ROVER: Routing Object-Centric Visual Evidence for Grounded Multi-Image Reasoning

Researchers introduce ROVER, a lightweight plugin that enhances multimodal large language models' ability to reason across multiple images by intelligently routing visual evidence to specific objects. The approach achieves significant performance improvements on grounded reasoning benchmarks while reducing computational overhead compared to existing methods.

ROVER addresses a fundamental challenge in multimodal AI: how to ground language reasoning in visual evidence across multiple images without computational bloat. Traditional approaches either inject all visual details into reasoning contexts—creating computational scaling problems—or use complex heuristics requiring extensive supervision. The ROVER method elegantly sidesteps these tradeoffs through an object-centric routing mechanism that activates only relevant visual evidence when needed, functioning as a learnable plugin compatible with existing large vision-language models.

The technical innovation centers on injecting step-specific token triplets that coordinate three tasks: aggregating context, distilling image-specific cues through differential attention, and routing historical evidence across objects and images. This architecture maintains holistic scene understanding while focusing computation precisely where reasoning requires visual grounding. The integration into Qwen2.5-VL-7B demonstrates practical viability at commercial model scales.

Performance gains across benchmarks—particularly the 14.6% grounding accuracy improvement on MM-GCoT—suggest ROVER meaningfully advances visual reasoning beyond current approaches. The strong transferability of VideoEspresso-trained models indicates the method generalizes across diverse visual reasoning tasks without task-specific tuning. For AI developers building multimodal systems, ROVER represents a replicable architecture pattern for efficient evidence routing that could influence how subsequent models handle visual grounding.

The research signals maturation in multimodal reasoning, moving beyond brute-force approaches toward computationally efficient, interpretable designs that respect both visual context and reasoning flow.

• →ROVER achieves 4.8% accuracy gains on multimodal reasoning by routing visual evidence selectively rather than injecting all visual details into context.
• →The method reduces computational scaling problems associated with region-of-interest approaches through lightweight object-centric differential attention mechanisms.
• →Strong transferability across benchmarks suggests the routing architecture generalizes beyond its training tasks, improving base model performance by 4.7% on average.
• →Integration with Qwen2.5-VL-7B demonstrates the approach scales to production-grade language models without architectural redesign.
• →Grounding accuracy improvements of 14.6% indicate ROVER effectively localizes visual evidence to specific objects during multi-image reasoning.