Result Analysis

The analysis uses the released results to ask a more central question: how far current MLLMs already satisfy the benchmark's dual-cognition requirement of reasoning both about the UAV itself and about the external world, when that requirement is tested under image and video observation settings. The emerging picture is clear: current models exhibit partial competence on both branches, yet neither capability is sufficiently strong in its own right, nor do the two appear to develop in a coordinated and consistent way. They therefore do not yet behave as if self-aware and environment-aware reasoning were one unified capability.

The sections below therefore move from media-specific evidence diagnosis to aggregate dual-cognition comparison. The first half asks how much dual-cognition performance survives once image reasoning must be grounded spatially and video reasoning must be localized temporally. The second half then asks whether overall dual-cognition ability remains stable across media and whether the two cognition branches stay associated and balanced after all released task families are aggregated.

Under image evidence, the central question is whether a model that can reason correctly about the UAV or the target landmark can also retrieve the spatial evidence that makes that reasoning verifiable. The current selected-model subset shows that semantic option accuracy remains consistently higher than spatial Acc@0.5 across nearly the entire comparison pool, indicating that dual-cognition reasoning under images is still easier at the answer level than at the grounding level.

This distinction matters because the image branch asks the model to support both self-aware and environment-aware decisions with explicit spatial evidence. Some systems remain competitive so long as only option selection is measured, yet lose a substantial fraction of their apparent advantage once dual-cognition reasoning must be grounded by a verifiable bounding box in the current view.

The bar chart below is constructed from the same image-model subset used by the benchmark examples, with one bar showing semantic option accuracy and the other showing spatial Acc@0.5 for each model. Read in that way, the figure is not only a ranking comparison: it summarizes how much of each model's apparent dual-cognition competence survives once image reasoning must be backed by explicit spatial grounding.

Under video evidence, the relevant question is whether a model that reaches the correct self-aware or environment-aware interpretation can also recover the temporal support that makes that interpretation inspectable. The current selected-model subset exhibits a related but not identical structure: semantic success and temporal localization remain more closely coupled than image semantics and spatial grounding, yet the interval metric still trails the semantic metric for most models.

This means that recognizing the correct behavior or visibility state is still not equivalent to delimiting the corresponding temporal support with sufficient precision. The resulting ranking changes are analytically useful because they reveal which models preserve temporal evidence quality after semantic recognition has already been achieved, and which models mainly gain from coarse recognition without matching boundary quality.

The bar chart below is constructed from the same video-model subset used by the benchmark examples, with one bar showing weighted semantic performance and the other showing temporal F1@0.5 for each model. The figure should therefore be read as a summary of how much dual-cognition performance remains once video reasoning is required to produce explicit interval evidence rather than only the correct semantic judgment.

The overall image-versus-video scatter shows a clear positive relation, but not a tight diagonal collapse. Across the paired models retained here, the aggregate cross-media relation remains only moderate (Pearson r = 0.59), so stronger image performance often carries into video reasoning, but far from uniformly.

The imbalance around the diagonal is equally informative. Among the displayed paired models, 18 lie above the diagonal and only 5 lie below it, so the present model pool more often improves or at least preserves its aggregate standing under video evidence than under image evidence. Cross-media transfer is therefore real, but it remains directional rather than symmetric.

The correlation summary confirms the same pattern numerically. Across the 23 paired models retained here, the overall image-video relation remains moderately positive (Pearson r = 0.59), which is strong enough to rule out a random pairing but still far from a fully stable cross-media transfer regime.

The aggregate means also sharpen the interpretation: image performance averages 19.48%, whereas video performance rises to 23.09%, leaving a positive gap of +3.62 points. Read together with the scatter, this indicates that the current model pool does not simply preserve one fixed level of dual-cognition quality across media; instead, many systems shift to a visibly different operating point once the evidence becomes temporally extended.

The overall self-versus-environment scatter again shows a positive but incomplete alignment (Pearson r = 0.39). Current models therefore do not split into two wholly unrelated branches, but neither do they collapse into one balanced capability.

The more striking feature is the direction of that spread. Among the displayed paired models, 24 lie above the diagonal, whereas only 0 lie below it. The dominant pattern is therefore not random variance but a systematic tilt toward stronger environment-aware scores, which suggests that current MLLMs still find self-aware reasoning harder to consolidate at the same level once both media are taken into account.

The correlation summary makes the same imbalance precise. Across the 24 paired models, self-aware and environment-aware scores remain positively related (Pearson r = 0.39), so the two branches do rise together to some extent rather than behaving as unrelated skills.

Yet the mean values show that this coupling is not balanced. Self-aware performance averages only 25.5%, whereas environment-aware performance reaches 43.92%, yielding a large positive gap of +18.42 points. The result is therefore not a unified dual-cognition competence, but a partially coupled system in which environment-aware reasoning is generally more mature than self-aware reasoning.