Mechanistic Interpretability (MI) has emerged as a vital approach to demystify the opaque decision-making of Large Language Models (LLMs). However, existing reviews primarily treat MI as an observational science, summarizing analytical insights while lacking a systematic framework for actionable intervention. To bridge this gap, we present a practical survey structured around the pipeline: "Locate, Steer, and Improve." We formally categorize Localizing (diagnosis) and Steering (intervention) methods based on specific Interpretable Objects to establish a rigorous intervention protocol. Furthermore, we demonstrate how this framework enables tangible improvements in Alignment, Capability, and Efficiency, effectively operationalizing MI as an actionable methodology for model optimization. With actionable mechanistic interpretability evolving at a fast pace, we pledge to keep this survey up to date, ensuring it reflects the cutting-edge advances in this area.

Key-Value (KV) cache quantization has become a widely adopted optimization technique for efficient large language models (LLMs) inference by reducing KV cache memory usage and mitigating memory-bound constraints. Recent studies have emphasized the importance of preserving the original precision of KVs for the first few tokens to ensure the protection of attention sinks. While this approach has proven effective in mitigating performance degradation, its underlying principles remain insufficiently understood. Moreover, it fails to address the recent discovery that attention sinks can emerge beyond the initial token positions. In this work, we elucidate the underlying mechanisms of attention sinks during inference by examining their role in the cross-layer evolution of extreme activation outliers. Additionally, we provide a comprehensive analysis of the interplay between attention sinks and KV cache quantization. Based on our enhanced understanding, we introduce \textit{\textbf{KVSink}}, a plug-and-play method that effectively predicts sink tokens with negligible overhead, enabling more thorough preservation. Extensive experiments demonstrate that KVSink outperforms the existing Preserve-First-N (PFN) strategy, offering more effective preservation of attention sinks during KV cache quantization. Moreover, when applied to the well-established KVQuant method, KVSink further improves perplexity (PPL) and reduces reliance on 16-bit numerical outliers.

Low-precision rollouts can reduce the cost of reinforcement learning for large language models, but numerical discrepancies between the sampler and learner can compromise policy optimization. Even with synchronized parameters and identical token histories, the two execution paths may assign different token probabilities, altering gradient estimation and update decisions. This survey examines how training–rollout mismatch arises, propagates through policy optimization, and evolves over training. We develop a component-level taxonomy that organizes existing approaches by intervention target into two complementary families: numerical design, which aligns execution across learner and sampler paths or preserves numerical fidelity and update visibility; and learning under residual mismatch, which corrects gradient estimation or controls policy updates. We compare the mechanisms, information requirements, and residual errors of these interventions to explain their limitations and how they can be combined. We further synthesize measurement practices and experimental evidence across precision regimes, architectures, and training configurations, examining how numerical discrepancies affect optimization stability, policy improvement, and computational cost. We clarify how differences in policy pairs, metric definitions, and training horizons shape the interpretation of reported results. Open challenges include predicting harmful mismatch, reducing mitigation overhead, sustaining policy improvement over longer training, and comparing methods at matched quality or end-to-end cost.