Neural Processing Units (NPUs) are critical for AI infrastructure, yet developing kernels remains a bottleneck due to the complexity of vendor-specific Domain-Specific Languages (DSLs). While LLMs excel in general coding, they fail to meet the stringent constraints of NPU development, showing a near-zero success rate on complex kernels in our preliminary study. To address these challenges, we present AscendKernelGen, the first comprehensive framework for NPU kernel development, marking a pioneering effort in this field. This framework consists of three interconnected components: (1) Ascend-CoT, the first dataset in the NPU kernel domain that incorporates chain-of-thought reasoning from real-world kernel implementations; (2) KernelGen-LM, a domain-adaptive model trained on this novel dataset using supervised fine-tuning and reinforcement learning; and (3) NPUKernelBench, the first benchmark platform designed to evaluate the compilation, correctness, and performance of generated NPU kernels. Experimental results demonstrate that our approach dramatically bridges the gap in hardware-specific coding: compilation success on complex Level-2 kernels improves from 0% to 95.5% (Pass@10), with 64% functional correctness. AscendKernGen is available at AscendKernGen and NPUKernelBench.

Diffusion Language Models (DLMs) have emerged as a promising alternative to autoregressive language models, showing strong non-autoregressive generation abilities. Although the masked denoising objective trains DLMs to recover tokens at arbitrary positions, current long-context usage still mostly follows a prefix-based paradigm: the full context is given as a prefix, and the answer is generated only at the end, leaving the model’s infilling ability underutilized. To investigate whether DLMs can read long contexts by infilling, we introduce Parallel Infilling Reading (PIR), a training-free method that inserts extraction slots into long contexts to generate intermediate reading states. Our empirical study shows that naive PIR improves over prefix-based DLM inference, but suffers from slot repetition, where multiple slots generate redundant evidence. To mitigate this issue, we propose Slot-Isolated PIR, which blocks direct slot-to-slot attention, and PIR-SFT, which uses fine-grained evidence supervision to teach DLMs effective evidence extraction while reading. Experiments on long-context question answering benchmarks show that our methods improve answer accuracy and generation efficiency, with increasingly larger gains as the context length grows.