Unverified50% confidenceFactExact time
Magnitude-normalized causal subspace steering first projects activations onto the causal subspace and then scales by the magnitude of the current activation, so that the intervention magnitude matches the actual energy distribution.
1
Sources
50%
Confidence
Long-term
Relevance
11/13/2026
First Seen
Sources
Related Entities
Related Claims
Unverified通过向前向传播激活中添加或减去特定方向向量,可以实时调节模型在该概念维度上的表现,无需修改权重参数66% similarUnverifiedThe method constructs quadratic models along push-forward curves, characterizing the curvature of energy in measure space using explicit self-adjoint second-order variational operators.65% similarUnverifiedStBR constructs a structured latent space that reflects transition dynamics and proves that divergence in this space is a theoretically bounded surrogate for the divergence of the true consequence distribution.64% similarUnverifiedProposition 1 proves that if the system is decomposable and each model is perfectly trained on isolated components with the context factor responsible for the remaining energy terms, the joint distribution converges to the true physical distribution in the zero-time limit63% similarUnverifiedSteering changes model behavior by adding a steering vector to the activation vector of a specific layer during the forward pass, and this vector is usually computed from the activation difference between contrastive positive and negative examples.63% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1165385API
curl https://kongchang.com/api/v1/knowledge/claims/1165385MCP
get_claim(id=1165385)