Three-slot breakdown · RS-VLMs

RS-VLM Comparison Matrix

Every model filed into three slots — Slot A: what encodes pixels · Slot B: what bridges to the LLM · Slot C: what grounds the output — plus training data and tasks. Amber = the model's distinguishing move. Click a header to sort.
All Pixel / ref-seg Reasoning / RL Grounding Temporal / change Multi-sensor Agentic / UHR
Model Slot AVision encoder Slot BBridge Slot CGrounding LLM Training data Tasks Distinguishing move
How to use this as an exercise (Phase 5c): before reading each paper, predict its three slots from the abstract alone; correct the row after dissecting. Notice the historical drift: 2024 innovated in Slot C (grounding mechanisms), late-2024 in data scale, 2025–26 in training method (RL/reasoning) and perception policy (agentic zoom) — architecture is largely frozen at the LLaVA recipe. That stability is itself a gap (see G11/G12).