Steered Generation via Gradient-Based Optimization on Sparse Query Features

In three linesPrototype-Based Sparse Steering applies Sparse Autoencoders to LLM attention query activations to decompose representations into interpretable features. Gradient-based optimization during inference aligns sparse representations with target behavior prototypes. Validated on Textualized Gridworld (planning constraints) and educational domain (cognitive complexity via Bloom's Taxonomy).

Summary generated by Claude — human-verified