PathoJev: Efficient Pathology Decisions from Shared Visual Context
arXiv:2610.40752 [cs.CV] · Submitted 29 October 2026 · Computer Vision and Pattern Recognition (cs.CV)
Authors: Chengyang Zhang, Wenchuan Zhang, Bo Li, Xinyu Liu, Jiaming Yang, Mengran Li, Jie Chen, Wei Ju, Yuhao Yi, Hong Bu, Jiancheng Lv
Overview
Pathology workflows demand multiple visual judgments about the same slide—from identifying tumor type to assessing grade and invasion. In conventional multimodal pipelines, generating and parsing a separate textual response for each structured question introduces latency between visual evidence and usable decisions.
Building on the System One paradigm, the authors introduce PathoJev, a pathology model that directly predicts answer probabilities for whole-slide and region-level assessment. Rather than producing free-form text per question, PathoJev encodes each slide once and evaluates independent question suffixes in batches over a shared visual context.
Method
PathoJev's efficiency stems from two design choices:
- Single-pass slide encoding. Each whole-slide image is encoded once, so its visual representation is reused across all downstream questions.
- Batched question suffixes over shared context. Independent question suffixes are evaluated in parallel against the same visual context, avoiding repeated prefix encoding and redundant computation.
- Recurrent state reuse. Reusing the hybrid backbone's recurrent states further limits the computation that grows with slide-context length as additional questions are added.
The result is a reusable probabilistic interface for combining multiple visual judgments across whole-slide and local tissue assessment.
Results
Across three whole-slide benchmarks, the 9B and 27B model variants achieve mean accuracies of 77.2% and 78.1%, exceeding SlideChat by 5.4 and 6.3 percentage points, respectively. Both scales also maintain competitive accuracy on five ROI benchmarks.
Efficiency gains are substantial:
- On WSI-Bench, the 9B model runs 9.1× faster than SlideChat.
- Shared context accelerates eight-question workloads by 6.9× and 5.9× over separate prefix encoding for the two scales, rising to 11.2× and 8.9× for 64-question workloads.
Significance
As computational pathology moves toward multimodal foundation models in 2026, inference cost and latency have become practical bottlenecks for clinical deployment, where a single slide typically triggers dozens of structured questions. PathoJev reframes pathology question answering as batched probabilistic prediction over a shared visual context rather than sequential text generation—a direction that aligns with the broader shift toward efficient, decision-oriented medical vision systems.
Availability
Code is available at github.com/ChyaZhang/PathoJev.
Cite as: arXiv:2610.40752 cs.CV] · DOI: [10.48550/arXiv.2610.40752
via ArXiv CV