Research map / Large language models / Agents, code and math
LLM as Dataset Analyst: Subpopulation Structure Discovery with Large Language Model
Yulin Luo, Ruichuan An, Bocheng Zou, Yiming Tang, Jiaming Liu, Shanghang Zhang
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 |
|---|---|---|---|
| add Criteria Refinement via self-consistency (ablation) | SSD-LLM · Stanford 40 Action (Action) | clustering accuracy (%): 75.6 -> 81.7 | Improved |
| reduce dimension suggestion rounds from 10 to 5 (ablation) | SSD-LLM · Stanford 40 Action (Action) | clustering accuracy (%): 75.6 -> 65.2 | Hurt |
| increase dimension suggestion rounds from 10 to 20 (ablation) | SSD-LLM · Stanford 40 Action (Action) | clustering accuracy (%): 75.6 -> 75.9 | No clear effect |
| use more captions per attribute-summarization round (40 vs 20) (ablation) | SSD-LLM · Stanford 40 Action (Action) | clustering accuracy (%): 75.6 -> 70.3 | Hurt |
| use more rounds with fewer captions per round (40x10) for attributes (ablation) | SSD-LLM · Stanford 40 Action (Action) | clustering accuracy (%): 75.6 -> 80.2 | Improved |
| use SSD-LLM-discovered attributes in generation prompts instead of unknown attributes (ablation) | diffusion augmentation · Waterbirds | worst group accuracy: 71.3 -> 76.2 | Improved |
| sample underrepresented subpopulations with weights instead of randomly (ablation) | diffusion augmentation · Waterbirds | worst group accuracy: 76.2 -> 77.4 | Improved |
| use LLM-generated sentences as diffusion prompts instead of direct attribute prompts (ablation) | diffusion augmentation · Waterbirds | worst group accuracy: 77.4 -> 78.3 | Improved |
| add LLM-suggested extra dimensions and attributes (ablation) | diffusion augmentation · Waterbirds | worst group accuracy: 78.3 -> 79.1 | Improved |
| double the number of generated images (ablation) | diffusion augmentation · Waterbirds | worst group accuracy: 78.3 -> 79.2 (worst group), 90.1 -> 91.1 (average) | No clear effect |
| use SSD-LLM for subpopulation shift instead of best prior method | image classification · Waterbirds, Metashift, Nico++ | worst group accuracy (average): 58.7 -> 62.0 | Improved |
| use SSD-LLM for slice discovery instead of Domino | ImageNet-1K · ImageNet-1K slice discovery | topic error rate (%): 43.44 -> 47.39 | Improved |
Closest papers by meaning
- Learning With Multi-Group Guarantees For Clusterable Subpopulations (ICML 2025 · similarity 0.84)
- Curated LLM: Synergy of LLMs and Data Curation for tabular augmentation in low-data regimes (ICML 2024 · similarity 0.84)
- Diverse Prototypical Ensembles Improve Robustness to Subpopulation Shift (ICML 2025 · similarity 0.84)
- RedPajama: an Open Dataset for Training Large Language Models (NeurIPS 2024 · similarity 0.83)
- Dataset Decomposition: Faster LLM Training with Variable Sequence Length Curriculum (NeurIPS 2024 · similarity 0.83)
- Prismatic Synthesis: Gradient-based Data Diversification Boosts Generalization in LLM Reasoning (NeurIPS 2025 · similarity 0.83)
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.
