Small AI models running on satellites keep ignoring the step-by-step plans they're supposed to follow. A new decoder framework fixes that without any retraining.
Researchers built PTC-Decoder, a plug-and-play system that sits between a small language model (SLM) and its output. It treats planning as a required first move rather than a suggestion, forcing the model to generate a plan by making that the very first tool it must call. A second piece, called TC-Decoder, then restricts which tool names the model can output at each step using a strict rule-based filter, while still letting it fill in the actual parameters freely. Tested on 200 real remote-sensing satellite tasks across seven small models, the combination produced a statistically significant improvement in task scores.
The catch with small models is that they run where big cloud models can't: offline, resource-constrained hardware like satellites, drones, or field sensors. Prior plan-then-solve approaches just asked models nicely, via prompt instructions, to make a plan first, and the paper found weak models mostly ignored that request. Constraining what a model can call, without touching how it reasons, is a narrow but genuine fix for a reliability problem that shows up whenever these agents work unsupervised.
The reported gain, +1.21 on the paper's own scoring scale, is modest, and the authors admit final-answer accuracy is still an open problem. This makes small models behave, it doesn't make them smarter.