← All concepts

visual grounding

Grounding means connecting language to specific parts of a visual scene, like drawing a box around the object a sentence names. Models trained for this are often called "grounding VLM" (vision-language models), meaning they process both text and images. The papers explore how these models learn this skill, whether it is reliable, and what it costs. Key questions include whether they truly use the image or just guess from text, how fast they can run, and whether they can be tricked into focusing on the wrong thing.

Together, the findings show that grounding is fragile and sometimes surprising. Models can learn to use images only if the training reward forces them to, and they may rely on small concrete cues to understand abstract ideas. They are just as vulnerable as older object detectors to physical tricks that make them point at the wrong person. Safety can also break when the scene context nudges a model toward harmful answers, even without explicit danger in the image. Speed and accuracy trade off, and failures can look like confident correct answers, which is risky in real-world use.

Papers
6
Released code
0
First seen
Oct 2026
Latest
Oct 2026

6 papers in the last two months, against 0 in the two before.

Start here

Three papers to read first, in the order they came out
  1. 1Changing reward requirements teaches vision-language models to actually use images

    Multimodal reinforcement learning with verifiable rewards fails to learn visual skills unless tasks require visual evidence.

    Oct 2026
  2. 2Parallel diffusion decoding rivals autoregressive visual grounding at 4B scale

    GroundAnything uses blockwise diffusion to deliver faster parallel visual grounding while retaining strong localization accuracy across diverse tasks.

    Oct 2026
  3. 3AI models struggle to distinguish objects surrounded by similar ones

    MLLMs fail at visual reasoning when tasks demand identifying one object among many similar ones.

    Oct 2026

The papers

Most central to this idea first, not most recent.

Concepts are extracted from each paper and reused across the corpus, so this page grows on its own as the desk reads.