Article
CommunityJev Models & Reimplementations
Improving JEV Entailment Reranking with Span Readouts and Selective Adaptation
arXiv paper testing whether the readout and adaptation choices used for direct selection still hold up in cross-request entailment reranking for JEV decision models. It combines a mean-span readout with selective backbone adaptation and reports accuracy on MC11, an average over eleven reranking configurations.
- Format
- Article
- Reading time
- 2 min
- Published by
- Community
- Added
- Oct 2026
Highlights
- Separates direct selection, which compares candidates in one request, from entailment reranking, which scores each candidate in its own NLI request.
- Mean-span pooling averages the contextualized states inside each NLI option instead of using an option-ending pointer.
- Selective adaptation keeps low-rank updates in attention and gated DeltaNet projections while removing feed-forward adapters.
- On MC11, selective adaptation improves NeoHorse-1-4B accuracy over the all-module baseline while cutting trainable parameters by more than half.
- Mean-span pooling adds no head parameters and yields further consistent gains at the same parameter count.
Questions about this entry
Short answers drawn from the linked project or page.
It asks whether the readout and adaptation choices used for direct selection remain effective when the same models are used for cross-request entailment reranking over fixed relation options.
Direct selection compares candidate answers within a single request, while entailment reranking evaluates each candidate in a separate natural language inference request over fixed relation options.
It is a readout that replaces the option-ending pointer with an average of the contextualized states inside each NLI option, and it does not add any parameters to the scoring head.
It retains low-rank updates in the attention and gated DeltaNet projections while removing the feed-forward adapters used by the all-module baseline.
MC11 averages accuracy over eleven reranking configurations drawn from seven sources, giving a single aggregate number for comparing the readout and adaptation variants.
Results are reported for NeoHorse-1-4B, where selective adaptation improved accuracy over the all-module baseline while cutting trainable parameters by more than half.
No. Mean-span pooling yields further consistent gains at the same parameter count as selective adaptation, because it adds no head parameters.
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
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
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
Hugging Face dataset of Jev-shaped retrieval and classification decisions built from MM-BRIGHT, holding Qwen3.6-35B-A3B-FP8 state, action, and hard-negative embeddings for post-training CLM heads.
Jev Models & Reimplementations#open-models#benchmark#evaluation
Community
Hugging Face dataset that renders HoVer claim-verification items as Jev Decision Index retrieval and classification requests, with Qwen3.6-35B-A3B-FP8 state, action, and hard-negative embeddings for post-training CLM heads.
Jev Models & Reimplementations#classification#benchmark#open-models
Community
Hugging Face dataset of Qwen3.6-35B-A3B-FP8 state/action embeddings for contrastive post-training on the Jev Decision Index's entailment and verification train splits: ANLI, ContractNLI, NLI4CT, HoVer and RAGTruth.
Jev Models & Reimplementations#open-models#calibration#research
Community
SecJev is a family of security-specialized System One decision models from 0.8B to 9B parameters, built on Kev's single-pass candidate scorer. It spans 14 tasks and eight sources, and SecJev-0.8B beats general Kev-9B by 20.51 points in task-macro accuracy.
Jev Models & ReimplementationsArticle#community#research#arxiv
Community
From 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.