PhaseWin: An Efficient Search Algorithm for Faithful Visual Attribution

Researchers have proposed PhaseWin, an efficient search algorithm for faithful visual attribution in computer vision models, and analyzed State Space Models (SSMs) for code understanding, finding they capture syntactic and semantic structure effectively.

Researchers have introduced PhaseWin, an algorithm that reorganizes greedy region selection into a phased window-search procedure, achieving controllable linear evaluation complexity with near-greedy faithfulness guarantees^[1]. This development addresses the exponential cost of exhaustive search over region subsets and the quadratic number of model evaluations required by greedy search. In a separate study, researchers analyzed SSMs, which have emerged as an efficient alternative to the Transformer architecture, and found that they can match or surpass Transformers on code understanding tasks^[2]. The analysis revealed that SSMs capture syntactic and semantic structure more effectively than Transformers during pretraining, but may forget certain relations during fine-tuning on some tasks. A frequency-domain framework called SSM-Interpret was used to expose a spectral shift toward short-range dependencies during fine-tuning. The researchers also proposed architectural modifications that significantly improve the performance of SSM-based code models.