Research is more than a paper title.

Training-free token pruning accelerates MLLMs, but replacing many visual tokens with a few representatives systematically suppresses accumulated visual attention.

Pruning is only useful when the reduced sequence preserves both the content and attention distribution of the original visual evidence.

Dual-view entropy pruning finds diverse, head-salient anchors; bias-aware token recycling records pruned evidence; logit-preserving rectification restores the missing attention mass.

Identifies Attention Logit Collapse, gives it a principled treatment, and validates the approach across single-image, multi-image, video, multiple MLLM architectures, and a vLLM serving environment.

High-resolution GUI observations generate substantial visual-token redundancy, while mobile devices impose strict latency, memory, and package-size constraints.

An agent that understands an interface but cannot respond within a practical device budget is not yet deployable.

A training-free large-model pruning strategy for GUI interfaces, paired with system-level optimization for terminal deployment.

Connects model-side redundancy reduction with end-to-end deployment constraints, without requiring additional model training.

Conventional multimodal ReID systems fuse heterogeneous visual features but underuse text semantics and retain redundant local observations.

Structured text can provide a shared semantic space for objects observed under different spectra and viewpoints.

IDEA combines an inverted multimodal feature extractor with cooperative deformable aggregation and text-enhanced benchmarks.

Introduces semantic guidance from MLLM-generated captions and adaptive local aggregation for robust multimodal retrieval.

Long multimodal sequences contain both modality-specific knowledge and substantial redundancy, making direct fusion brittle and expensive.

Efficiency should come from better information organization, not merely from shrinking the model.

A connected line of work spanning token selection, decoupled mixture-of-experts, synergistic prompting, and Mamba aggregation.

Builds a progression from diverse token discovery to adaptive feature routing and scalable long-sequence interaction.

Identity cues change across RGB, near-infrared, thermal, aerial, and text-enhanced observations.

Reliable retrieval requires both shared semantics and respect for modality-specific evidence.

Token permutation, diverse token selection, hybrid convolution–Transformer representation, Mamba aggregation, mixture-of-experts, and diffusion priors.

A continuous body of work that connects foundational multimodal recognition with emerging vision-language and efficiency questions.

Fast-moving research areas are difficult to navigate when methods, datasets, and resources are scattered.

Good research infrastructure begins with a legible map of the field.

Continuously organized paper lists, datasets, taxonomies, and implementation resources.

Two community-facing knowledge bases connecting multimodal perception and efficient multimodal foundation models.