DAGR: State-Conditioned Goal Representations via Difference-Aware Goal Cross-Attention
Abstract
Goal-conditioned reinforcement learning hinges on how the goal is encoded. Contrastive, metric, temporal-distance and information-theoretic encoders disagree on the objective. They agree on one thing. None of them sees the current state, so the embedding cannot mark which part of the goal still needs action, and the policy must recover that cue by inverting both encoders. We propose DAGR, which refines the static embedding of any late-fusion encoder into a state-conditioned one through multi-scale gated cross-attention. A gated residual holds the refinement near the base, and a difference-aware attention rule biases the scores by a per-token state-goal mismatch. A single condition decides what such a refinement can guarantee, namely whether the block returns its input at closed gates. We prove that the usual post-norm placement violates it, measure the consequence on frozen checkpoints, and recover part of the resulting loss by restoring the condition. On OGBench DAGR improves navigation and matches or trails the base elsewhere. Our ablations trace the gain to the gated residual rather than to the difference bias that names the method. Code is available at https://github.com/leixingxing1/DAGR
Related
- Multiple representations in difference sets under polynomial density constraints
- Nodal obstruction to conditioned-diffusion representations of the weak momentum
- Finite-State Time-Inhomogeneous Markov Chains Conditioned to Survive
- When Should a World Model Move? Loss-Conditioned State Execution
- Learning Topology-Aware Representations for Mechanics-Relevant TPMS Metamaterials
BibTeX
@misc{lei2026dagr,
title = {{DAGR: State-Conditioned Goal Representations via Difference-Aware Goal Cross-Attention}},
author = {Xing Lei and Wenyan Yang and Xuetao Zhang and Donglin Wang},
howpublished = {arXiv},
year = {2026},
month = oct,
doi = {10.48550/arxiv.2607.13731},
eprint = {2607.13731},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2607.13731},
}
RIS
TY - UNPB
TI - DAGR: State-Conditioned Goal Representations via Difference-Aware Goal Cross-Attention
AU - Xing Lei
AU - Wenyan Yang
AU - Xuetao Zhang
AU - Donglin Wang
PY - 2026
DA - 2026/10/02
PB - arXiv
DO - 10.48550/arxiv.2607.13731
UR - https://arxiv.org/abs/2607.13731
AB - Goal-conditioned reinforcement learning hinges on how the goal is encoded. Contrastive, metric, temporal-distance and information-theoretic encoders disagree on the objective. They agree on one thing. None of them sees the current state, so the embedding cannot mark which part of the goal still needs action, and the policy must recover that cue by inverting both encoders. We propose DAGR, which refines the static embedding of any late-fusion encoder into a state-conditioned one through multi-scale gated cross-attention. A gated residual holds the refinement near the base, and a difference-aware attention rule biases the scores by a per-token state-goal mismatch. A single condition decides what such a refinement can guarantee, namely whether the block returns its input at closed gates. We prove that the usual post-norm placement violates it, measure the consequence on frozen checkpoints, and recover part of the resulting loss by restoring the condition. On OGBench DAGR improves navigation and matches or trails the base elsewhere. Our ablations trace the gain to the gated residual rather than to the difference bias that names the method. Code is available at https://github.com/leixingxing1/DAGR
ER -