A clear lead on the benchmark
An AI system built to identify social features in urban scenes recorded the strongest overall scores in a test of boundary-aware image segmentation. On the SocioSeg test set, EVEREST had the best reported average rank, with a cIoU of 50.4 and an F1 score of 61.4.
Those two measures capture different parts of a mask. cIoU measures how much the predicted region overlaps the target, while F1 summarizes the balance between correctly included and missed pixels. SocioSeg reports both for each task group and as overall averages.
Against SocioReasoner, EVEREST's reported overall cIoU was 50.4, compared with 47.9, while F1 was 61.4 compared with 59.7.
What the benchmark asks the model to do
SocioSeg samples pair a satellite image with a spatially aligned digital map and a pixel-level mask marking a target social entity. The dataset reports more than 13,000 samples, over 5,000 entity names, 90 classes and 10 functions.
A search-and-check design
The paper frames the problem around two weaknesses: inaccurate target boundaries and too little active exploration. EVEREST responds with an egocentric self-correction process, in which the model is guided by pseudocode-regularized execution while it works through the image.
The workflow has two coupled stages. Enumerate looks across the scene for candidate instances. Verify then explores each candidate's boundary using rendered feedback, deciding whether to keep it, adjust it or drop it.
Because parsing, rendering and SAM2 execution are non-differentiable, the vision-language model's policy is trained with GRPO-style reinforcement learning, a reward-based approach.
In the reported implementation, the vision-language backbone was Qwen2.5-VL-3B and the promptable segmenter was a frozen SAM2 model. Training used a rollout batch size of 128, groups of eight, and AdamW with a learning rate of 1 × 10−6.
Where the scores moved
The pattern was not identical across task levels. For Socio-class, EVEREST reported cIoU 50.5 and F1 61.5, 2.9 and 1.4 points above SocioReasoner respectively.
On Socio-function, the reported cIoU was 40.6 for SocioReasoner and 44.4 for EVEREST, while F1 was 50.3 and 54.9 respectively.
Socio-name was less one-sided: SocioReasoner had a slightly higher F1, while EVEREST had the best cIoU.
In a qualitative case study, the paper describes SocioReasoner as producing a single-instance segmentation with ambiguous boundaries, while EVEREST produced more accurate and coherent segmentation through its staged process.
The internal checks were mixed
Changing the number of refinement points did not produce a simple upward trend. The full model reported cIoU 50.4 and F1 61.4; with one point, the scores fell by 0.9 and 0.8. Three points produced cIoU 50.6 and F1 61.2, while four points produced 47.7 and 59.6.
Removing either of the two stages was associated with lower reported scores. Removing Enumeration was associated with declines of 6.9 cIoU points and 8.0 F1 points; removing Verification was associated with declines of 4.1 and 2.6 points.
A result with a narrow boundary
The paper's testing protocol adds some context but not a measure of statistical certainty. Hyperparameters were selected on the validation set with a single seed, then each method was trained across five different random seeds. The preprint does not report confidence intervals, significance tests or per-seed variability summaries.
That leaves the main conclusion appropriately narrow: EVEREST ranked highest on the reported SocioSeg test metrics, but the evidence does not establish that the same result would hold on other datasets, regions, sensors or model backbones. Nor did the study test whether better masks improve urban resource allocation, navigation or planning.
The document is an arXiv preprint dated 15 September 2026.
Paper data and sources
Original title: EVEREST:Endogenous Vision-Language Reinforcement Reasoning Exploration for Urban Socio-Semantic Segmentation
Authors: Qixiu Li, Zhongzhi He, Xiang Zhu et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-25
DOI: Not available
Original paper · Full text