Get Started
Home
Topics
Search
Library
Research questionHow can we select visual instruction-tuning examples under a fixed budget while preserving alignment and dataset coverage?Visual instruction-tuning corpora are growing quickly, but training on only a budgeted subset can discard examples that support alignment and instruction following. Selection must account for both the coherence of individual visual instruction-response samples and coverage of the broader data distribution.
AI
Computer Vision
LLM Pretraining & Post-training
Machine Learning
Multimodal Models
Latest papersRecent research connected to this question, newest first.DICS: Exploring Data Intrinsic Consistency for Visual Instruction SelectionApplies to visual instruction tuning for vision-language models and evaluates selection across dataset scales, data budgets, and model architectures. The reported evidence includes experiments using substantially smaller subsets, including roughly 25% of certain training corpora; it does not establish applicability beyond the tested datasets and models.research paper · Sep 2, 2026
Related questions
How can instruction-tuned LLMs learn corpus-specific knowledge without exhaustive synthetic QA or instruction fine-tuning?How do visual-space alignment and task balance determine whether joint VLM training helps or conflicts?How should LLM pre-training allocate a fixed token budget between repetition and auxiliary views when prior knowledge is incomplete?How can we robustly compress LLM KV caches across open-domain inputs without input-specific budget thresholds?