Snacks Are All You Need

1. Introduction

Modern pug training pipelines are elaborate. A typical curriculum spans four stages, nine specialists, and a clicker, and still fails when the dog notices that the person holding the clicker is also holding a phone. We ask whether the pipeline is necessary, or whether it is the snack doing the work.

Prior work has treated the snack as a reward signal. We treat it as the whole signal.

2. Method

We introduce Single-Snack Training (SST). The model receives one snack, delivered inside a 0.4-second window that begins when the desired behavior is first approximated and ends when the model loses interest. No other supervision is provided. The curriculum baseline received the standard four-stage program over six weeks.

• Delivery window: 0.4 seconds, measured by a witness who was also hungry.
• Snack: one (1), unbroken, of a size the model could identify from across the room.
• Evaluation: Sit, Stay, Paw, and Come, each requested once, in a quiet room, with the door closed.

3. Results

SST matched or exceeded the curriculum baseline on all four tasks. The largest gain was on Come, where the baseline lost most of its trials to a squirrel that was not part of the experimental design. SST was unaffected, because SST had already received the snack.

We call the remaining gap the Dinner Bottleneck. In every condition, performance dropped sharply within an hour of the evening meal, regardless of training method. We were unable to remove the bottleneck, and we are not certain we want to.

4. Limitations

The snack was eaten. This makes the experiment impossible to reproduce without a second snack, and a second snack is a different experiment. We encourage replication, and we encourage it to be funded in snacks.

5. Conclusion

A single, well-timed snack is sufficient to elicit the full range of behaviors we measured. We discuss implications for scaling, which are large, and for dinner, which is at six.