Research map / Large language models / Instruction and fine-tuning
Bridging Explainability and Embeddings: BEE Aware of Spuriousness
Cristian Daniel Paduraru, Antonio Barbalau, Radu Filipescu, Andrei Liviu Nicolicioiu, Elena Burceanu
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 changed | Where | Measured | Effect |
|---|---|---|---|
| regularize class weights to be equidistant from BEE-found SCs | linear probe, Waterbirds without counterexamples · Waterbirds (no counterexamples) | worst-group accuracy: 43.2 -> 57.9 | Improved |
| regularize class weights to be equidistant from BEE-found SCs | linear probe, CivilComments without counterexamples · CivilComments (no counterexamples) | worst-group accuracy: 18.6 -> 31.3 | Improved |
| regularize class weights to be equidistant from BEE-found SCs | linear probe, CelebA without counterexamples · CelebA (no counterexamples) | worst-group accuracy: 9.6 -> 10.4 | Improved |
| use BEE-identified SCs instead of random class-neutral concepts for regularization (ablation) | SC regularization, Waterbirds · Waterbirds (no counterexamples) | worst-group accuracy: 46.6 -> 57.9 | Improved |
| use BEE-identified SCs instead of random class-neutral concepts for regularization (ablation) | SC regularization, CivilComments · CivilComments (no counterexamples) | worst-group accuracy: 19.1 -> 31.3 | Improved |
| use BEE SCs instead of Lg's SCs for regularization (ablation) | SC regularization, CelebA · CelebA (no counterexamples) | worst-group accuracy: 8.3 -> 10.4 | Improved |
| train with GroupDRO instead of ERM when no minority groups exist | linear probe, Waterbirds · Waterbirds (no counterexamples) | worst-group accuracy: 43.2 -> 38.9 | Hurt |
| add BEE-found spurious concept firemen to peafowl image | CLIP ViT-L/14 fine-tuned on ImageNet · FLUX-generated peafowl images | correct class rate (%): 100.0 -> 5.3 | Hurt |
| add BEE-found spurious concept reading a newspaper to Mexican hairless dog image | CLIP ViT-L/14 fine-tuned on ImageNet · FLUX-generated Mexican hairless dog images | correct class rate (%): 47.5 -> 0.9 | Hurt |
| add BEE-found spurious concept shrimp to Bernese Mountain Dog image | CLIP ViT-L/14 fine-tuned on ImageNet · FLUX-generated Bernese Mountain Dog images | correct class rate (%): 99.8 -> 10.6 | Hurt |
| add firemen to peafowl prompt, evaluated on alexnet | alexnet trained on ImageNet-1k · Generated peafowl + firemen (transfer) | accuracy (%): 100.0 -> 4.6 | Hurt |
| add firemen to peafowl prompt, evaluated on swin_b | swin_b trained on ImageNet-1k · Generated peafowl + firemen (transfer) | accuracy (%): 100.0 -> 81.5 | Hurt |
Closest papers by meaning
- BEE: Metric-Adapted Explanations via Baseline Exploration-Exploitation (AAAI 2025 · similarity 0.86)
- Let Samples Speak: Mitigating Spurious Correlation by Exploiting the Clusterness of Samples (CVPR 2025 · similarity 0.84)
- Diagnosing Transformers: Illuminating Feature Spaces for Clinical Decision-Making (ICLR 2024 · similarity 0.84)
- SANEval: Open-Vocabulary Compositional Benchmarks with Failure-mode Diagnosis (NeurIPS 2026 · similarity 0.84)
- Complexity Matters: Feature Learning in the Presence of Spurious Correlations (ICML 2024 · similarity 0.84)
- Local vs. Global Interpretability: A Computational Complexity Perspective (ICML 2024 · similarity 0.84)
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.
