NTH

Let Samples Sing in Chorus: Mitigating Spurious Correlations via Cross-view Consensus

AuthorsWeiwei Li, Junzhuo Liu, Tong Chu, Hengfu Yu, Yinjie Lei, Wen Li

AffiliationsUniversity of Electronic Science and Technology of China · Sichuan University

November 20, 2026 2 min read
Watch on YouTube
The one-line take

CHORD combines clues from multiple representation spaces to help models rely less on misleading shortcuts and perform better on underrepresented groups.

Key results

84.49%
Mean WGA on four standard benchmarks

CHORD’s mean worst-group accuracy, compared with 82.01% for NSF.

82.01%
NSF mean WGA on four standard benchmarks

Baseline mean worst-group accuracy reported for comparison with CHORD.

59.93%
Speech-task mean WGA

CHORD’s mean worst-group accuracy on CREMA-D and MEAD, versus 53.99% for ERM.

53.99%
ERM speech-task mean WGA

ERM comparison score on CREMA-D and MEAD.

71.43%
CheXpert WGA

CHORD’s worst-group accuracy on CheXpert, versus 31.21% for ERM.

What the paper found

Deep models can achieve high average accuracy by relying on shortcuts—such as a bird’s background—while failing on underrepresented groups. CHORD addresses this without group labels by letting frozen representation views propose within-class sample partitions, then testing each partition in the target model’s feature space. A constrained diagonal transformation suppresses differences between partition subsets; samples whose target-space predictions return to their labeled class become reversal evidence. CHORD aggregates that evidence into soft sampling weights and retrains only the classifier head, discarding auxiliary views at inference. Across Waterbirds, CelebA, MultiNLI, CivilComments, CREMA-D, MEAD, and CheXpert, it reports a mean worst-group accuracy of 84.49% on four standard benchmarks, versus 82.01% for NSF. On speech tasks, mean worst-group accuracy is 59.93%, compared with 53.99% for ERM; on CheXpert it reaches 71.43%, versus 31.21% for ERM. The method combines pretrained views such as DINO and CLIP with a geometry-enhanced target view, and applies across vision, language, speech, and medical imaging models including ResNet-50, BERT, and wav2vec 2.0. The paper also discloses that OpenAI Codex assisted with manuscript and code preparation and experiment-related work.

Original abstract

Deep models often rely on spurious correlations that fail on underrepresented groups. Existing methods either correct human-specified bias attributes or infer unknown biases from the learned representation space. However, annotated attributes may differ from the shortcuts actually used by a model, while a single representation provides incomplete and potentially noisy evidence. Our key insight is that a bias obscured in one feature space may be clearer in another. In this work, we propose CHORD (Cross-view Harmonic Outlier Reversal for Debiasing), a multi-view framework that allows samples to "sing in chorus" across diverse representation spaces to achieve robust cross-view consensus. CHORD aggregates aggregates within-class partitions in each view, which is a proxy for specifying candidate biases, to debias the target representation and classifier without bias supervision. Experiments on seven benchmarks across vision, language, speech, and medical imaging show consistent improvements in worst-group accuracy. Code and checkpoints are available at https://github.com/davelee-uestc/chord_debiasing_public .

Read the original paper

More in Multimodal AI

Browse all 79 papers →
01Multimodal

ALIVE: Making Streaming Video Models Interactive

Jinzhao Li, Yuan Xu, Jinxuan Wu, Yukun Wang, Yijia Lei, Hanyu Ji, Yinuo Chen, Zining Huang, Chi Zhang, Hang Ye, Yiwu Zhong, Honglei Yan, Panwang Pan, Yizhou Wang, Yin Li, Miao Liu

ALIVE aims to make video-streaming AI an interactive partner that can monitor events, coordinate tasks, and ask for feedback.

Read analysis