Answer Subspace
A low-dimensional subspace in the activation space of a given model layer that is spanned by the answer directions of multiple samples. It is used to separate answer-related components from the remaining representations and serves as a tool for mechanistic interpretability analysis.
Timeline (last 90 days)
A recent study proposes estimating a low-dimensional answer subspace from answer-position representations, separating answer-related components from the rest of the representation.
Experiments show that the information predicting cross-modal inconsistency lies mainly outside the answer subspace, concentrated in a reproducible two-dimensional representation, and that this representation cannot be explained by fixed modality offsets.
Matched 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.
Cross-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.
Improving cross-modal consistency may require interventions targeting modality alignment within the answer subspace, rather than merely adjusting high-level fusion strategies.
The answer subspace refers to a low-dimensional subspace in the activation space of a given layer of the model (usually the position corresponding to the answer token) spanned by the answer directions of multiple input-output samples.
All Facts (6)
A recent study proposes estimating a low-dimensional answer subspace from answer-position representations, separating answer-related components from the rest of the representation.
50%UnverifiedExperiments show that the information predicting cross-modal inconsistency lies mainly outside the answer subspace, concentrated in a reproducible two-dimensional representation, and that this representation cannot be explained by fixed modality offsets.
50%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.
50%UnverifiedCross-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.
50%UnverifiedImproving cross-modal consistency may require interventions targeting modality alignment within the answer subspace, rather than merely adjusting high-level fusion strategies.
50%UnverifiedThe answer subspace refers to a low-dimensional subspace in the activation space of a given layer of the model (usually the position corresponding to the answer token) spanned by the answer directions of multiple input-output samples.
50%