Article
CommunityJev Models & Reimplementations
Seeing the Options: Typed Decisions over Visual Candidates with Reusable Option Memory
arXiv paper introducing Iris, a System One design that accepts image, text, or mixed candidates while leaving Jev's decision module untouched, with order-invariance, null-probability, and candidate-reuse results on a Multimodal-Mind2Web benchmark.
- Format
- Article
- Reading time
- 2 min
- Published by
- Community
- Added
- Oct 2026
Highlights
- Iris admits image, text, or mixed candidates and leaves Jev's decision module unchanged.
- Independent encoding reaches 78.4/72.1/74.8% across three splits with a flip rate of exactly zero.
- Joint encoding with OneJev-4B reaches 90.6/89.5/91.6% but changes answers on 8-10% of reordered records.
- A third design reaches 92.1/87.0/90.2% at a third of OneJev's latency with null AUROC 0.93-0.97.
- Reusing image candidates across requests cuts latency 37x at K=256 and 191x at K=2048.
Questions about this entry
Short answers drawn from the linked project or page.
It asks what happens to the four properties System One decision models need - a none-of-the-above probability, invariance to candidate order, scalability in candidate count, and reuse of candidate computation - when the candidates themselves are images rather than text in letter slots.
OneJev and Jev-Omni put images in the request while candidates stay text in letter slots. Iris instead admits image, text, or mixed candidates and leaves Jev's decision module untouched.
It builds a visual-candidate benchmark from Multimodal-Mind2Web in which every candidate is the element's screenshot crop, with K=16 candidates and three splits.
Independent encoding reports a flip rate of exactly zero across the three splits. Joint encoding with a fine-tuned OneJev-4B changes its answer on 8-10% of records when the crops are reordered.
The design that encodes each candidate after the state, but never after another candidate, reaches 92.1/87.0/90.2% at a third of OneJev's latency, with a calibrated null probability at AUROC 0.93-0.97.
No. Set attention, token-level reads, and multi-vector memory buy nothing, because the encoder maps visually distinct elements to vectors within cosine 0.999 of each other.
The paper points to code in the lllfx/Iris repository on GitHub, alongside the benchmark built from Multimodal-Mind2Web screenshot crops.
Reactions & coverage
Posts, threads, and videos about this entry from around the web.
X: Standard One SH (StandardThinking)
LinkedIn: What multimodal decision models actually change, and what they do not.
LinkedInJevDirectory.org
Hacker News: Decision Index 0.2.1 scores open reproductions of Jev on a frozen suite.
jackbravo4 pointsDecision Index 0.2.1 – 70 open reproductions of TypeSafe Jev's Decision ModelRead the thread on Hacker NewsHacker News: Pointer to a DiffusionGemma plus vLLM tutorial for a Jev-shaped classifier.
1 pointHow to build a Jev-style classifier with DiffusionGemma and vLLMRead the thread on Hacker NewsMore Jev resources like this
Open-weight vision-language judge that takes a task instruction plus 10 frames of a robot manipulation attempt and predicts success or failure with calibrated probabilities in a single forward pass. Released with inference code and a 20,628-prediction evaluation set.
Jev Models & ReimplementationsArticle#arxiv#open-models#evaluation
Community
Apache-2.0 Hugging Face repository for RSI-JEV-27b, a 27B multimodal decision model based on Qwen3.8-27B that answers supplied multiple-choice questions over images, returning a selected option and probabilities for all options in one forward pass.
Jev Models & Reimplementations#open-models#benchmark#calibration
Community
A family of 0.8B-9B ophthalmic decision models that score candidate answers without autoregressive decoding, trained on RetiJev-Data: 78,158 clinical-decision and diagnosis questions and 50,074 guideline-grading questions. RetiJev-9B answers a case in 158 ms on one GPU.
Jev Models & ReimplementationsArticle#arxiv#open-models#benchmark
Community
Shows that open instruction-tuned models already read decisions from answer-letter logits without training, and that LoRA fine-tuning with a decision head lifts them past Jev: Gemma 4 31B goes from 79.6 to 84.4 on the new 2,352-question OpenDecisions set and scores 81.5 on JevBench v1.7.
Jev Models & ReimplementationsArticle#community#research#arxiv
Community
A System One model for Chinese-language decisions trained on a 10-million-example corpus and fine-tuned for medicine, legal, and finance. Pre-training exceeds closed-source Jev by 1.24% on general tasks with a 20.3x speedup; domain fine-tuning beats Jev in medicine by 4.0%.
Jev Models & ReimplementationsArticle#community#research#arxiv
Community
1.4kGitHub stars
Reads a typed decision from one prefill of a pretrained instruction-tuned model by restricting the next-token distribution to option tokens, then corrects label and position bias with no gradient steps. Averaging over cyclic rotations lowers order-flip rates and raises accuracy on 11 of 11 models.
Jev Models & ReimplementationsArticle#community#research#arxiv
Community
nokia-applied-research/AnyJevPython
nokia-applied-research/AnyJevPythonFrom the Jev guides
Original write-ups on what this entry is for.
From the Jev AI model community
Posts from builders shipping with Jev right now.