$ cat wiki/papers/2026/2610.12419-onesearch-vl.md
OneSearch-VL — traceable evidence for visual research
TL;DR
OneSearch-VL connects visual anchors, entity relations, cited facts and answer-producing operations in a shared evidence graph. The authors use that graph for training data, RL supervision and operation-level evaluation. (source)
Authors & Org
Hongyu Li, Manyuan Zhang, Kaituo Feng, Shu Chen, Dian Zheng, Hao Li, Hao Yu, Zhangquan Chen, Zoey Guo, Ray Zhang, Shaofei Huang, Tianrui Hui, Linjiang Huang and Si Liu. First submitted 2026-10-08. Affiliations are not stated in the captured abstract page. (source)
Method
The Visually Grounded Evidence Graph (VGEG) preserves dependencies across visual grounding, external retrieval and fact composition. Its data engine constructs and verifies multi-image and video questions and filters expert trajectories, producing OneSearch-VL-SFT-110K and OneSearch-VL-RL-10K. The Evidence-aware Visual-Grounded Rubric reward (EVGR) supervises evidence traceability and visual grounding during RL. (source)
Results
OneSearch-VL-8B improves over Qwen3-VL-8B with tool access by 20.2 percentage points on OneSearch-MI-Bench and 17.6 percentage points on OneSearch-Video-Bench, both newly introduced by this work. The abstract also reports gains across 7 image benchmarks and VideoDR without numerical effect sizes. (source)
Significance
For Agents (LLM Agents), this makes the relation between an observation, a retrieved source and an answer an explicit training target. Interpretation: the method addresses evidence traceability rather than treating a correct final answer as sufficient evidence of a sound research process. (source)
Open Questions
Absolute scores, inference costs and the effect of each training component are not stated in the abstract. The largest quantified gains are on the authors’ own benchmarks, so independent replication and evaluation on other tasks remain important. Full paper and implementation were not reviewed. (source)
Cite
arXiv:2610.12419 · Code · (source)