Curiosity Driven Knowledge Retrieval for Mobile Agents

Table 1 presents the overall comparison on the AndroidWorld benchmark. Compared with existing state-of-the-art methods, DroidRun with AppCards using GPT 5 in version v0.3.9 achieves a success rate of 88.8%, surpassing the previously reported 84.5% of MobileUse and establishing a new publicly available best result. In terms of relative improvement, GPT 5 increases from 84.5% to 88.8%, corresponding to an absolute gain of 4.3 percentage points and a relative increase of approximately 5.1%. For Gemini 2.5 Pro, the performance in version v0.3.3 improves from 63.0% to 69.0%, and in version v0.3.9 improves from 65.5% to 71.6%. Both versions yield gains of around 6 percentage points, with relative improvements in the range of 9–10%. In contrast, Grok 4 fast decreases from 61.2% to 60.3% in version v0.3.9, showing no benefit from AppCards.

Taken together, these results indicate that the effectiveness of AppCards varies across models and versions. When the backbone model has strong reasoning and knowledge integration capabilities, AppCards consistently translate into performance improvements, with GPT 5 achieving breakthrough results. For weaker or stylistically mismatched models, however, AppCards may not be fully utilized and can even introduce negative effects. This suggests that AppCards not only contribute to performance enhancement but also expose critical interactions between external knowledge and model characteristics, offering insights for designing more robust knowledge injection mechanisms.

Table 2 reports detailed performance across different task difficulty levels. In version v0.3.3, Gemini 2.5 Pro combined with AppCards improves from 78.7% to 83.6% on easy tasks, from 55.6% to 61.1% on medium tasks, and from 26.3% to 36.8% on hard tasks. Although improvements appear across all levels, the most substantial gain is observed in hard tasks, with an increase of 10.5 percentage points, highlighting the value of AppCards in guiding models through complex task structures. In version v0.3.9, GPT 5 exhibits even more pronounced improvements. Easy tasks remain nearly unchanged, moving from 90.2% to 91.8%. Medium tasks remain stable at 88.9%. Hard tasks, however, increase dramatically from 57.9% to 78.9%, yielding a 21.0 percentage point improvement. This underscores the critical role of AppCards in bridging knowledge gaps under complex conditions. Gemini 2.5 Pro in the same version shows consistent improvements of 4.9, 8.3, and 5.2 percentage points across easy, medium, and hard tasks, respectively, reflecting balanced benefits. By contrast, Grok 4 fast shows modest gains on easy and medium tasks but experiences a sharp decline on hard tasks, dropping from 52.6% to 31.6%, a decrease of 21.0 percentage points, indicating that its knowledge integration ability is insufficient to leverage AppCards in demanding scenarios.

Overall, these stratified experiments demonstrate that AppCards provide the greatest benefit in complex tasks, substantially enhancing long-horizon reasoning and cross-application operations. At the same time, the experiments reveal clear divergences among models. Strong backbones are able to reliably absorb external knowledge and achieve significant gains, whereas weaker models may encounter instability. This contrast shows that AppCards not only improve average performance but also provide concrete evidence for understanding the mechanisms of interaction between external knowledge and model capabilities.