) and can be arbitrarily far below R*(s); without an additional condition linking μ(s) to high-reward subgoals, the bound is vacuous. The proof of Proposition 3.4 assumes that high-Q subgoals in the buffer cluster around μ(s), that μ(s) approximates g* within error ϵ, that Q_h is Lipschitz with constant L, and that Lϵ < Δ; these assumptions are essentially the desired conclusion restated as conditions. The near-optimal diffusion policy assumption (Assumption A.4) also assumes the core property the regret bound is meant to quantify. Please reformulate these results with explicit coverage and smoothness conditions that make the bounds non-vacuous, or clearly label them as heuristic motivation rather than formal guarantees.","section":"Appendix A.2, Theorem 3.3 and Proposition 3.4"},{"comment":"The identity in Eq. (27), D_KL(q(g0|s)||p_θ(g0|s)) = L_VLB(θ_h) − H(q(g0|s)), is incorrect as an equality: it drops the nonnegative term E_q[D_KL(q(g1:N|g0,s)||p_θ(g1:N|g0,s))]. The correct relation is an inequality, D_KL ≤ L_VLB − H(q), and consequently D_KL ≤ L_VLB. The intended bound can be repaired because D_KL ≤ L_VLB whenever the L_VLB bound holds, but the proof as written does not establish Theorem A.1. Please correct Eq. (27) and the surrounding argument.","section":"Appendix A.1.2, Eq. (27)"}],"minor_comments":[{"comment":"The caption of Table 2 says the results are for policies obtained after 5M steps with sparse rewards, but the table also contains dense-reward rows and Figure 1 shows 10M-step x-axes for Stochastic Ant Fall and Stochastic Ant FourRooms. Please make the training horizon and reward settings consistent between the table and figures.","section":"Table 2 and Figure 1"},{"comment":"The labels in Fig. 2(d) and the surrounding text render the ε probability as a blank symbol in the caption (e.g., 'HIDI ( = 0.05)'). Please fix the typography so the subgoal selection probability is displayed correctly.","section":"Section 4.2, Fig. 2"},{"comment":"Line 14 stores the transition (s_{t−1}, g_{t−1}, a_t, r_t, s_t, g_t), but the low-level transition should be indexed consistently with the notation in Section 2, e.g., (s_t, a_t, r_t, s_{t+1}) with the current subgoal g_t. Please clarify the indexing.","section":"Algorithm 1, line 14"},{"comment":"Remark A.3 states that minimizing L_gp reduces the KL divergence between p_θh(·|s) and the GP predictive distribution, but the displayed identity includes an additional entropy term H(p_θh(·|s)). Minimizing L_gp therefore also encourages lower entropy, which is in tension with the claim that the diffusion model retains expressiveness. Please state this trade-off explicitly.","section":"Appendix A.1.3, Remark A.3"},{"comment":"The paper relies on the conjecture in Section 2 that HIRO-style relabeled subgoals asymptotically approximate an optimal high-level policy. This is an important assumption for the diffusion training target; it is acknowledged as a conjecture, but the paper should explicitly discuss what could go wrong if it fails and whether any experiments test the stationarity of the relabeled target. Also, 'Diffsuion-QL' in Section 4 is a typo for 'Diffusion-QL', and the Mish activation is used in Table 3 without a reference.","section":"Section 2 and related work"},{"comment":"The numbering of theoretical results is inconsistent between the main text and the appendix: Theorem 3.1 and Proposition 3.2 in the main text correspond to Theorem A.2 in Appendix A.1.3, but the connection is never stated. Proposition 3.4's proof is in Appendix A.2 without a matching label. Please align the numbering and add cross-references.","section":"Theorem numbering"}],"recommendation":"major_revision","confidential_remarks":"The empirical evaluation is a genuine strength: the method wins across all reported settings, the ablations are informative, and the qualitative subgoal visualization supports the claims. However, the inconsistency between the main-text GP loss and the appendix's sparse predictive NLL is a serious reproducibility problem, and the theoretical guarantees in §3.3/A.2 are largely vacuous as stated. If the authors can clarify the implemented objective and either repair or downgrade the theoretical claims, the paper could be suitable for publication. I would encourage the editor to ask for a clear statement of code or an exact pseudocode for the GP loss computation, since without it the central mechanism cannot be checked."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nHere's my take on Wang et al., arXiv:2505.21750. The empirical result is real and the method is genuinely new: a conditional diffusion model for online high-level subgoal generation in HRL, regularized by a GP prior, with a hybrid diffusion/GP-mean selection rule. The experiments are extensive — 13 task/reward settings, 10 seeds, ablations — and HIDI wins all of them. That is a solid within-subfield contribution.\n\nThe soft spot is the GP regularization. The paper's central novelty is the GP prior, but the written objective does not match the theory. Eq. 9/11 define Lgp using a joint prior N(g;0,KN+σ²I) over N training outputs. That is not a conditional density for a single subgoal given s, and the gradient formula in Prop. 3.2 / Eq. 31 only follows for the sparse GP predictive NLL N(g|μ*(s),σ*²(s)I), which is a different loss. So the stated theorem does not apply to the stated objective. Because no code is released, the implemented loss is not recoverable. That is load-bearing: the roughly 15% ablation gain attributed to GP regularization cannot be checked.\n\nThe theory also leans on assumptions close to the conclusion: Theorem 3.3 assumes the diffusion policy is near-optimal, and Prop. 3.4 assumes the GP mean sits in high-Q regions. The appendix Theorem A.1 has a further issue — Eq. 27 equates a KL divergence with L_VLB minus entropy, which is at best an inequality. None of this kills the empirical story, but the theoretical guarantees are overstated.\n\nWhat survives is solid: diffusion is a good online subgoal generator, and combining it with GP-style uncertainty guidance helps in practice. The paper deserves a serious referee — the empirical work is too strong to desk reject — but it needs major revision on the specification of Lgp, a code release, and a theory section that only claims what it actually proves.","headline":"Strong empirical HRL paper undermined by an inconsistent GP loss specification and overclaimed theory.","tokens_in":24552,"tokens_out":3590,"would_cite":false,"duration_ms":34942,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["68T05","68T07"],"pacs":[],"model":"deepseek-v4-flash","headline":"A conditional diffusion model regularized by a Gaussian process prior, plus a GP-mean selection rule, generates achievable subgoals for off-policy hierarchical RL, beating prior HRL methods on continuous control benchmarks.","keywords":["hierarchical reinforcement learning","diffusion model","subgoal generation","Gaussian process","uncertainty quantification","continuous control","hindsight relabeling","off-policy reinforcement learning"],"falsifier":"Track, during HIDI training on a stochastic Ant Maze task, the divergence between the relabeled-subgoal distribution used to train the diffusion model and the subgoal distribution that maximizes the trained high-level Q-function: if that divergence does not shrink as training progresses, the relabeling conjecture the method rests on is false and the reported gains cannot be credited to learning a near-optimal subgoal distribution.","tokens_in":23461,"feed_emoji":"🎯","tokens_out":13057,"duration_ms":112595,"temperature":0.7,"pith_summary":"Hierarchical reinforcement learning is unstable because the low-level policy keeps changing, so subgoals that used to be reachable stop being reachable and the high-level policy chases a moving training target. This paper proposes HIDI, which turns the high-level policy into a conditional diffusion model that generates subgoals by denoising noise conditioned on the current state, and regularizes it with a Gaussian process prior that adds uncertainty quantification. A hybrid selection rule occasionally substitutes the GP's predictive mean for a diffusion sample, trading expressiveness for reliability. If the paper is right, off-policy hierarchical RL can be stable and sample-efficient without constraining the subgoal space, and the method beats prior HRL approaches on MuJoCo continuous-control benchmarks, including stochastic and image-observation variants.","feed_headline":"Beats prior HRL methods with a diffusion-GP subgoal generator","feed_subtitle":"Maze, arm and four-room benchmarks show higher success rates and better sample efficiency.","key_machinery":"The load-bearing object is the generative high-level policy: a conditional diffusion model treated as a density over subgoals given the state, $\\pi_h(g|s) = p_{\\theta_h}(g^{0:N}|s)$, trained with the standard noise-prediction objective of denoising diffusion probabilistic models. Around it sits a sparse Gaussian process prior whose predictive mean $\\mu_*(s)$ and variance $\\sigma_*^2(s)$ define the uncertainty signal; the GP loss gradient takes the explicit form $\\nabla_{\\theta_h} L_{gp} = \\mathbb{E}\\left[\\frac{g - \\mu_*(s)}{\\sigma_*^2(s)} \\cdot \\nabla_{\\theta_h} g\\right]$, so parameter updates are strongest exactly where the GP is most confident. The third element is the hybrid selection rule $g^* = \\mu_*$ with probability $\\varepsilon$, otherwise a sample from $\\pi_h(\\cdot|s)$, which the theory analyzes as an $\\varepsilon$-mixed policy whose single-step regret is bounded by $\\varepsilon(R^*(s) - R_{\\min}) + (1-\\varepsilon)\\delta$.","core_discovery":"In HIDI, the high-level policy is not a neural actor mapping states to subgoals but the reverse process of a conditional diffusion model, $\\pi_h(g|s)$, which iteratively denoises Gaussian noise into a subgoal conditioned on the current state. The generator is trained jointly in three directions: a denoising objective over HIRO-relabeled subgoals, so the generated distribution tracks what the current low-level policy can actually reach; a Gaussian process prior whose negative log marginal likelihood penalizes subgoals inconsistent with the smooth structure of previously successful state-subgoal pairs; and a TD3-style objective that backpropagates through the denoising chain to maximize the high-level Q-function. At decision time, with probability $\\varepsilon$ the GP's predictive mean is used as the subgoal instead of a diffusion sample, a hybrid the authors argue combines the diffusion model's expressiveness with the GP's reliability. The paper claims this outperforms HIRO, HRAC, HIGL, SAGA, and HLPS on MuJoCo continuous-control benchmarks, with ablations attributing roughly 15% gains to the diffusion component, 15–16% to the GP regularization, and 7–8% to the selection rule, and it supports the selection rule with a bounded single-step regret result and a single-step policy improvement result.","pith_inferences":["The $\\varepsilon$-mixture rule is effectively an exploration thermostat: at $\\varepsilon = 0$ the high level is pure diffusion expressiveness and at $\\varepsilon = 1$ it is a deterministic GP-mean policy, so an adaptive $\\varepsilon$ schedule that starts high and decays would be a natural extension the paper does not test.","The GP's predictive variance doubles as a coverage measure for the state space, so a direct extension is to use high-uncertainty regions to drive high-level exploration bonuses, converting the regularizer from a passive anchor into an active subgoal-seeking signal.","If the relabeling conjecture is correct, the diffusion-plus-GP module is a drop-in replacement for the high-level actor in any goal-conditioned off-policy HRL scheme, which predicts the gains transfer beyond the TD3/HIRO scaffolding used in the experiments.","The 0.64-versus-0.00 gap on the image-observation FourRooms variant hints that diffusion subgoal generation tolerates high-dimensional observations that adjacency-constrained methods cannot; re-running the comparison on larger visual navigation domains would stress-test that reading."],"forward_implications":["Subgoal generation no longer needs an artificially constrained action space: the diffusion model directly captures complex subgoal distributions, so stability mechanisms like HRAC's adjacency constraint and HIGL's landmarks become unnecessary.","The GP supplies principled uncertainty quantification: the gradient analysis shows GP regularization pulls generated subgoals toward the predictive mean with strength inversely proportional to predictive variance, anchoring learning to well-supported regions.","The reported ablations are roughly additive: replacing the relabeling baseline with diffusion gains about 15%, adding GP regularization gains about 15–16%, and adding the selection rule gains about 7–8% on the hard tasks.","On the image-observation Stochastic Ant FourRooms task, HIDI reaches 0.64 success where HIRO and HRAC score 0.00, so the method scales to settings where adjacency-based subgoal generation collapses.","The selection strategy carries theoretical guarantees: under a near-optimal diffusion assumption its single-step regret is bounded by $\\varepsilon(R^*(s) - R_{\\min}) + (1-\\varepsilon)\\delta$, and a policy-improvement argument shows mixing in the GP mean does not degrade single-step performance."],"supporting_citations":[{"why":"Supplies the HIRO hindsight-relabeling scheme that produces the diffusion model's training targets and the two-layer off-policy HRL setup the method builds on.","marker":"(Nachum et al., 2018)"},{"why":"Supplies the denoising diffusion objective and noise-prediction parameterization used to generate subgoals.","marker":"(Ho et al., 2020)"},{"why":"Supplies the sparse GP pseudo-input approximation that makes uncertainty quantification tractable at replay-buffer scale.","marker":"(Snelson & Ghahramani, 2005)"},{"why":"Supplies the TD3 algorithm used as the actor-critic backbone at both hierarchy levels.","marker":"(Fujimoto et al., 2018)"},{"why":"HRAC baseline whose architecture HIDI follows and whose adjacency-constraint approach HIDI argues against.","marker":"(Zhang et al., 2020)"},{"why":"HIGL baseline; landmark-guided subgoal generation that HIDI compares against and outperforms.","marker":"(Kim et al., 2021)"},{"why":"SAGA baseline; adversarial subgoal generation that HIDI compares against for stability and sample efficiency.","marker":"(Wang et al., 2023a)"},{"why":"HLPS baseline; GP-based probabilistic subgoal-representation method HIDI must beat.","marker":"(Wang et al., 2024)"},{"why":"Supplies hindsight experience replay, the mechanism behind the relabeling that the diffusion training target relies on.","marker":"(Andrychowicz et al., 2017a)"}],"fun_headline_variants":["Diffusion subgoals with GP regularization beat prior HRL methods","Uncertainty-guided diffusion subgoals for hierarchical RL","Hybrid diffusion-GP subgoal selection outperforms in HRL","Subgoal diffusion plus GP uncertainty beats existing HRL","Diffusion model with GP prior improves hierarchical RL"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole pipeline learns from HIRO-style relabeled subgoals, and the paper's Section 2 conjecture — that relabeled subgoals are asymptotically drawn from a distribution approximating an optimal high-level policy — is stated but not proved; if it is false, the diffusion model's target is biased and the GP regularization inherits the same bias.","fun_headline_variants_meta":{"raw":{"variants":["Diffusion subgoals with GP regularization beat prior HRL methods","Uncertainty-guided diffusion subgoals for hierarchical RL","Hybrid diffusion-GP subgoal selection outperforms in HRL","Subgoal diffusion plus GP uncertainty beats existing HRL","Diffusion model with GP prior improves hierarchical RL"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000228,"raw_usage":{"total_tokens":1473,"prompt_tokens":942,"completion_tokens":531,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":558,"completion_tokens_details":{"reasoning_tokens":449}},"tokens_in":558,"tokens_out":531,"duration_ms":5407,"temperature":1.0,"reasoning_tokens":449,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T13:24:04.262313+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Track, during HIDI training on a stochastic Ant Maze task, the divergence between the relabeled-subgoal distribution used to train the diffusion model and the subgoal distribution that maximizes the trained high-level Q-function: if that divergence does not shrink as training progresses, the relabeling conjecture the method rests on is false and the reported gains cannot be credited to learning a near-optimal subgoal distribution.","supporting_citations":[{"cited_title":"Data-efficient hierarchical reinforcement learning","cited_arxiv_id":null,"evidence_quote":"Supplies the HIRO hindsight-relabeling scheme that produces the diffusion model's training targets and the two-layer off-policy HRL setup the method builds on."},{"cited_title":"Addressing function approximation error in actor-critic methods","cited_arxiv_id":null,"evidence_quote":"Supplies the TD3 algorithm used as the actor-critic backbone at both hierarchy levels."},{"cited_title":"Generating adjacency-constrained subgoals in hierarchical reinforcement learning","cited_arxiv_id":null,"evidence_quote":"HRAC baseline whose architecture HIDI follows and whose adjacency-constraint approach HIDI argues against."},{"cited_title":"Landmark-guided subgoal generation in hierarchical reinforcement learning","cited_arxiv_id":null,"evidence_quote":"HIGL baseline; landmark-guided subgoal generation that HIDI compares against and outperforms."},{"cited_title":"a m \\\" a r \\","cited_arxiv_id":null,"evidence_quote":"HLPS baseline; GP-based probabilistic subgoal-representation method HIDI must beat."}],"review_version":1}