Saliency-Aware Multi-Route Thinking: Grounding and Reasoning on Vision-Language Agents
Published in NeurIPS, 2026
Mingjia Shi, Yinhan He, Yaochen Zhu, Jundong Li.
Accepted at NeurIPS 2026. Mingjia Shi is the first author.
Saliency-Aware Principle Selection (SAP) is a training-free inference-time method for vision-language agents. It evolves high-level reasoning principles, explores short parallel reasoning routes, and uses visual grounding tools as advisory evidence to reduce object hallucination under a comparable token budget.
The linked arXiv preprint is the earlier version, titled Saliency-Aware Multi-Route Thinking: Revisiting Vision-Language Reasoning.
Recommended citation: M. Shi, Y. He, Y. Zhu, J. Li. Saliency-Aware Multi-Route Thinking: Grounding and Reasoning on Vision-Language Agents. NeurIPS 2026.
Download Paper