Unverified50% confidenceFactExact time
Matched intervention applies perturbations of the same dimension and similar norm to two conditions, changing only the direction in which the perturbation lies, thereby separating the effect of perturbation magnitude from that of direction/position.
1
Sources
50%
Confidence
Long-term
Relevance
11/18/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedMatched intervention experiments show that perturbing the answer subspace significantly changes model predictions, whereas perturbing the disagreement representation (even with the same dimensions) barely changes predictions, with an effect similar to a random subspace.71% similarUnverifiedThe fixed-point score matching framework has limitations: the score is inherently a local quantity, and when the modes of the target distribution are far apart, the method struggles to correctly weigh the relative weights of the modes.68% similarUnverifiedCross-modal disagreement can be induced by misaligning modality-specific answer representations: perturbing the answer subspace of only a single modality is enough to induce disagreement, whereas applying the same perturbation to both modalities largely preserves consistency.67% similarUnverifiedLinear weighted fusion directly assigns different weights to the similarity scores of two retrieval streams and adds them together, but requires that the score distributions of both streams be similar; otherwise, additional score alignment processing is needed.66% similarUnverifiedContrastive activation probing uses cosine similarity to measure the directional deviation of intermediate-layer activations between the base model and the fine-tuned model under the same probe prompts, with lower similarity indicating greater drift.66% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1182940API
curl https://kongchang.com/api/v1/knowledge/claims/1182940MCP
get_claim(id=1182940)