GMM-Optimal-Transport steering (CHaRS)

methods / Causal Validation / Causal interventions (steering) / GMM-Optimal-Transport steering (CHaRS)

Techniqueadvanced

Models a concept's activations as a Gaussian mixture (not a single centroid), clusters source and target activations into semantic sub-groups, then steers by an entropy-regularized (Sinkhorn) discrete optimal-transport map between cluster centroids, applied as a kernel-weighted barycentric combination of per-cluster difference-in-means translations — reducing to plain diff-in-means steering only in the degenerate one-cluster case.

Used in (1 observation)

Concept classes are multi-cluster; GMM-OT steering beats single directionsmeasured

structure: Concept Cluster Heterogeneity · models: Gemma-2-9B-it, Llama-3.1-8B-Instruct, Llama-3.2-3B-Instruct, Qwen2.5-3B-Instruct, Qwen2.5-7B-Instruct, Qwen2.5-14B-Instruct, Qwen2.5-32B, Gemma-2-2B, Llama-3-8B, FLUX.1, Mistral-7B · paper: Concept Heterogeneity-aware Representation Steering