Abstract:Feature steering has emerged as a promising approach for controlling LLM behavior through direct manipulation of internal representations, offering advantages over prompt engineering. However, its practical effectiveness in real-world applications remains poorly understood, particularly regarding potential trade-offs with output quality. We show that feature steering methods substantially degrade model performance even when successfully controlling target behaviors, a critical trade-off. Specifically, we evaluate Goodfire's Auto Steer against prompt engineering baselines across 14 steering queries (covering innocuous and safety-relevant behaviors) on 171 Massive Multitask Language Understanding (MMLU) questions using Llama-8B and Llama-70B, measuring accuracy, coherence, and behavioral control. Our findings show that Auto Steer successfully modifies target behaviors (achieving scores of 3.33 vs. 2.98 for prompting on Llama-8B and 3.57 vs. 3.10 on Llama-70B), but causes dramatic performance degradation: accuracy on the MMLU questions drops from 66% to 46% on Llama-8B and 87% to 73% on Llama-70B, with coherence falling from 4.62 to 2.24 and 4.94 to 3.89 respectively. Simple prompting achieves the best overall balance. These findings highlight limitations of current feature steering methods for practical deployment where task performance cannot be sacrificed. More broadly, our work demonstrates that mechanistic control methods face fundamental capability-behavior trade-offs that must be empirically characterized before deployment.
Abstract:This paper presents the first comprehensive interpretability analysis of a Transformer-based Sign Language Translation (SLT) model, focusing on the translation from video-based Greek Sign Language to glosses and text. Leveraging the Greek Sign Language Dataset, we examine the attention mechanisms within the model to understand how it processes and aligns visual input with sequential glosses. Our analysis reveals that the model pays attention to clusters of frames rather than individual ones, with a diagonal alignment pattern emerging between poses and glosses, which becomes less distinct as the number of glosses increases. We also explore the relative contributions of cross-attention and self-attention at each decoding step, finding that the model initially relies on video frames but shifts its focus to previously predicted tokens as the translation progresses. This work contributes to a deeper understanding of SLT models, paving the way for the development of more transparent and reliable translation systems essential for real-world applications.