In recent years, multimodal large language models have achieved substantial progress, yet their performance remains limited in specialized domains such as circuit diagram understanding. A key bottleneck lies in the lack of diverse and large-scale training data tailored to the structural characteristics of circuit diagrams, together with the high cost of expert annotation. To address this challenge, we propose CircuitPile, a systematic data synthesis framework for circuit visual-language modeling. CircuitPile integrates heterogeneous circuit resources from multiple sources and constructs standardized label spaces and multi-granularity visual representations. Based on this framework, we build 738K domain-specific pre-training samples and 100K task-oriented instruction-tuning samples for the pre-training and supervised fine-tuning stages, respectively. Furthermore, to systematically evaluate fundamental visually grounded circuit perception, we introduce CircuitPerception-Bench, a diagnostic benchmark comprising 2,700 multiple-choice samples across three dimensions: component recognition, spatial understanding, and quantitative reasoning. CircuitPerception-Bench focuses on fundamental visually grounded perception and complements external benchmarks such as AMS Benchmark. Experimental results show that multimodal models trained with CircuitPile achieve significant performance gains on the internal benchmark and demonstrate favorable transferability on the external public AMS Benchmark. For example, applying CircuitPile-based SFT to Qwen3-VL-8B improves the scores on CircuitPerception-Bench and AMS Benchmark from 0.5759 and 0.5067 to 0.7703 and 0.5988, respectively.

Language models can be steered by modifying their internal representations to control concepts such as emotion, style, or truthfulness in generation. However, the conditions for an effective intervention remain unclear and are often validated through heuristics and trial-and-error. To fill this gap, we demonstrate that intervention efficacy, measured by linear steerability (i.e., the ability to adjust output via linear transformations of hidden states), emerges during intermediate stages of training. Moreover, even closely related concepts (e.g., anger and sadness) exhibit steerability emergence at distinct stages of training*.To better interpret the dynamics of steerability during training, we adapt existing intervention techniques into a unified framework, referred to as the “Intervention Detector” (ID), which is designed to reveal how linear steerability evolves over the course of training through hidden state and representation analysis. ID reveals that concepts become increasingly linearly separable in the hidden space as training progresses, which strongly correlates with the emergence of linear steerability. We further introduce ID-based metrics, such as heatmaps, entropy trends, and cosine similarity, to help interpret how linear steerability evolves throughout training. In addition, we apply ID across different model families to ensure the generality of our findings on steerability dynamics.