benchmarks evals
The Counterfactual Must Be Close in Every Modality
CRANE retrieves real training examples whose text, image or other components jointly resemble the query while belonging to the alternate class.
Summary
CRANE retrieves real training examples whose text, image or other components jointly resemble the query while belonging to the alternate class.
Most counterfactual explanations are built for one data type, while multimodal attribution methods often explain each input independently. CRANE is a post-hoc, model-agnostic retrieval framework that searches real training examples from the target class. Its strongest strategy intersects nearest-neighbor sets from each modality, grounding the explanation in a jointly observed example rather than synthesizing an unconstrained alternative. Across four heterogeneous datasets and three fusion strategies, multimodal consensus retrieval produced the best or joint-best result on three datasets, while single-modality methods remained competitive when one channel dominated the decision. The approach inherits the coverage and biases of the training set and does not show that a nearby example is a causal explanation.
Why it matters
CRANE retrieves real training examples whose text, image or other components jointly resemble the query while belonging to the alternate class.
Limits and context
- The approach inherits the coverage and biases of the training set and does not show that a nearby example is a causal explanation.
Key claims
CRANE retrieves real training examples whose text, image or other components jointly resemble the query while belonging to the alternate class.
Qualification: The approach inherits the coverage and biases of the training set and does not show that a nearby example is a causal explanation.
Evidence: source-2026-10-25-014
Sources
- arXiv preprint 2610.31594, original v1arXiv · primary research
Corrections
No corrections have been recorded for this story.