Nils A. Herrmann, Leander Girrbach, Kirill Bykov, Zeynep Akata
原文摘要
Recent work shows that large language models (LLMs) encode behavioural traits ("personas") as linear directions in activation space, often called "persona vectors". Prior work has used such directions as static handles for behavioural steering. We instead treat them as dynamic signals: probes we can monitor and intervene on as reasoning unfolds. We use the term polylogue to denote the time series of alignments between persona vectors and hidden activations over the course of generation. Experiments across four open-weight models show that polylogue features predict correctness on MMLU-Pro on par with low-dimensional activation baselines, while remaining interpretable through their associated persona directions. They also suggest concrete steering targets, namely which latent directions to modulate at different stages of a response. We instantiate this as a simple paragraph-conditioned intervention that improves accuracy on three of four models but degrades the fourth, suggesting that stage-aware latent steering is possible but not yet robust. Together, this positions the polylogue as an interpretable tool for reasoning-time monitoring and intervention.
Leander Girrbach, Alexander von Recum, Kirill Bykov, Zeynep Akata
原文摘要
Detecting whether and in particular how two data distributions differ is a fundamental statistical problem. For example, we might be interested in which concepts (expressed in natural language) differ between two classes, the outputs of different generative models, or the same model with different prompt specifications. When working with large and complex image datasets, we intuitively think that first selecting concepts from a broad list using fine-grained patch embeddings, and then optimizing the concepts to better fit the data is an effective workflow. However, prior work instead uses MLLMs to propose concepts and image-level embeddings to select them. Therefore, we systematically compare these techniques and find that selecting concepts from broad lists can replace MLLM-proposals at obviously lower cost. MLLM inference can then be spent on optimizing selected concepts instead to achieve substantial improvements. Additionally, by designing a new benchmark that focuses on low-prevalence and multi-concept distinctions, we find that selecting nonredundant concepts instead of only the most distinguishing concepts matters, and we introduce techniques for how to effectively do this. Finally, we show how our methods address important problems like detecting spurious correlations and biases. Overall, our work presents a fundamental perspective shift on how we analyze interpretable distributional differences between datasets.