SwordBench: New Benchmark Reveals Hidden Costs of Steering AI Vision

A team of researchers from (likely) Warsaw University of Technology has released SwordBench, a new benchmark designed to evaluate the orthogonality of steering image representations in vision models. Steering—intervening on model representations at inference time to correct biases or concepts—is critical for AI safety and interpretability, but existing evaluation protocols are limited to ambiguous language tasks. SwordBench fills this gap by providing a unified suite for testing steering across multiple vision backbones and concept removal tasks.

The benchmark introduces two key evaluation concepts: cross-concept robustness, which measures how stable concept detection remains when inputs are orthogonalized against unrelated concepts, and collateral damage, which quantifies whether steering inadvertently harms downstream task performance on bias-free inputs. Testing with linear SVMs, sparse autoencoders, and optimization-based methods on several backbones, the authors found that linear SVMs exhibit superior separability and orthogonality but fail to achieve zero collateral damage, often trailing sparse autoencoders. In simpler regimes, even standard and optimization-based methods fail to achieve perfect steering, highlighting the need for more robust techniques. The source code is expected on GitHub soon.