Sinie van der Ben

Sinie van der Ben, Raphaël Baur, Yannick Metz et al.

Recent work identified emotion vectors in Claude Sonnet 4.5, which are internal representations that encode emotion concepts, causally influence behavior, and exhibit geometry mirroring human psychological structure. We test the generality of these findings in two open-weight models, Apertus-8B-Instruct-2509 and Gemma-4-E4B-it, extracting emotion contrast vectors across all layers, using two model-generated corpora. We recover valence geometry for both models, with peak PC1--valence correlations of $r = 0.76$ and $r = 0.83$, approaching the $r = 0.81$ reported for Claude.Beyond replication, we observe notable differences in how valence representations emerge across model depth. In Gemma-4-E4B-it, valence is strongly encoded in early layers but collapses towards later layers, whereas Apertus-8B-Instruct-2509 exhibits the opposite pattern, with valence representations absent in early layers, but emerging at mid depths. Arousal encoding, in contrast, is sensitive to the extraction corpus: both models show stronger PC2--arousal alignment with Gemma-generated stories ($r$ up to $0.45$) than Apertus-generated ones ($r \leq 0.21$), suggesting arousal-relevant cues are unevenly distributed across generated corpora. We open-source our experiment code and dataset for reproducible investigation of emotion representations across language model architectures.

Sinie van der Ben, Neele Roch, Anna Hedström et al.

Cross-paper comparison of sparse autoencoder (SAE) interpretability often relies on autointerpretability scores. In this evaluation pipeline, a language model (LM) explains each feature, and another LM scores the explanation. For these comparisons to be meaningful, scores must reflect stable properties of the features rather than confounding aspects of the evaluation pipeline. Through systematic experiments across four metrics (simulation, detection, fuzzing, purity), two models (Pythia-160M, Apertus-8B), and four axes of methodological variation, we show that this assumption does not hold. Specifically, we find that R1) methodological variance collectively exceeds architectural variance across all metrics and tested models; R2) each metric exhibits a distinct instability profile, with detection being the most stable and fuzzing unreliable across all conditions; R3) top-k feature rankings do not stay consistent across corpus and draw conditions, masking per-feature instability behind stable mean scores; a failure that cannot be detected by monitoring explanation similarity alone. These findings suggest that cross-paper comparisons based on autointerpretability scores may reflect pipeline differences rather than architectural differences, with implications for the ongoing debate on SAE utility. More broadly, unreliable evaluation slows progress in interpretability research at a time when reliable tools for understanding AI systems are needed. To support evaluation, we contribute a variance decomposition approach, a Stability Check, and a Minimum Reporting Checklist.

Xinyuan Yan, Rita Sevastjanova, Sinie van der Ben et al.

Large language models (LLMs) produce high-dimensional embeddings that capture rich semantic and syntactic relationships between words, sentences, and concepts. Investigating the topological structures of LLM embedding spaces via mapper graphs enables us to understand their underlying structures. Specifically, a mapper graph summarizes the topological structure of the embedding space, where each node represents a topological neighborhood (containing a cluster of embeddings), and an edge connects two nodes if their corresponding neighborhoods overlap. However, manually exploring these embedding spaces to uncover encoded linguistic properties requires considerable human effort. To address this challenge, we introduce a framework for semi-automatic annotation of these embedding properties. To organize the exploration process, we first define a taxonomy of explorable elements within a mapper graph such as nodes, edges, paths, components, and trajectories. The annotation of these elements is executed through two types of customizable LLM-based agents that employ perturbation techniques for scalable and automated analysis. These agents help to explore and explain the characteristics of mapper elements and verify the robustness of the generated explanations. We instantiate the framework within a visual analytics workspace and demonstrate its effectiveness through case studies. In particular, we replicate findings from prior research on BERT's embedding properties across various layers of its architecture and provide further observations into the linguistic properties of topological neighborhoods.