EviRover: Reinforcing Agentic Perception Beyond a Glance
Kaixuan Fan, Kaituo Feng, Tianshuo Peng, Yilei Jiang, Manyuan Zhang, Junke Wang, Xiangyu Yue
The Chinese University of Hong Kong · Fudan University
Published September 30, 2026 · on the feed October 1, 2026 · ▲ 3 · arXiv 2609.40230
The abstract
Visual perception is conventionally formulated as a one-shot prediction from a single glance at the image, under the assumption that the image content and the model's parametric knowledge suffice to resolve the query. This assumption often fails in real-world scenarios that hinge on fine-grained visual details or require knowledge-intensive and up-to-date information. We term such cases perception under insufficient evidence and formulate perception as an agentic process that can obtain information beyond a single glance. To address the absence of data for this setting, we design two dedicated data generation pipelines, yielding EviRover-SFT-5K and EviRover-RL-12K for training. We further construct EviLens, a human-verified benchmark comprising 688 instances across five perception categories. Building on these data, we present EviRover, to our knowledge the first perception agent explicitly trained to resolve perceptual queries through interaction, using supervised fine-tuning followed by agentic reinforcement learning. Experiments show that the 4B EviRover outperforms its backbone by 30 points on average on EviLens, reaching performance comparable to advanced proprietary models. The gains transfer beyond EviLens to WebEyes, conventional perception benchmarks, and general multimodal benchmarks, including a 15-point improvement on BrowseComp-VL. All code, models, and data are released.
The Brief
2:22A short podcast summary of the paper. Written and voiced by AI.
Transcript
EviRover. Reinforcing Agentic Perception Beyond a Glance. By Kaixuan Fan and six others, from the Chinese University of Hong Kong and Fudan University. Published on the thirtieth of September. Its code repository is public and real: about thirty-five source files.
Visual perception is conventionally treated as a one-shot prediction. The model takes a single glance at the image and answers, on the assumption that the image plus the model's parametric knowledge are enough to resolve the query. In real-world scenarios that assumption often breaks. Some questions hinge on fine-grained visual details. Others need information that is knowledge-intensive or up to date. In both kinds of case, the answer is simply not contained in what the model sees at first plus what it already knows, and staring harder at the same pixels will not put it there. The authors call these cases perception under insufficient evidence.
Their response is to change the formulation itself. Perception becomes an agentic process, one that is allowed to go and obtain information beyond the first glance, rather than a single prediction that must succeed or fail on what was in front of it. EviRover is, to the authors' knowledge, the first perception agent explicitly trained to resolve perceptual queries through interaction.
Training an agent for this setting needs data that did not exist, and the authors name that absence as the obstacle. So they designed two dedicated data generation pipelines. They produce two training sets, EviRover-SFT-5K and EviRover-RL-12K. The model is trained with supervised fine-tuning first, followed by agentic reinforcement learning. To measure the result they built EviLens, a human-verified benchmark of six hundred and eighty-eight instances across five perception categories.
The four billion parameter EviRover beats its own backbone by thirty points on average on EviLens, which brings it to performance comparable with advanced proprietary models. The gains are not confined to the benchmark the authors built. They transfer to WebEyes, to conventional perception benchmarks, and to general multimodal benchmarks, including a fifteen-point improvement on BrowseComp-VL. All code, models and data are released.
A single glance is a choice, not a law of perception, and this work treats it as one. When the pixels and the weights do not hold the answer, EviRover is trained to go looking, with purpose-built data to learn from and a human-checked benchmark to be judged on. A small model doing that well enough to sit near proprietary systems is the headline result.
Paper Quest · 7 chapters · about 39 min
Play this paper
Learn it by doing: make the calls the authors made, guess the result before you see it, and defend it in a viva graded on your own words. Chapter 1 is right here.
Chapter 1 🧩 · The problem
One glance is all most vision models get. Here, one glance is not enough.
You are asked to outline one person in a group photo. The query never names them: it describes someone who led a biology institute and later coauthored a study.
The road ahead
- 1 · The problem ~6 min Some perception queries cannot be answered from one look at the image, because the needed detail or fact is not in what the model sees or knows.
- 2 · What came before ~3 min
- 3 · The key idea ~8 min Turn perception into a step-by-step search for evidence, ending with a location, boundary or count only once the missing evidence has been gathered.
- 4 · Choose your own research ~3 min
- 5 · Build the method ~8 min Build the missing training data with two pipelines, teach tool use by imitation with supervised fine-tuning, then refine when and how to search with agentic reinforcement learning.
- 6 · The experiment ~6 min You’ll guess: On EviLens, by how many points on average did the 4B EviRover beat its own backbone?
- 7 · The viva ~5 min Three questions from the examiners, answered in your own words. Win it for the paper’s trophy.
Free, with Google or an email address. Your stars and progress are kept.