Ideal Contribution Estimation: Taylor Expansion and Gradient Cache
4 Dec 2025
The ideal contribution of generated data is approximated using a first-order Taylor expansion based on the gradient dot product.
Formalizing Generative Active Learning for Instance Segmentation
3 Dec 2025
Proposes BSGAL, a Generative Active Learning algorithm that uses gradient cache to filter unlimited synthetic data.
Active Learning and Data Influence: Core Concepts and Evolution
3 Dec 2025
Highlights the novelty of applying these methods to generated data in complex instance segmentation.
From GANs to Diffusion: GDA for Perception Tasks
19 Nov 2025
Highlights its use in perception (segmentation, detection) but notes the limited exploration of filtering and effective utilization.
Enhancing Long-Tailed Segmentation with Gradient Cache and BSGAL
19 Nov 2025
Proposes BSGAL, a Generative Active Learning algorithm that uses gradient cache to filter unlimited synthetic data for long-tailed instance segmentation.
Cross-Model Validation: MIVPG's Efficacy on Encoder-Decoder vs. Decoder-Only LLMs
19 Nov 2025
MIVPG's CSA module remains effective when switching from FLAN-T5-XL to the OPT-2.7b LLM architecture.
Theoretical Proof: CSA Module Maintains MIL Properties
19 Nov 2025
Provides the theoretical proof for Proposition 2, establishing that the Correlated Self-Attention (CSA) module in MIVPG maintains permutation equivalence.
Visual Prompt Generation: Cross-Attention in Q-Former
19 Nov 2025
Details the Q-Former architecture: a 12-layer BERT-based model using 32 learnable query embeddings.
Future MLLMs: Contribution of MIL-Based Techniques and Enriched Visual Signals
18 Nov 2025
This paper concludes that MIVPG is a general, powerful component for fusing enriched visual representations in MLLMs.