PathoJev-27B Beats SlideChat by 6.3 Percentage Points in Average Accuracy on Three WSI Benchmarks

Researchers from Sichuan University, West China Hospital, the National University of Singapore and Sun Yat-sen University have introduced PathoJev, a digital pathology model that answers questions about a histology slide with a ready-made probability distribution over the answer options rather than with text. The whole-slide image (WSI) is read once, and any number of questions about it are then processed on top of the shared visual context. On three whole-slide benchmarks, the 9B and 27B versions beat SlideChat, the previous model for slide-level questions, in average accuracy by 5.4 and 6.3 percentage points respectively. Against WSI-LLaVA, another model in the same class, the lead is 7.0 and 7.9 points.

Why pathology needs direct probability output

A pathologist makes several judgments from a single slide at once: tumor type, grade, presence of invasion. Current pathology vision-language models, such as Quilt-LLaVA, PathChat, SlideChat and Patho-R1, answer through autoregressive text generation. For a question with predefined options, this means sequential decoding followed by extracting the answer from the generated string. Such generation does not directly produce a normalized distribution over the options. If there are several questions about a slide, the whole cycle repeats for each one, even though they share the same tissue data.

PathoJev is built on the System One paradigm: the input is mapped directly to a typed probabilistic judgment. Each question has an output type and an ordered set of answers, and the model picks the option with the highest probability. Binary questions use a two-answer set; multiple-choice questions use text alternatives.

One slide, many questions

A frozen CONCH v1.5 encoder extracts features from the slide's tissue patches. These are averaged over a spatial grid, and a trainable multilayer perceptron projects the result into the embedding space of the language backbone. This shortens the visual sequence while preserving its spatial organization. For ROIs, meaning local tissue regions, the base model's native visual pathway is kept. Training uses questions on both whole slides and ROIs, so the shared decision interface covers both levels.

Computation is split into a shared prefix containing the slide and separate suffixes for individual questions. The questions branch off the prefix independently and are processed as a batch. The hybrid backbone uses recurrent states, so the share of compute that grows with slide context length shrinks with each additional question. Prefix caching saves re-reading the slide, but standard softmax attention still accesses keys and values for every new question. To compare the cost of an additional question, the authors used Qwen3-8B, which relies only on softmax attention.

What the measurements showed

Beyond the lead over SlideChat and WSI-LLaVA, both versions retain competitive accuracy on five ROI benchmarks. On WSI-Bench, PathoJev-9B runs faster than SlideChat. Latency here is the GPU forward-pass time per question, compared on WSI-Bench and PathMMU-tiny. The zero-shot WSI baselines in the comparison worked with slide thumbnails.

A separate control tests the interface itself. The authors trained an autoregressive variant on the same data, and PathoJev-9B proved more accurate than it on six of eight benchmarks, with lower latency on both WSI and ROI tasks. On a workload of eight questions, the shared context cuts total time relative to encoding the prefix separately, for both the 9B and 27B models, and the gain grows at 64 questions. The exact speedup percentages are best checked in the paper itself.

What this means in practice

For developers of medical AI systems, the main takeaway is the architectural pattern. If one input comes with many structured questions, it is more efficient to encode the context once and read answer probabilities directly. The model's confidence in each option is available immediately, and it can be used for thresholds and for routing cases to a physician.

The limitations are also clear. Direct probability output suits questions with a predefined set of answers and does not replace free-form descriptions or extended diagnostic reasoning, as in Patho-R1. The results come from academic benchmarks, not from real streams of clinical reports. The project code is published on GitHub in the ChyaZhang/PathoJev repository.