01 / AI · Analysis
GeoJeV trades 1.12 extra seconds for 5.32 points of accuracy
GeoJeV steers image search with probability distributions instead of generated text, but the abstract does not name the method behind its 7.8-fold speed comparison.

GeoJeV, a method that searches ultra-high-resolution remote sensing images for evidence regions, raised the average accuracy of the Qwen3-VL-8B model by 5.32 percentage points for 1.12 extra seconds per question, according to figures its authors published on arXiv on 9 Nov 2026. The average covers three benchmarks. Remote sensing images are pictures of the Earth's surface taken from a distance.
Ultra-high-resolution imagery holds rich detail but is hard to interpret accurately and efficiently, the researchers wrote. The paper reports accuracy and processing time together.
The central question is what makes the decision at each step of the search. The paper moves that decision from text generation to a probability distribution, and its speed claim rests on a comparison method the abstract does not name.
GeoJeV accuracy by benchmark (author-reported)
| LRS-GRO | 60.59 | |
|---|---|---|
| MME-RealWorld-RS | 52.79 | |
| LRS-VQA | 34.32 |
Probability distributions replace generated text in the search
Existing evidence-search methods rely heavily on generative language models that handle images and text together for intermediate search control, the researchers wrote. Their first objection is that reinforcement learning from human feedback (RLHF) tunes these models mainly towards answers people prefer, not towards judging whether a region is relevant. Their second is the computing cost of repeated autoregressive decoding, which generates output one token at a time.
According to the paper, GeoJeV recasts evidence search as a decision problem driven by spatial probability, and it requires no training. At each region evaluated, a frozen JeV model, meaning one that receives no further training, outputs a probability distribution over candidate regions. The abstract does not give the JeV model's architecture or size.
The paper describes three components that share this distribution. Region Rank prioritises candidate regions, and Sparse Branch sets the branching width according to how many candidates carry probability. Depth Gate adjusts refinement depth using branch sparsity and how concentrated the distribution is.
The abstract states that sharing the distribution "reduces reliance on RLHF-aligned generation" and "eliminates intermediate decoding". The first phrase describes less dependence on generation tuned by reinforcement learning. The second states outright that the intermediate step is removed.
Accuracy gaps across three benchmarks
The three benchmarks are LRS-GRO, MME-RealWorld-RS and LRS-VQA. In the researchers' published figures, GeoJeV scored highest on LRS-GRO and lowest on LRS-VQA.
The same figures put the average accuracy gain at 5.32 percentage points and the added time at 1.12 seconds per question. Those 1.12 seconds are the price of the gain. The abstract does not give Qwen3-VL-8B's accuracy on each benchmark without GeoJeV.
The comparison behind the 7.8-fold speedup
GeoJeV's inference was 7.8 times faster than a recent method, according to the researchers' published figures. The abstract does not name that method. The meaning of the 7.8-fold figure depends on what it is measured against.
The abstract also omits the hardware used to measure the 1.12 seconds and the 7.8-fold speedup. Its conclusion is that replacing text generation with a probability distribution delivered gains in both accuracy and speed. The weight of that conclusion depends on two items: the name of the comparison method and the measurement setup.