LADDER (Language-Driven Slice Discovery and Error Rectification) is a new framework from our lab that uses Mammo-CLIP to detect and correct biases in vision classifiers, including those trained on mammograms.
Instead of relying on manually defined subgroups or attributes, LADDER discovers failure modes and bias slices using natural language and interpretable reasoning from Large Language Models (LLMs).
- 🧠 Automatically identify performance disparities across latent subgroups
- 🩻 Evaluate alignment of radiology reports with model predictions
- ⚙️ Use pseudo-labels and debiasing to correct classifier errors — no extra annotation needed
Example: If your model underperforms on younger patients or dense breast cases, LADDER helps surface those slices using textual probes (e.g., "dense breast with asymmetry") and suggests retraining paths to reduce bias — all without needing protected attribute labels.
This work was presented at ACL 2025.