Jev in a driving simulator: collisions fell once it knew the consequences

Three researchers at Purdue University published a preprint on 3 November 2026 in which jev-1.13.0, called through its HTTP API, selects maneuvers for a car in a simulator. S83 Driving is a long way from a Slovak online shop, but one difference stands out: the same model collided in 26.0% of episodes in one test, and in 2.4% once the predicted consequences of each option were added. S83 The model stayed the same; what changed was its input. This article is based on the full text of the preprint, which contains no information about peer review.

Three layers: command, foresight, choice

The system is called Jev-AD. When a passenger says something, Jev answers four typed questions in one call: whether the command asks for something unsafe, whether it is off-topic, how aggressive it is on a scale of 0 to 4, and which maneuver it requests. With a probability of unsafe of at least 0.8 the system refuses the command, in the band from 0.2 to 0.8 it asks for clarification, and with a probability of at least 0.5 that the command is off-topic it ignores it. The authors state that they set these thresholds before the evaluation. S83

The second layer is the prediction. For each candidate action, a Qwen2.5-0.5B encoder with two MLP heads estimates how nearby vehicles will move over the next 2 seconds. Ordinary code then works out collisions, minimum time to collision (TTC), and the final lane and speed, and writes the result as short phrases. A TTC below 1.5 s is labeled Danger. S83

In the third layer, Jev picks one of five actions every 0.5 s: Left, Idle, Right, Faster or Slower. It is told to avoid options marked Collision or Danger when a safe choice exists. The authors say explicitly that this is not a hard filter. In merging, the model followed it in 98.7% of decisions where a safe choice existed. S83

What the authors measured

The simulated merging in highway-env covered 500 episodes, counted per car (50 scenes, 5 commands, 2 cars). Jev with the text scene description from the KoMA reference agent collided in 26.0% of episodes. With foresight it dropped to 2.4%, with a 95% confidence interval of 0.6 to 4.8%. In traffic replayed from NGSIM recordings, the ratio was 48.2% against 2.8%. S83

What did not improve also deserves attention. A simple constant-velocity prediction gave 2.8% in merging and 1.8% on the NGSIM recordings, so in the second case less than the trained model. According to the authors, the drop came from having a prediction at all, not from the method used to compute it. With the same foresight, GPT-6-luna reached 0.6% in merging and GPT-4o-mini 2.2%, so Jev was not the safest choice in this test. S83

For command understanding, the authors used a set of 138 commands: 23 unsafe, 105 safe and 10 borderline with no correct answer. In this test, Jev refused all 23 unsafe commands and none of the safe ones, and identified the maneuver with an accuracy of 98.9%. S83

Jev's median latency was 136 ms, against 770 ms for GPT-4o-mini and 1.49 s for GPT-6-luna; foresight added a median of 124 ms. The table gives a cost per 1,000 calls of USD 0.030 for Jev, USD 0.065 for GPT-4o-mini and USD 0.079 for GPT-6-luna. The verbal comparisons of speed and cost in the preprint's text do not match this table, so only the table values are given here. S83

What an online shop can take from it

The authors explain the good result by the fact that the model no longer had to derive gaps and speeds from coordinates on its own. In their view, a typed decision model judges well what the input says and fails when the right choice depends on what the input does not say. S83 What follows is a proposal by JEV.sk, not a conclusion of the preprint.

For a Slovak online shop, this suggests a simple approach. If Jev is to choose how to handle a return request, let code first determine whether the deadline is still running, what state the order is in and what each option would mean. Send these results in state as short sentences, not as raw dates and database rows. Jev then judges the meaning, and Choice, Score and Noul get a question they have the material to answer.

The split into three bands can be adopted too. For a question such as "is this fraud?", a high probability can lead to rejection, a medium one to clarification or escalation to a human, and a low one to continuing. Set the thresholds before the evaluation, on your own data, as the authors did. S83

The third lesson is caution with prohibitions. In the preprint, the ban on unsafe options was only an instruction, and the model did not always follow it. S83 Refunds or order changes in an online shop should therefore be guarded by code, not by text in the question.

Limits of the study and the model version

The authors ran no test on a real vehicle. The agent reads exact states from the simulator instead of sensor perception, and it chooses among only five actions with 2 seconds of foresight. The prediction model was trained separately for each setting and was not tried across settings. Adversarial commands were checked only through the unsafe requests in the HG-Cmd set. S83 The numbers therefore say nothing about safety in real traffic.

The preprint studied jev-1.13.0, while the alias jev-latest now points to jev-1.14.0. S01 The authors themselves note that the behavior of a proprietary model can change between versions. S83 The documentation recommends pinning a specific version if you need stable results. S01