Stephen Cheng

Stephen Cheng, Sarah Wiegreffe, Dinesh Manocha

Applying steering vectors to large language models (LLMs) is an efficient and effective model alignment technique, but we lack an interpretable explanation for how it works-- specifically, what internal mechanisms steering vectors affect and how this results in different model outputs. To investigate the causal mechanisms underlying the effectiveness of steering vectors, we conduct a comprehensive case study on refusal. We propose a multi-token activation patching framework and discover that different steering methodologies leverage functionally interchangeable circuits when applied at the same layer. These circuits reveal that steering vectors primarily interact with the attention mechanism through the OV circuit while largely ignoring the QK circuit-- freezing all attention scores during steering drops performance by only 8.75% across two model families. A mathematical decomposition of the steered OV circuit further reveals semantically interpretable concepts, even in cases where the steering vector itself does not. Leveraging the activation patching results, we show that steering vectors can be sparsified by up to 90-99% while retaining most performance, and that different steering methodologies agree on a subset of important dimensions.

Samarth Chopra, Jing Liang, Gershom Seneviratne et al.

We present Splatblox, a real-time system for autonomous navigation in outdoor environments with dense vegetation, irregular obstacles, and complex terrain. Our method fuses segmented RGB images and LiDAR point clouds using Gaussian Splatting to construct a traversability-aware Euclidean Signed Distance Field (ESDF) that jointly encodes geometry and semantics. Updated online, this field enables semantic reasoning to distinguish traversable vegetation (e.g., tall grass) from rigid obstacles (e.g., trees), while LiDAR ensures 360-degree geometric coverage for extended planning horizons. We validate Splatblox on a quadruped robot and demonstrate transfer to a wheeled platform. In field trials across vegetation-rich scenarios, it outperforms state-of-the-art methods with over 50% higher success rate, 40% fewer freezing incidents, 5% shorter paths, and up to 13% faster time to goal, while supporting long-range missions up to 100 meters. Experiment videos and more details can be found on our project page: https://splatblox.github.io