AI Paper Atlas

Research map / Probabilistic and scientific modeling / Causal discovery

Causal Fine-Tuning under Latent Confounded Shift

Jialin Yu, Yuxiang Zhou, Haoxuan Li, Junchi Yu, Mengyue Yang, Yulan He, Nevin L. Zhang, Philip Torr, Ricardo Silva

ICML 2026 · Probabilistic and scientific modeling › Causal discovery · approach: Causal representation identifiability · 0 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
add Causal Fine-Tuning to standard fine-tuningBERT SFT · Amazon stop-word injection, OOD spurious 10%F1: 49.33 -> 56.40Improved
add Causal Fine-Tuning to standard fine-tuningBERT SFT · data-source injection, OOD spurious 10%F1: 37.78 -> 49.22Improved
add Causal Fine-Tuning to standard fine-tuningBERT SFT · in-distribution test spurious 90%F1: 92.39 -> 92.37 (Amazon); 91.39 -> 90.59 (data source)No clear effect
CFT versus SWA under mild shiftBERT · Amazon stop-word injection, OOD spurious 10%F1: SWA 59.75 vs CFT 56.40Hurt
CFT versus SWA under 8x scaled spurious noiseBERT · Amazon stop-word injection 8x noise, OOD spurious 10%F1: 33.05 -> 33.11Improved
condition on Phi without causal adjustment (CFT-N) (ablation)CFT · data-source injection, OOD spurious 10%F1: 49.22 -> 15.05Hurt
add causal adjustment over Phi (CFT versus CFT-N) (ablation)CFT · Amazon stop-word injection, OOD spurious 10%F1: 48.00 -> 56.40Improved
use only causal representation C to predict (CFT-C) (ablation)CFT · data-source injection, OOD spurious 10%F1: 49.22 -> 42.25Hurt
use only low-level features Phi to predict (CFT-Phi) (ablation)CFT · Amazon stop-word injection, OOD spurious 10%F1: 56.40 -> 12.40Hurt
replace paired views (R0,R1) with identical views (R1,R1) (ablation)CFT · data-source injection, OOD spurious 10%F1: 49.22 -> 37.24Hurt
interpolate pretrained and fine-tuned weights (WISE) over SFTBERT SFT · data-source injection, OOD spurious 10%F1: 37.78 -> 31.83Hurt
add more training data (ablation)CFT vs SFT · data-source injectionF1No clear effect

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy