Enterprise AI Analysis

Consistent Performance Gaps Identified

Evaluations across seven advanced LVLMs reveal consistent non-English performance gaps relative to English, particularly in reasoning-intensive tasks. The benchmark demonstrates a significant capability imbalance across the eight dimensions, with models achieving near-saturation in basic perception tasks (e.g., WI) but diverging sharply in spatial reasoning tasks.

A high correlation between fundamental P&R capabilities and end-to-end competence indicates that the FPR-ACC score effectively reflects both basic and advanced performance.

GUI-XLI: Cross-lingual Intervention Method

To bridge cross-lingual P&R gaps, the paper proposes GUI Cross-Lingual Intervention (GUI-XLI). This method leverages the superior P&R capabilities of English by steering non-English representations toward their English counterparts at critical layers sensitive to linguistic factors during inference. GUI-XLI achieves an average performance gain of 6.5% in non-English settings with negligible inference latency, aligning cross-lingual reasoning patterns at the representational level.

The approach involves constructing a GUI Cross-Lingual Memory to store discrepancy vectors, enabling adaptive retrieval and application during inference as optimization directions.

Visualizing Cross-lingual Alignment

Analysis of intermediate layers shows that they serve as English-centric reasoning hubs, and cross-lingual distributional differences reflect P&R discrepancies. Using t-SNE visualization, the paper demonstrates that without GUI-XLI, representations form distinct language-specific clusters. After applying GUI-XLI, non-English representations become more concentrated and aligned with their English counterparts, qualitatively confirming the bridging of GUI P&R gaps.