acceptodds
Under review as a conference paper at ICLR 2027

Compositional Generalization via Geometric Abstractions in Sequential Decision-Making

Abstract

Compositional generalization, the ability to recombine known components in new settings, in sequential decision-making requires identifying which parts of earlier rollouts remain useful for new tasks. Existing methods often consider reusing temporally extended skills and predictive models, but not rich local geometry and dynamics. We propose a geometric abstraction to quantify local transition geometry and dynamics and enable efficient compositional generalization in sequential decision-making. The proposed geometric descriptor aggregates both first- and second-order statistics of lifted one-step transitions within a trajectory segment in a positive semidefinite cone. We show that these descriptors make shared structure explicit, distinguish algebraic composition in an abstract representation space, and reveal opportunities for generalization. In particular, we prove that they are well defined up to coordinate gauge, complete for the induced low-order additive signal class, additive under valid segment composition, and minimally sufficient among admissible additive descriptors. Under explicit matrix-sufficiency, Bellman-compatibility, and local-smoothness assumptions, we further derive a local first-order approximation of action-value variation in matrix space, motivating reuse of matrix increments learned from existing geometry and dynamics. The proposed matrix-to-value mapping is plug-in compatible with standard model-free and model-based baselines to bootstrap value-function learning in new tasks, while obstruction filtering rejects implausible compositions. Empirically, MSRL achieves the best average finite-budget target AUC (0.73), improving over MSRL from scratch (0.65), the strongest source-initialized transfer baseline TD-MPC-PT+FT (0.63), and the strongest from-scratch baseline TD-MPC (0.57).

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
Reinforcement Learning, Compositional Generalization, Geometric Representation Learning

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_ddYMOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a positive semidefinite matrix representation for trajectory segments, which aggregates local state changes, actions, rewards, and related information into matrix descriptors, while using matrix addition to compose trajectory segments. Based on this representation, the authors propose MSRL, which pretrains a matrix-to-value branch on source tasks and combines segment retrieval with obstruction filtering for policy learning on target tasks.

Critical strengths
  1. The matrix-based representation of trajectory segments has some novelty, and the overall method is relatively unified in its design.

  2. The experiments include negative controls such as random initialization, shuffled source value, and unaligned lift, which help assess whether the observed gains actually come from meaningful pretraining.

Critical weaknesses
  1. The composition property essentially follows from the additivity of M(τ)=∑tψtψt⊤,M(\tau)=\sum_t \psi_t\psi_t^\top, while completeness is established only for the specifically defined affine-quadratic function class Faq.F_{aq}. The obstruction theorem is also largely equivalent to stating that a point does not belong to the reachable set if its distance to that set is greater than zero. These results are formally complete, but they do not provide sufficiently strong guarantees on RL performance or generalization.

  2. The experiments explicitly use a simplified ψexp\psi^{exp} and remove the Kronecker action-position blocks from the theoretical definition. In addition, the segment length in the target-control library is actually fixed to 1 transition. Therefore, the multi-step segment composition, gluing, and related theory emphasized in the main text are not directly validated in the main control experiments.

  3. The paper states that it reports final performance, success rate, and steps-to-threshold, but I did not find the corresponding complete results.

  4. The main source-transfer comparisons include Traj2Value-RNN-FT, Traj2Value-Transformer-FT, and Subgraph-GNN-to-Value-FT, which are adapted variants constructed by the authors rather than standard algorithms from prior work.

  5. There are multiple clear editing and formatting issues in both the references and the main text. For example, the HER authors are listed as “OpenAI Pieter Abbeel,” which is clearly incorrect. Conference and journal names are also inconsistent, such as “International Conference on Machine Learning / International conference on machine learning,” “PMLR / Pmlr,” and “Nature / nature.” The Ravindran (2003) entry also lacks complete publication information. Appendix E.1 further contains nearly duplicated environment-description paragraphs.

Questions for authors

See Weaknesses.

2 comments

Response to the mathematical contribution and its control implications.

We would like to clarify the distinct roles of the structural results and the empirical control evaluation. The former characterize the descriptor under explicit assumptions rather than establishing unconditional RL improvement. The contribution is an explicit representation in which useful operations on observed transitions are exact and compatible: prefix accumulation, valid concatenation, and nonduplicating overlap assembly occur in one fixed-dimensional PSD space. Completeness and minimal sufficiency establish a precise information boundary: equality of descriptors is equivalent to equality of every additive affine-quadratic functional induced by the chosen lift, and any other admissible descriptor sufficient for that class must determine MM. This is functional sufficiency, not minimum storage dimension or preservation of arbitrary history.

The value-theoretic results play a different role. Given matrix sufficiency, lifted Bellman compatibility, and differentiability, the matrix-conditioned value admits a local first-order expansion. We do not claim global linearity for finite segments, a guarantee that arbitrary compositions improve return, or universally superior transfer. In fully observed Markov tasks, the optimal value may be independent of history; accordingly, practical finite-sample benefits must come from the learned inductive bias and are an empirical question. Likewise, the exact reachable-set obstruction is a mathematical criterion; its learned implementation gωg_\omega is a heuristic filter, not a certified oracle. The theory states what the representation guarantees and what must additionally hold for the value interface. The experiments separately measure learning performance.

Manuscript revision. Sections 3--4 distinguish exact descriptor properties, conditional value statements, and empirical control outcomes, without promoting a representation identity into an unconditional RL theorem.

Response to multi-step composition and the one-step library.

A key distinction is between the one-step retrieval library and the multi-step history summarized by the critic. The reported library contains one-transition entries, whereas the critic state is Zt=∑s=0t−1ψsψs⊤Z_t=\sum_{s=0}^{t-1}\psi_s\psi_s^\top and therefore aggregates the entire executed prefix. In particular, Zt+1=Zt+ψtψt⊤Z_{t+1}=Z_t+\psi_t\psi_t^\top performs exact repeated composition at every environment step. Thus the evaluated method does use multi-step historical statistics, even though candidate retrieval/lookahead is limited to one transition. A one-step library is a computationally bounded design choice; it is not a test of the performance gains from retrieving arbitrary long segments or performing learned overlap gluing, and we do not claim it is.

Independent source-side checks do provide descriptor-level evidence: GraphMotif directed-motif/reversal retrieval is 0.9390.939, MicroGrid obstruction AUROC is 0.8890.889, and additive-consistency errors are small (Appendix E). These support the structural interpretation, not an unreported multi-step planning experiment. The appropriate conclusion is that the current targets validate prefix accumulation, one-step retrieval, and source initialization, while the more general multi-segment operations are established algebraically.

Manuscript revision. Sections 5--6 and Appendix E now distinguish executed-prefix horizon, retrieval horizon, and general representation-level gluing.

Response to numerical reporting and the baseline design.

We agree that complete final-return, success-rate, and steps-to-threshold tables are not included. The primary reported metric is normalized few-shot AUC (Table 2, five target families and five seeds), complemented by full-budget learning curves for AntMaze, Walker2d, and Hopper (Figure 2). Source-initialized MSRL obtains 0.730.73 mean AUC, compared with 0.650.65 for the matched from-scratch model and 0.620.62 for TD-MPC-PT+FT. We have revised the manuscript to restrict numerical claims to these reported results.

Because success rates and steps-to-threshold require task-specific criteria or thresholds, we do not infer them from the reported AUC values or mean learning curves.

We make no separate superiority claim for these unreported metrics.

We clarify the distinction between published RL algorithms and controlled representation adaptations used in our experiments. PPO, SAC, TD3, MBPO, DreamerV3, and TD-MPC are published RL algorithms. Traj2Value-RNN-FT, Traj2Value-Transformer-FT, and Subgraph-GNN-to-Value-FT are intentionally constructed representation controls, not claimed to be standalone published algorithms. Their role is to hold the source-to-target protocol fixed while varying the representation; their validity as controls does not depend on each variant having been separately published. We clarify the groups and initialized components in the table captions and Appendix E.

Manuscript revision. We align every metric statement with the displayed numerical results and explicitly label controlled representation baselines as such.

Response to citation and editorial accuracy.

We thank the reviewer for identifying these bibliographic and editorial inconsistencies. We have corrected the HER author field, the Ravindran--Barto IJCAI 2003 entry, and inconsistent venue capitalization, and have streamlined the explanatory wording in Appendix E.1.

Reviewer_85XXOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates compositional generalization in reinforcement learning by exploiting shared local geometric structures across different environments. The authors propose Matrix-Space Reinforcement Learning (MSRL), which represents trajectory segments using geometric matrices that capture local transitions, dynamics, and reward information. A matrix-to-value mapping is pretrained on source tasks and transferred to target tasks to improve learning efficiency.

Critical strengths

1.The paper provides an interesting perspective on compositional generalization by exploiting local geometric information rather than relying solely on contextual representations or reusable skills. 2.The proposed matrix representation has clear mathematical properties, supported by theoretical analysis. Experiments across multiple environments demonstrate improvements in sample efficiency.

Critical weaknesses
  1. The validity of Assumption 4.1 requires further justification. The assumption states that the geometric matrix is sufficient for optimal decision-making. However, the matrix is constructed from only a limited set of manually selected geometric features. Since the goal is to improve generalization across environments, it is unclear whether these features can capture sufficient task-relevant information. In contrast, using the complete historical context avoids discarding potentially important information. More detailed experiments and visualizations are needed to justify this assumption.

  2. The proposed representation ignores important temporal ordering information. The matrix mainly captures local geometric statistics, such as positions and displacements. However, execution order is crucial in sequential decision-making. Trajectories containing the same local transitions in different orders can produce identical matrices, even though their decision-making implications may differ significantly. This limitation deserves further investigation.

  3. The transferability of the learned matrix-to-value mapping (F_\eta) is questionable. The authors propose learning (F_\eta) as a shared mapping between geometric matrices and value functions across different tasks. However, identical local transitions may have significantly different values depending on the task objectives, rewards, and global environment structure. Such inconsistencies may make (F_\eta) difficult to train and transfer effectively. Further analysis and experiments are needed to validate this shared value representation.

  4. More comparisons with recent reinforcement learning and domain adaptation methods are needed. Since the paper focuses on cross-environment generalization, the experimental evaluation should include recent representative methods, such as Cross-Domain Policy Optimization via Bellman Consistency and Hybrid Critics (ICLR 2026), Cross-Domain Policy Adaptation by Capturing Representation Mismatch (ICML 2024), Towards Generalizable Reinforcement Learning via Causality-Guided Self-Adaptive Representations (ICLR 2025), and Dual-Robust Cross-Domain Offline Reinforcement Learning Against Dynamics Shifts (ICLR 2026), where the experimental settings are comparable.

Questions for authors

Please refer to the Weaknesses section for my questions and concerns.

4 comments

Response to matrix sufficiency and temporal order.

It is useful to distinguish preservation of the full ordered history from sufficiency of the pair (ot,Zt)(o_t,Z_t) for action values. MSRL uses Qθ(ot,Zt,a)Q_\theta(o_t,Z_t,a), not Qθ(Zt,a)Q_\theta(Z_t,a): the current observation remains available. Assumption 4.1 is a conditional statement about optimal value, not a theorem asserting that hand-selected features reconstruct every environment state. In a fully observed Markov task, Qe⋆(ot,ht,a)=Qe⋆(ot,a)Q_e^\star(o_t,h_t,a)=Q_e^\star(o_t,a) regardless of the matrix, so the condition holds. This does not by itself imply a nontrivial population-value contribution from ZtZ_t; the benefit of an auxiliary matrix branch for finite-sample training and cross-task initialization is assessed empirically. In partially observed tasks, sufficiency genuinely restricts the environments covered by the Bellman analysis.

The existing target-suite comparisons provide empirical context for this representation choice: the protocol-matched Traj2Value-RNN-FT and Traj2Value-Transformer-FT controls obtain mean few-shot AUCs of 0.580.58 and 0.600.60, respectively, compared with 0.730.73 for source-initialized MSRL (Table 2).

These results support the utility of the matrix representation under the evaluated transfer protocol; they are not a proof of matrix sufficiency in arbitrary partially observed tasks or a capacity-matched comparison.

The ordering observation is already a property of the definition: M(τ)=∑tψtψt⊤M(\tau)=\sum_t\psi_t\psi_t^\top is invariant to reordering the same lifted transition multiset. The descriptor is therefore not intended as an injective encoding of chronological order. For a concrete boundary case, take closed segments A=(0→1→0)A=(0\to1\to0) and B=(0→2→0)B=(0\to2\to0). Histories A⋆BA\star B and B⋆AB\star A have identical matrices and final observations. Their optimal values agree in a fully observed Markov task; they may differ if an unobserved variable remembers which segment occurred last and changes subsequent rewards, even when the rewards within both segments are unchanged. Such a partially observed task can violate Assumption 4.1 and is not covered by the exact matrix-Bellman reduction. This example makes explicit a limitation of the conditional Bellman analysis.

Manuscript revision. Section 4 now specifies this scope and includes the closed-segment counterexample, separating the role of the current observation from that of aggregated transition statistics.

Response to value transfer across tasks with different objectives.

MSRL does not posit a task-independent optimal matrix-to-value function. Two tasks can attach different values to the same ZZ because their rewards, dynamics, and objectives differ. We transfer a parameter initialization, ηtgt(0)=ηsrc\eta_{\mathrm{tgt}}^{(0)}=\eta_{\mathrm{src}}, and then optimize the branch with target interactions; the target observation pathway, action head, and policy are also trained in the target task. Equality Vsrc(Z)=Vtgt(Z)V_{\mathrm{src}}(Z)=V_{\mathrm{tgt}}(Z) is neither assumed nor necessary.

The manuscript already reports matched tests of this narrower, concrete claim:

Target initialization/controlMean few-shot AUC
Source-pretrained MSRL0.73\mathbf{0.73}
MSRL from scratch0.650.65
Random branch initialization0.640.64
Shuffled source-value pretraining0.590.59
Unaligned-lift pretraining0.580.58
Observation-only pretraining0.620.62

The 0.080.08 absolute AUC-point gain over the identical architecture trained from scratch supports a practical transfer benefit. The negative controls directly test whether merely performing additional source updates, transferring arbitrary weights, or removing channel alignment produces the same gain; none does in the reported target suite. The best reported source-initialized comparator, TD-MPC-PT+FT, achieves a mean few-shot AUC of 0.620.62. These observations support useful aligned source initialization, not a universal guarantee against negative transfer.

Manuscript revision. Section 5 and the experiments explicitly distinguish parameter transfer from value-function invariance and foreground the matched negative controls.

Response to recent cross-domain baselines.

The suggested papers are relevant but use different transfer interfaces. QAvatar combines source and target critics and is the closest suggested online value-transfer comparison. PAR assumes a dynamics-adaptation setup with shared state, action, and reward spaces; CSR targets adaptive causal representations; DROCO is explicitly cross-domain offline RL. Matching a benchmark name alone would not equalize source data access, reward changes, target interaction budgets, or what is transferred. The submitted evaluation already includes PPO, SAC, TD3, MBPO, DreamerV3, TD-MPC, their applicable source-initialized counterparts, and controlled representation alternatives. These results support the claims within the evaluated transfer protocol, while a direct protocol-matched QAvatar comparison remains outside the reported experiments.

Manuscript revision. The Introduction now discusses all four recent methods, with their protocol differences made explicit. We do not present results for an unrun baseline.

Reviewer_85XXOpenReview ↗

Remaining Concerns on Geometric Representation and Experimental Validation

Thank you for the clarification. However, my main concerns remain insufficiently addressed.

1. Geometric matrix representation. The authors argue that Assumption 4.1 holds in fully observed Markov environments because the current observation is already sufficient for optimal decision-making. However, this does not address my core concern. If the current observation is already sufficient, the geometric matrix is not needed to ensure decision sufficiency. The key motivation of MSRL is to capture generalizable knowledge across environments, making the properties and effectiveness of the geometric matrix particularly important. However, the matrix is constructed from a limited set of manually designed features, without sufficient analysis of why these features can capture transferable knowledge or facilitate decision-making. The current experiments do not adequately justify this representation choice.

2. Temporal ordering. The authors acknowledge that the geometric matrix loses temporal ordering information. This confirms my original concern, and the potential impact on sequential decision-making and generalization remains unresolved.

3. Experimental comparisons. Although the authors explain the differences in transfer protocols, comparisons with recent cross-domain RL and domain adaptation methods are still missing. In particular, QAvatar is acknowledged as a closely related approach, but no direct comparison is provided.

Overall, while the rebuttal clarifies some theoretical assumptions, it does not fully address the fundamental concerns regarding the effectiveness of the geometric representation and the completeness of experimental validation. Therefore, I maintain my original rating.

Response to the follow-up on geometric representation, temporal ordering, and experimental comparisons.

We appreciate the reviewer's follow-up. We would like to clarify the distinction between the theoretical scope of the representation and the empirical evidence for its transferability.

1. Geometric representation and transferability.

The Markov sufficiency argument was intended to clarify the scope of Assumption 4.1, rather than to establish the practical usefulness of the matrix. The latter is a separate question addressed by our representation design and experiments.

The matrix aggregates first- and second-order statistics of state changes, actions, and rewards, including their cross-moments. These features are chosen to expose local transition geometry and action-reward relationships through an additive representation. Completeness is established for the induced affine-quadratic functional class, not for arbitrary task-relevant information.

Importantly, the reported experiments directly examine whether this representation supports meaningful transfer. Across five target families, source-initialized MSRL achieves a mean few-shot AUC of 0.73, compared with 0.65 for the identical architecture trained from scratch, 0.64 with random initialization, 0.59 with shuffled source-value pretraining, and 0.58 with unaligned-lift pretraining. Under the same transfer protocol, the trajectory-based RNN and Transformer controls achieve 0.58 and 0.60, respectively.

These results indicate that the benefit is not explained merely by additional parameters or source pretraining. The alignment and structure of the transferred representation matter. We view this as empirical evidence for a useful inductive bias, rather than a claim of universal representation sufficiency.

2. Temporal ordering.

We agree that the matrix does not preserve arbitrary temporal ordering. This is an explicit property of the representation, whose purpose is to aggregate reusable local transition statistics rather than reconstruct complete histories.

However, MSRL conditions its value function on both the current observation and the accumulated matrix, Qθ(ot,Zt,a)Q_\theta(o_t,Z_t,a). In fully observed Markov environments, the current observation retains the information required for optimal decision-making, while the matrix provides an auxiliary representation for learning and transfer.

Temporal ordering can become essential when unobserved history affects future rewards or dynamics. We explicitly address this boundary through the closed-segment counterexample added to Section 4. Such environments are not covered by the matrix-sufficiency assumption without additional conditions. Our empirical claims are therefore restricted to the evaluated settings, rather than arbitrary partially observed environments.

3. Experimental comparisons and transfer protocols.

We would like to emphasize that a meaningful transfer comparison requires matching the transfer setting, source information, and target interaction budget, rather than grouping methods solely under the broad category of cross-domain RL.

The suggested methods address different transfer mechanisms. QAvatar combines source and target critics, PAR focuses on dynamics adaptation under shared state, action, and reward spaces, CSR develops adaptive causal representations, and DROCO considers cross-domain offline RL. These differences affect the information available to each method and the learning problem being evaluated. Consequently, directly comparing results obtained under their respective original protocols would not isolate the contribution of our geometric representation.

Our evaluation instead uses a controlled transfer protocol with established model-free and model-based algorithms, including PPO, SAC, TD3, MBPO, DreamerV3, and TD-MPC, together with applicable source-initialized variants and representation-specific controls. These comparisons examine both the benefit of source transfer and the additional contribution of the geometric representation under consistent experimental conditions.

The revised Introduction further distinguishes these recent approaches from our setting. Our empirical contribution is the demonstrated improvement in finite-budget target learning through structured geometric transfer, supported by controlled comparisons rather than cross-protocol performance claims.

We hope these clarifications make the representation's intended contribution, empirical support, and remaining evaluation boundaries more precise.

Reviewer_JddEOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper focus on the compositional generalization problem: whether useful segments can be extracted from rollout trajectories collected in simple tasks and reused in more complex tasks to speed up learning. The paper proposes a new approach to represent such segments: each segment of experience is represented as a matrix, so that concatenating experiences corresponds to matrix addition. Transfer across tasks is achieved by pretraining a “matrix → value” module.

Critical strengths

Compositional generalization is a valuable problem. Rather than following the ideas of existing work, the authors designed a new approach to represent “experience segments” and gave several theoretical advantages of this design.

Critical weaknesses
  1. The definition “step lift” and “Matrix Proxy” are the foundation of the paper’s design, yet the motivation and physical meaning behind them are not clearly conveyed. For example, why use the midpoint c_{t}? Why \xi \left(a_{t}\right)\otimes c_{t};\ \xi \left(a_{t}\right)\otimes {\delta }_{t} is focused in the step lift vector?

  2. The authors should organize the paper’s logic from the reader’s perspective. First, the statement that “existing methods reuse temporally extended skills, predictive occupancies, goal-conditioned values, or behaviorally compressed states” should be accompanied by a figure or table, as these methods are not intuitive to understand. Second, current writing order is not friendly for reading. Sections 2-4 presents a series of properties and proofs without explaining why these properties need to be discussed, making it hard to follow.

  3. The experiments use a simplified lift that omits the “action ⊗ position” Kronecker blocks of the theoretical version. The reason should be discussed.

Questions for authors

Please respond to the several questions raised in “Critical Weaknesses”.

1 comment

Response to the geometric design and its compact experimental realization.

The step lift is designed to expose interpretable moments of directed, action-conditioned transitions, rather than to recover an unrestricted trajectory history. The midpoint--displacement coordinates are not arbitrary: with ct=(xt+xt+)/2c_t=(x_t+x_t^+)/2 and δt=xt+−xt\delta_t=x_t^+-x_t, we recover xt=ct−δt/2x_t=c_t-\delta_t/2 and xt+=ct+δt/2x_t^+=c_t+\delta_t/2 exactly. Reversing a transition preserves ctc_t but reverses the sign of δt\delta_t, separating transition location from oriented transport. The blocks ξ(at)⊗ct\xi(a_t)\otimes c_t and ξ(at)⊗δt\xi(a_t)\otimes\delta_t refine this distinction by action; for one-hot actions, their quadratic matrix blocks contain within-action second-order geometry and transport moments. The constant component makes segment length and first-order statistics accessible, and the reward component adds reward--geometry couplings.

The experiments deliberately use ψtexp=[xˉt;Δxˉt;aˉt;rˉt;1]\psi_t^{\mathrm{exp}}=[\bar x_t;\Delta\bar x_t;\bar a_t;\bar r_t;1], not the full tensorized lift. Although the compact lift omits the explicit Kronecker action--position blocks, its outer product still contains the cross-moments ∑taˉtxˉt⊤\sum_t\bar a_t\bar x_t^\top and ∑taˉt(Δxˉt)⊤\sum_t\bar a_t(\Delta\bar x_t)^\top. Moreover, since cˉt=xˉt+Δxˉt/2\bar c_t=\bar x_t+\Delta\bar x_t/2, the action--midpoint cross-moment is recoverable. The compact version omits higher-order within-action quadratic moments, not all action-conditioned geometry. For dx=8d_x=8 and da=10d_a=10, its lift has 2828 dimensions versus 178178 for the full version, reducing dense matrix storage by about 40.4×40.4\times. PSD structure and exact segment addition hold for either lift; completeness is always relative to the function class induced by the chosen lift.

Manuscript revision. We clarify the midpoint, action-tensor blocks, compact-lift trade-off, and precise scope of the inherited properties in Section 2 and Appendix E.1.

Response to the role of structural guarantees and differences from previous approaches.

Each structural property serves a specific purpose in the proposed representation and control interface. Gauge covariance specifies transformation under aligned coordinate changes; completeness and minimal sufficiency identify exactly which lift-induced statistics can be recovered; composition supports prefix updates and stored increments; gluing characterizes overlap correction; and obstruction separates algebraic matrix sums from environment-realizable trajectories. We do not claim that gluing was independently evaluated by the one-step target library, nor that these representation-level facts alone ensure improved RL return. Bellman closure and local value behavior require separately stated assumptions.

MSRL also differs from options, successor features, goal-conditioned methods, and bisimulation in the object being reused. It uses additive transition statistics as a common representation interface and transfers the source-trained matrix-value branch as a target initialization, rather than transferring a fixed skill, discounted occupancy prediction, goal-conditioned policy, or compressed state. This is a distinction in the interface and algebra, not a claim that all prior approaches are interchangeable or inferior.

Manuscript revision. The Introduction includes a compact comparison table, and the theoretical section explicitly maps each guarantee to its intended representation or control operation.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.