Replay sampling is central to off-policy task-oriented dialogue policy learning, as it determines which past interactions are revisited during policy updates. Existing priority-based replay methods often rely on a monolithic priority distribution to form minibatches, which entangles coverage across error magnitudes with emphasis on high-error transitions. Under sparse dialogue success and heavy-tailed temporal-difference (TD) errors, this design overexposes extreme residuals and suppresses informative medium-error transitions, a failure mode we call mid-magnitude starvation. To address this problem, we propose Hierarchical Prioritized Experience Replay (HPER), a two-level replay sampler that separates coverage from prioritization. HPER first samples TD-error quantile strata according to occupancy, ensuring that different error magnitudes enter training, and then applies local prioritized replay within the selected stratum to retain emphasis on large residuals. This coverage-first design restores intermediate learning signals without diluting high-error samples. Experiments on multiple dialogue datasets show that HPER improves replay efficiency and yields faster competence onset and higher terminal success than uniform replay and monolithic priority-based replay under matched budgets.

Embedding-as-a-Service (EaaS) has emerged as a critical paradigm for commercializing large language models (LLMs). However, existing backdoor watermarking techniques are fundamentally limited to “zero-bit” detection, which prevents user-level traceability in multi-user EaaS scenarios. To address these limitations, we propose RShield, a multi-bit backdoor watermarking that enables reliable user-level attribution of LLMs for EaaS under model extraction attacks. RShield integrates Reed-Solomon error-correcting codes with orthogonal feature mapping to introduce highly-structured redundancy, constructing fault-tolerant symbol sequences for multi-bit watermark space, thereby staying recoverable even after aggressive extraction noise condition.To mitigate semantic distortion under the interference of noise channel, RShield employs a lightweight Adapter to adaptively inject multi-bit watermarks in the feature space, preserving the quality of EaaS while achieving a user-level traceability.Extensive experiments on four NLP benchmarks demonstrate that RShield efficiently achieves 100% multi-bit watermark recovery and high semantic fidelity under model extraction attacks compared to existing methods, while significantly reducing the degradation of watermarking on downstream task performance.