Official Review · ICLR 2026

This paper proposes a method for learning decoupled state-to-state and state-to-goal representations for goal-conditioned reinforcement learning. While the idea is interesting and potentially valuable, the paper suffers from very poor presentation, missing key baseline comparisons, unclear notation, and numerous writing errors. Several claims are unsupported, and the experimental results lack statistical analysis. Overall, the work needs substantial rewriting and stronger empirical validation before it can be considered for acceptance. The paper presents an interesting idea for learning decoupled state-to-state and state-to-goal representations for goal-conditioned reinforcement learning (RL). This is a promising perspective and is shown to perform better than the baselines. **Presentation** - The paper contains numerous typos, incomplete sentences, and inconsistent or incorrect notation. - In addition, several symbols are used before being introduced (or never introduced at all). For example, $p_g$ used in equation (2) and $B$ used in equation (6) are never defined, $\phi$ used in equation (3) is introduced only later. Overall, these issues make the paper extremely confusing to read. **Missing baselines** - Several important baselines are missing from the comparisons, such as CURL: Contrastive Unsupervised Representations for Reinforcement Learning, HIQL: Offline Goal-Conditioned RL with Latent States as Actions, Contrastive Learning as Goal-Conditioned Reinforcement Learning, and Offline Goal-Conditioned Reinforcement Learning via f-Advantage Regression, among others. - The paper ignores these baselines entirely and only compares against the simplest one, CRL. Even if the proposed approach does not outperform these stronger baselines, their inclusion would still be valuable, at least in the appendix. **Unsupported claims** - The paper also includes unsupported statements such as “This setting is beneficial to enhance the generalization of RL and is meaningful to sim2real,” and “In some papers, this sparse reward function is equal to…”. Such claims should be substantiated with references or evidence. - Finally, the authors do not report the statistical significance of their results. There are no standard deviation plots, min–max regions, or p-value analyses, which raises concerns about the robustness of the reported performance. - Why did the authors choose the baselines they did, and why not include more relevant ones?