Research map / Vision-language and multimodal / Multimodal LLMs
GENIXER: Empowering Multimodal Large Language Models as a Powerful Data Generator
Henry Hengyuan Zhao, Pan Zhou, Mike Zheng Shou
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 915K Genixer synthetic VQA data to LLaVA-1.5 pretraining | LLaVA-1.5 7B · VizWiz | accuracy: 50.0 -> 53.8 | Improved |
| add Genixer-915K synthetic VQA data | LLaVA-1.5 7B · MME | score: 1465.0 -> 1502.7 | Improved |
| add Genixer-915K synthetic VQA data | LLaVA-1.5 7B · SEED-Bench image | accuracy: 66.2 -> 66.6 | No clear effect |
| add Genixer-915K synthetic VQA data | LLaVA-1.5 7B · LLaVA-Bench (in the wild), MM-Vet | GPT-4 judged score: 65.4 -> 64.0; 31.1 -> 30.1 | Hurt |
| add Genixer-350K synthetic REC data to Shikra training | Shikra · RefCOCO/+/g average | accuracy: 82.92 -> 83.49 | Improved |
| add Genixer-350K synthetic REC data to Shikra training | Shikra · RefCOCOg val | accuracy: 82.27 -> 81.99 | Hurt |
| use the Genixer-L generator model directly (zero-shot) as a VQA model | Genixer-L · GQA, ScienceQA, SEED | accuracy: GQA 62.0 -> 61.3; SQA 66.8 -> 65.6; VQAv2 78.5 -> 79.7 | Mixed |
| retrain Genixer-L with the 665K LLaVA finetuning data mixed in (MixT) (ablation) | Genixer-L · six benchmarks | accuracy: VizWiz 50.0 -> 54.1; GQA 62.0 -> 63.1 | Improved |
| scale synthetic data from 300K to 915K (ablation) | LLaVA-1.5 pretraining · ScienceQA-IMG | accuracy: 68.5 -> 69.7 | Improved |
| filter samples with Fuyu-8B probability threshold 0.7 instead of none (ablation) | Genixer-915K · ScienceQA-IMG | accuracy: 69.6 -> 69.7 (VizWiz 53.5 -> 53.8) with 1.4M -> 0.9M samples | Improved |
Closest papers by meaning
- Adapt-$\infty$: Scalable Continual Multimodal Instruction Tuning via Dynamic Data Selection (ICLR 2025 · similarity 0.90)
- VIGC: Visual Instruction Generation and Correction (AAAI 2024 · similarity 0.90)
- CoIN: A Benchmark of Continual Instruction Tuning for Multimodel Large Language Models (NeurIPS 2024 · similarity 0.90)
- Scaling Instruction-tuned LLMs to Million-token Contexts via Hierarchical Synthetic Data Generation (ICLR 2025 · similarity 0.89)
- MAmmoTH2: Scaling Instructions from the Web (NeurIPS 2024 · similarity 0.89)
- MIA-Bench: Towards Better Instruction Following Evaluation of Multimodal LLMs (ICLR 2025 · similarity 0.89)
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.
