PlayJev: A Multimodal JEV-Like Model
for Small Games
Gallery
Ten games, one set of weights, ten different sets of moves to choose from.
Results
Sixteen held-out episodes per game, the same seeds for every player, the most probable move taken with no sampling. The bar is where the model lands between random play and the scripted teacher that labelled its training data.
the same numbers as a table
| game | random | the model | teacher | vs teacher |
|---|
Does it know when it is right?
Each dot takes the steps where the model gave its chosen move about the same probability, and asks how often that move was the teacher's. On the diagonal, the number it reports is the rate it achieves.
Knowing when to ask for help
The model plays every step it is sure about and hands the rest to the teacher. Two boards of the same game from the same seed: on the left the model plays alone, on the right it keeps the steps it is confident about and the teacher takes the others, shown in coral.
What the confidence is worth
The horizontal axis is the share of steps the teacher actually took, so the left end is the model alone and the right end is the teacher alone. Picking those steps by confidence is the line that has to beat handing over the same share at random, or the confidence means nothing.
Learning a new game
Two games were held out of an eight-game model. Fine-tuning that model on a few thousand frames of a held-out game is compared against fine-tuning the raw base model on the same frames. The line is how often the fine-tuned model picks the teacher's move on 2000 frames of that game it never trained on.
every run as a table, with the scores
| game | starts from | frames | agreement | score | vs teacher |
|---|
How it works
The model sees the frame and the option list, rendered with the frozen OpenJev prompt, and its answer is the next token after "Answer:". The probability over moves is the softmax over the option letters at that position: one forward pass, no sampling, nothing generated.
The most probable move is played, with one exception: a move that leaves the picture unchanged, such as a blocked direction in 2048 or Sokoban, is not repeated on that same picture, and the next most probable move goes instead.
Each board here is the vendored game running with the shim and hook used in training, a virtual clock, a seeded random source and synthetic key events, so the recorded action sequence reproduces the episode step for step.
The games
Ten open-source browser games, vendored with their own licence and every change we made recorded next to them. The art in Infinite Mario, Floppy Bird and the Racer belongs to its original owners, and the licence below covers the code each author wrote.