Answers are grounded in this paper and cite their sources.
Sparse Autoencoders (SAEs) have emerged as an important tool in mechanistic interpretability, enabling the decomposition of internal activations in large language models (LLMs) into sparse, interpretable feature representations. Despite recent advances in SAE-based analysis, understanding how these features are organized across transformer layers and how individual features respond to diverse semantic contexts remains an important research challenge. In this study, we investigate the internal feature representations of Qwen2.5-3B through two complementary experiments: cross-layer feature correspondence and polysemantic activation analysis. First, we employ the Hungarian algorithm to examine feature activation patterns across 36 transformer layers, identifying potential correspondences between features at different depths. Our analysis reveals a sparse and depth-dependent correspondence structure, with only 17 matches identified among 8,960 potential feature pairs. Second, we investigate the contextual activation behavior of individual SAE features to examine their responses across semantically diverse inputs. In particular, Feature F374 in Layer 33 exhibits substantial activations across sentences associated with different religious and philosophical traditions, including Christian, Jewish, Catholic, and Chinese philosophical contexts. In contrast, its activation is considerably lower on selected technology- and sports-related control texts, suggesting a degree of contextual selectivity. Together, these findings provide an empirical examination of cross-layer feature correspondence and polysemantic activation behavior in Qwen2.5-3B. Our results highlight the sparse nature of feature relationships across transformer depth and demonstrate how individual SAE features can exhibit activation patterns spanning multiple semantic domains. This study contributes to the broader understanding of feature organization, contextual selectivity, and internal representations in large language models.
No takes yet. Share an insight, caveat, or question.
Sahil Adit (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: