Article
CommunityJev Demos & Experiments
Jev-AD: Safe Language-Guided Autonomous Driving via a Language World Model
Jev-AD pairs Jev, a typed decision model that answers predefined questions with probabilities rather than generated text, with a language world model that predicts how nearby vehicles move over the next two seconds for every candidate action. Evaluated in simulated highway merging and 500 replayed NGSIM episodes.
- Format
- Article
- Reading time
- 2 min
- Published by
- Community
- Added
- Nov 2026
Highlights
- Pairs Jev's typed probabilistic decisions with a language world model predicting nearby vehicle motion two seconds ahead.
- In 500 simulated highway merging car-episodes, foresight lowered the same decision model's collision rate from 26.0% to 2.4%.
- Reported decision latency is 0.25 seconds per decision.
- Jev-AD achieved the passenger's intent in 93% of free-form command trials.
- In 500 replayed real-world NGSIM episodes, collision rate fell from 48.2% to 2.8%, and the world model foresaw 93% of dangerous actions.
Questions about this entry
Short answers drawn from the linked project or page.
Jev-AD is a language-guided autonomous driving system that pairs Jev, a typed decision model answering predefined questions with probabilities, with a language world model. Code turns the world model's predictions into foreseen consequences, which Jev compares instead of inferring them from coordinates.
For every candidate action, the language world model predicts how nearby vehicles move over the next two seconds. Code converts those predictions into foreseen consequences that Jev compares directly, rather than having Jev infer each action's consequences implicitly from the scene.
Across 500 car-episodes of simulated highway merging, adding foresight lowered the same decision model's collision rate from 26.0% to 2.4%. Jev-AD also achieved the passenger's intent in 93% of free-form command trials.
In 500 replayed real-world NGSIM episodes, foresight lowered the collision rate from 48.2% to 2.8%. The world model, learned from this traffic, foresaw 93% of dangerous actions.
The digest reports 0.25 seconds per decision for the simulated highway merging evaluation. No other latency figures are stated for the system.
No. Jev is described as a typed decision model that answers predefined questions with probabilities, not generated text. The language world model supplies predicted consequences for candidate actions, which Jev compares instead of inferring them from coordinates.
The digest references a project page for Jev-AD but states no license, repository, or installation instructions. It also reports no on-road deployment, so the described results come from simulated highway merging and replayed NGSIM episodes.
More Jev resources like this
Hugging Face dataset of 80 generated rollout videos (422 MB) supporting studies on Jev-routed temporal granularity for causal video generation, produced with Causal-rCM on Tesla T4 and A100 GPUs.
Jev Demos & ExperimentsVideo#research#video#demo
Community
A Japanese-language experiment workspace that asks Jev, through typesafe-sdk-python's jev-latest, why a dummy restaurant's sales fell, and compares Jev against a local Ollama model on accuracy, speed, answer variance, confidence, and cost.
Jev Demos & Experiments#github#python#demo
Communitytaiki00wo/jev-playground
arXiv study that evaluates ride-hailing rebalancing policies under the EU Platform Work Directive's disclosure requirement, using self-interested LLM driver agents built on Claude Haiku 5.5 and Jev in a three-zone model and in Manhattan.
Jev Benchmarks & EvaluationsArticle#arxiv#research#agents
Community
A Python demo that screens PubMed abstracts for one systematic review question using Jev. Six narrow questions per abstract go to Jev in a single call, with rule-based tiers and confidence gates deciding include, exclude, or human review.
Jev Demos & Experiments#demo#python#madewithjev
Communityanondo1969/jev-evidence-triage
arXiv paper on a small candidate formalizer trained on offline teacher-generated lists that builds explicit, validated alternatives from task context for a frozen decision model, evaluated on CALVIN-D episodes and annotations.
Jev Models & ReimplementationsArticle#arxiv#research#benchmark
Community
arXiv paper introducing S²-MAS, a society-scale multi-agent simulation framework built on 500 million synthetic household personas from survey data. Its Genese engine separates decisions from state updates, and sharp-421m-mlx is benchmarked against Jev 1.13 on financial-choice tasks.
Jev Benchmarks & EvaluationsArticle#arxiv#evaluation#benchmark
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.