RetiJev: Fast and Source-Grounded Ophthalmic Decision Models
RetiJev tackles the latency, cost, and privacy barriers of autoregressive LLMs in point-of-care ophthalmic decision support. The authors introduce RetiJev-Data, a 128k-question source-grounded corpus, and a family of 0.8B–9B compact models that score predefined candidate answers in a single pass instead of generating free-form text. This yields sub-200 ms inference per case on one GPU while reportedly surpassing a 27B LLM and a commercial decision model on guideline grading and clinical-action accuracy.
RetiJev presents a practically compelling case for replacing generative LLMs with discriminative scoring in structured ophthalmic tasks, and the dataset construction is impressively thorough. However, the absence of human clinician adjudication at any stage of label curation or evaluation remains a critical liability for a system intended for clinical use.
The efficiency gains are sharply quantified and the shared pointer-head architecture is a principled fit for variable candidate sets. Most notably, guideline-grading labels are computed deterministically from published clinical criteria and verified by independent parameter re-extraction, giving that portion of the benchmark the strongest grounding, while explicit train-test partitioning by rule families creates a meaningful unseen-decisions generalization test.
The entire pipeline is synthetic: patient descriptions, questions, and the majority of labels are generated by LLMs, with verification performed only by other LLMs. This circular supervision is vulnerable to the very self-preference biases and correlated errors the authors note in prior work. Additionally, the comparison to the 27B LLM baseline is confounded by the use of that same Qwen3.8-27B model as a distillation teacher for clinical-decision training.
Finally, reducing complex clinical judgments to a fixed catalog of 53 predefined actions risks oversimplifying nuanced management, and the near-total lack of human clinician review beyond catalog approval undermines claims of clinical reliability.
The guideline-grading gains are the most credible because labels are rule-computed rather than model-generated, and the large accuracy margins over Jev 1.13 are consistent across model sizes. The clinical-decision results are harder to interpret: the benchmark itself is synthetic and model-verified, and the 0.8B model exceeds the 27B baseline on seen-decisions while only matching it on unseen-decisions, suggesting some memorization of synthetic templates. The paper also omits comparisons against simpler encoder-based medical classifiers or retrieval-augmented pipelines, leaving it unclear how much of the gain is unique to the Jev paradigm versus task-specific fine-tuning.
Model weights and code are publicly released, which is a strong positive. Yet independent reproduction of the full training pipeline would be difficult: the 94k clinical-decision examples depend on specific LLM prompts, a multi-stage filtering and cross-model adjudication protocol, and a 60-entry answer catalog whose clinical validity is asserted but not detailed in the main text. Hyperparameters and input encodings are relegated to Appendix J, and the exact LLM versions and temperature settings used during synthetic data generation are not specified here.
Ophthalmic clinical decision support requires fast, accurate, and privacy-preserving inference, yet autoregressive large language models (LLMs) incur substantial latency and computational overhead for decisions among predefined actions. We therefore investigate a decision-oriented alternative to generative reasoning. Specifically, we first introduce RetiJev-Data, a source-grounded ophthalmology corpus comprising 78,158 clinical-decision and diagnosis questions derived from articles in PubMed Central and a MeSH vocabulary covering 531 eye diseases, and 50,074 guideline-grading questions spanning 11 grading systems, along with a 24,662-question evaluation benchmark. Based on this, we develop RetiJev, a family of 0.8B-9B decision models that directly score candidate answers without autoregressive decoding. Across all model sizes, RetiJev outperforms both a 27B LLM and the commercial Jev 1.13 decision model on guideline grading by 10.4-23.0 points. RetiJev-9B further achieves the highest accuracy on clinical decisions, improving over both baselines by 5.4-6.5 points while matching them on diagnosis. It answers a case in 158 ms on one GPU, 3.5x faster than LLM generation and 470x faster than reasoning. Models are released at https://github.com/Awenbocc/RetiJev.
Pick a starting point or write your own. Challenges run in the background, so you can keep reading while the AI investigates.
No challenges yet. Disagree with the review? Ask the AI to revisit a specific claim.