一句话：将可解释性作为一等编译原语，基于 MLIR/JAX 统一稀疏自编码器、归因与策略引导，消除冗余计算。
现代 AI 可解释性工具彼此孤立，每个工具都需要单独执行前向传播，带来显著延迟，使实时 AI 安全审计难以落地。本文提出 IAM（Interpretable AI with MLIR），把可解释性视为一等编译原语，在 JAX 的垂直集成架构上构建统一的编译器集成框架。作者为稀疏自编码器、归因与策略引导定义自定义 MLIR dialect，并借助 XLA 的算子融合与内存规划进行编译优化，从而消除冗余计算。框架还引入 skip-transcoder 架构以支持实时引导，同时维持模型精度。该工作为通过严格编译抽象来工程化可解释性、构建可信 AI 系统奠定基础。
Modern AI interpretability tools operate in isolation, each requiring separate forward passes and adding significant latency, making real-time AI safety auditing impractical. This research addresses fragmentation by treating interpretability as a first-class compilation primitive. IAM (Interpretable AI with MLIR) unifies fragmented tools into a single compiler-integrated framework built on JAX’s vertically integrated architecture. By defining custom MLIR dialects for sparse autoencoders, attribution, and policy steering, and leveraging compiler optimization through XLA’s fusion and memory planning, IAM eliminates redundant computation. The framework introduces skip-transcoder architectures for real-time steering while maintaining model accuracy. This work establishes the foundation for trustworthy AI systems where interpretability is engineered through rigorous compiler abstractions.