TICDA: Tabular In-Context Data Attribution
TICDA scores which in-context table examples are helping or hurting a tabular foundation model’s prediction.
The paper targets TFMs that make predictions from labeled demonstrations without updating weights. Its method trains linear surrogates on the model’s latent embeddings, giving per-demonstration influence in one forward pass. The authors test it on label-error detection, context curation, cross-model attribution transfer, and active learning, where they say it offers the best overall compromise against competing methods. ArXiv · AI/CL/LG's note
The paper targets TFMs that make predictions from labeled demonstrations without updating weights. Its method trains linear surrogates on the model’s latent embeddings, giving per-demonstration influence in one forward pass. The authors test it on label-error detection, context curation, cross-model attribution transfer, and active learning, where they say it offers the best overall compromise against competing methods. ArXiv · AI/CL/LG's note
score 5