From the source
Elorian researchers studied when grounded reasoning — where a vision language model emits text coordinates during chain-of-thought — actually helps, and found it pays off only for references that cannot be described in words.
- The team introduced a test that randomizes each coordinate as it is decoded and resets the KV-cache, giving a cleaner signal than comparing answers before and after randomization.
- Applying it to TreeVGR, iVGR and DeepEyes, they report that these models generally do not significantly rely on their grounding for perception and spatial reasoning tasks such as MultiHopSpatial and SpatialMQA.
- Because many existing VQA benchmarks can be solved with simple textual references, the authors built BrickVision, a benchmark for reasoning with undescribable visual references.
- In training experiments initialized from Qwen3-VL 8B, they found that a reflection template — "Let's look at [coordinate]" — produced significantly more grounding usage and better counting performance than appending coordinates when objects are mentioned.
- The work is an abridged version of a paper, with full results on arXiv.

