Xiangxian Li 李象贤

Uses a benchmark plus iterative reasoning so the system can inspect child-oriented videos and explain why a clip may be risky.

Learns when to send a prompt to different expert branches, so vision-language tuning stays balanced when long-tailed examples are rare.

Shows that a shorter dialogue can still capture enough behavioral signal for personality assessment, without making the conversation drag on.

Tracks what users are likely trying to select when the target keeps moving, which is useful for dynamic interaction scenes.

Rebuilds missing table information by borrowing signals across modalities and reducing the bias that incomplete columns usually introduce.

Turns EEG patterns into a simple estimate of when people are entering or leaving flow during interactive tasks.

Learns causal links between aligned visual and semantic features so the classifier relies less on spurious correlations.

Balances learning across common and rare legal motives so long-tailed text classification is less biased.

Looks for shared flow moments across participants from EEG in ubiquitous-computing settings.

Uses privileged cross-modal hints during training to learn stronger image features for long-tailed classification.

Uses extra cross-modal information during training to guide image features toward harder classes.

Predicts which interface elements each person is likely to prefer, enabling more personalized UI recommendations.

Compares adversarial training choices for long-tailed recognition and identifies the ones that stay stable in practice.