PhaseWin Search Framework Enable Efficient Object-Level Interpretation

2025-11-14

AI Summary

Existing submodular subset selection methods achieve high fidelity in object-level interpretability of foundation models but suffer from prohibitive computational overhead, limiting practical deployment. To address this, we propose PhaseWin—a highly efficient phase-wise window search algorithm that approximates greedy selection performance with near-linear time complexity. PhaseWin employs a coarse-to-fine multi-stage search, adaptive pruning, window-based fine-grained selection, and dynamic supervision. Theoretical analysis rests on the monotone submodular function assumption, ensuring strong provable guarantees and scalability. Experiments demonstrate that PhaseWin attains over 95% greedy attribution fidelity on Grounding DINO and Florence-2 using only 20% of the computational budget required by standard greedy selection. Moreover, it consistently outperforms state-of-the-art baselines across multiple attribution tasks, establishing new efficiency–fidelity trade-off benchmarks in foundation model interpretability.

Technology Category

• Search and Optimization: Learning to Search
• Computer Vision: Interpretability, Explainability, and Transparency
• Machine Learning: Other Foundations of Machine Learning

Application Category

• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs
• User Modeling, Personalization and Recommendation: Fairness-aware retrieval and ranking
• Search and Retrieval-Augmented AI: Efficiency and scalability of Web search engines

Abstract

Attribution is essential for interpreting object-level foundation models. Recent methods based on submodular subset selection have achieved high faithfulness, but their efficiency limitations hinder practical deployment in real-world scenarios. To address this, we propose PhaseWin, a novel phase-window search algorithm that enables faithful region attribution with near-linear complexity. PhaseWin replaces traditional quadratic-cost greedy selection with a phased coarse-to-fine search, combining adaptive pruning, windowed fine-grained selection, and dynamic supervision mechanisms to closely approximate greedy behavior while dramatically reducing model evaluations. Theoretically, PhaseWin retains near-greedy approximation guarantees under mild monotone submodular assumptions. Empirically, PhaseWin achieves over 95% of greedy attribution faithfulness using only 20% of the computational budget, and consistently outperforms other attribution baselines across object detection and visual grounding tasks with Grounding DINO and Florence-2. PhaseWin establishes a new state of the art in scalable, high-faithfulness attribution for object-level multimodal models.

Problem

• Enables efficient object-level interpretation for foundation models
• Reduces quadratic complexity of attribution methods to near-linear
• Maintains high faithfulness while dramatically cutting computational costs

Innovation

• PhaseWin algorithm enables near-linear complexity attribution
• Combines adaptive pruning with windowed fine-grained selection
• Achieves 95% greedy faithfulness with 20% computational cost