# Elorian — Where and Why is Grounded Reasoning Useful?

- Company: Elorian (elorian.ai)
- Announced: 2026-11-06
- Category: research-paper
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://elorian.ai/blog/why-grounding
- Record: https://forck.live/items/32903-where-and-why-is-grounded-reasoning-useful
- Subject: Glimpse visual reasoning models
- Models affected: TreeVGR, iVGR, DeepEyes, Qwen3-VL 8B, Gemini 3.5 Flash

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.

## Evidence

Verbatim from https://elorian.ai/blog/why-grounding:

> David Yan 1,2 , Le Xue 1 , Forrest Huang 1 , Qiuyi Zhang 1 , Dustin Tran 1 , Andrew M. Dai 1 , Yinfei Yang 1
> 
> 1 Elorian · 2 Princeton University
> 
> This is an abridged version of the paper.
> 
> Some sections, experiments and details are left out here. Read the full paper on arXiv for the complete results.
> 
> Summary
> 
> We study the effectiveness of grounded reasoning strategies, where a model outputs text coordinates during chain-of-thought (CoT).
> 
> We introduce a technique to measure grounding usage in existing reasoning models and find that many open models do not rely on their generated coordinates when reasoning.
> 
> We find that grounded reasoning provides advantages on tasks that require using references undescribable with text, but performs no better than plain textual reasoning when objects are describable in language.
> 
> Because many existing VQA benchmarks are solvable with simple textual references, we introduce BrickVision, a challenging benchmark designed to measure a model's capability to reason with undescribable visual references.
> 
> Grounded reasoning models
> 
> While solving visual tasks, humans shift their gaze between objects pertaining to the problem.

---

Record: https://forck.live/items/32903-where-and-why-is-grounded-reasoning-useful
Catalogue: https://forck.live/llms.txt
Current issue: https://forck.live/feed.md
