Research map / Label-efficient and robust learning / Recognition and long-tail
AUXILIARY TASK UPDATE DECOMPOSITION: THE GOOD, THE BAD AND THE NEUTRAL
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| guide cross-task masked-language-modeling updates using supervised-task gradients | ATTITTUD with RoBERTa · Imdb + Amazon MLM | F1: 95.1 $\pm$ 0.10 -> $\mathbf{95.4 \pm 0.03}$ | Helped |
| guide cross-task masked-language-modeling updates using supervised-task gradients | ATTITTUD with RoBERTa · Amazon + Imdb MLM | F1: 67.8 $\pm$ 0.46 -> $\mathbf{68.5 \pm 1.01}$ | Helped |
| guide same-task masked-language-modeling updates using supervised-task gradients | ATTITTUD with RoBERTa · Imdb | F1: 96.1 $\pm$ 0.11 -> 96.1 $\pm$ 0.09 | No clear effect |
| guide same-task masked-language-modeling updates using supervised-task gradients | ATTITTUD with RoBERTa · Amazon | F1: 70.3 $\pm$ 0.87 -> 70.1 $\pm$ 1.13 | Mixed |
| decompose and reweight auxiliary-task gradients for primary-task-aware training | ATTITTUD · MultiCifar100 | Average accuracy: 75.6 -> $\mathbf{76.1}$ | Helped |
| decompose and reweight auxiliary-task gradients for primary-task-aware training | ATTITTUD · Cifar10-Cat-vs-Dogs | Average accuracy: 65.3 $\pm$ 1.35 -> $\mathbf{67.1 \pm 1.31}$ | Helped |
| apply ATTITTUD using ImageNet auxiliary data after ImageNet pretraining | Pretrained ResNet18 · ChexPert-5k | Average AUC across 5 pathologies: 81.4 $\pm$ 1.34 -> $\mathbf{83.3 \pm 0.71}$ | Helped |
| replace the canonical per-parameter basis with top-k randomized-SVD directions (ablation) | ATTITTUD · Cifar10 Cat-vs-Dog | Average accuracy: 51.42 $\pm$ 2.09 -> $\mathbf{62.2 \pm 4.00}$ | Helped |
| replace randomly chosen orthogonal basis vectors with top-k randomized-SVD directions (ablation) | ATTITTUD · Cifar10 Cat-vs-Dog | Average accuracy: 58.72 $\pm$ 2.68 -> $\mathbf{62.2 \pm 4.00}$ | Helped |
| replace the average primary-task gradient basis with top-k randomized-SVD directions (ablation) | ATTITTUD · Cifar10 Cat-vs-Dog | Average accuracy: 59.13 $\pm$ 2.08 -> $\mathbf{62.2 \pm 4.00}$ | Helped |
| use a larger subspace dimension (ablation) | ATTITTUD · low-resource Cifar10 Cat-vs-Dog | Average accuracy | Hurt |
| use a larger subspace dimension (ablation) | ATTITTUD · high-resource MultiCifar100 | Average accuracy | Helped |
Closest papers by meaning
- Aux-NAS: Exploiting Auxiliary Labels with Negligibly Extra Inference Cost (ICLR 2024 · similarity 0.87)
- Modeling Multi-Task Model Merging as Adaptive Projective Gradient Descent (ICML 2025 · similarity 0.85)
- When is Task Vector Provably Effective for Model Editing? A Generalization Analysis of Nonlinear Transformers (ICLR 2025 · similarity 0.84)
- Distribution-Aligned Decoding for Efficient LLM Task Adaptation (NeurIPS 2025 · similarity 0.84)
- Task Groupings Regularization: Data-Free Meta-Learning with Heterogeneous Pre-trained Models (ICML 2024 · similarity 0.84)
- AdaMerging: Adaptive Model Merging for Multi-Task Learning (ICLR 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.