Coordinate-free visual reasoning

Thinking with Images without Coordinates

Meng Cao^1*, Haoze Zhao^1*, Yi Wang², Ian Reid¹, Xiaodan Liang^1,3

¹Mohamed bin Zayed University of Artificial Intelligence, ²Shanghai AI Lab, ³Sun Yat-sen University

Comparison between coordinate-based reasoning and TiCO coordinate-free evidence discovery.

TiCO replaces explicit coordinate prediction with semantic evidence discovery in visual token space. A probe token estimates task-relevant similarity over visual patches, connected components become evidence rollouts, and the model answers from semantically sampled regions instead of fragile numerical boxes.

HR-Bench 4K

78.0%

overall accuracy

HR-Bench 8K

76.3%

best overall

Generated tokens

6.2

vs. 527.7 for Mini-o3

Inference time

1.9s

on HR-Bench 4K

#How TiCO Works

Coordinate prediction is replaced by a semantic probe, token-level similarity maps, connected components, and distribution-guided evidence rollouts.

01Probe

Semantic anchoring

The model emits a dedicated <GROUND> token whose hidden state acts as a semantic anchor for task-relevant visual patches.

02Evidence

Region rollouts

Patch similarities are thresholded into connected components and sampled by aggregated attention mass to preserve confident and diverse evidence.

03Answer

Grounded reasoning

Sampled evidence regions are fed back to the policy model, enabling interleaved visual reasoning without explicit coordinate regression.

#Accuracy and Efficiency

TiCO improves high-resolution visual reasoning while sharply reducing the token and time overhead of coordinate-based thinking-with-images methods.

Coordinate-based methods versus coordinate-free TiCO and their performance efficiency comparison.

High-resolution reasoning

On HR-Bench 4K, TiCO reaches 78.0% overall accuracy, outperforming Mini-o3, DeepEyes, and TreeVGR. On HR-Bench 8K, it reaches 76.3%.

General visual grounding

TiCO obtains 66.9% on MME-RealWorld and 89.1% overall on POPE, improving grounded reasoning beyond the Qwen2.5-VL baseline.

Efficient inference

On HR-Bench 4K, TiCO uses only 6.2 generated tokens and 1.9s per query, compared with 527.7 tokens and 20.2s for Mini-o3.

#More Results

Beyond the main benchmark numbers, we highlight the experiments that explain why coordinate-free grounding works: how supervision is built, how robust the evidence extractor is, and where the model succeeds or fails qualitatively.

Robustness study

We perturb bounding-box supervision to test whether evidence extraction collapses under imperfect localization. TiCO remains stable because boundary noise affects only a subset of patch labels, while coordinate regression methods directly inherit numerical target noise.

Robustness and threshold analysis for TiCO.

Optimization dynamics

The training curves show stable SFT convergence for grounding and language modeling losses, followed by RL improvements in both accuracy reward and format reward. The attention visualization illustrates how the semantic probe becomes more focused after reinforcement learning.

Attention visualization after SFT and RL.

Qualitative comparison

In high-resolution scenes, coordinate-based crop-and-zoom pipelines can lock onto the wrong region or lose context after a small localization error. TiCO samples semantic evidence directly from visual tokens, keeping the selected regions aligned with the question.

#Citation

BibTeX will be updated when the public preprint is available.

@article{tico2026thinking,
  title   = {Thinking with Images without Coordinates},
  author  = {Cao, Meng and Zhao, Haoze and Wang, Yi and Reid, Ian and Liang, Xiaodan},
  journal = {arXiv preprint},
  year    = {2026}
}