A new training framework called DuoMem shows that a small language model can learn complex, multi-step reasoning from a larger one — well enough to run usefully on a phone or edge device.
Researchers introduced DuoMem as a way to close the gap between large cloud-side AI agents and the compact models that can actually fit on constrained hardware. The framework works along two axes: it prepends high-quality memories generated by a large "teacher" model to the smaller "student" model's input, and it fine-tunes lightweight LoRA adapters on successful teacher problem-solving trajectories. On ALFWorld, a benchmark for embodied decision-making tasks, a 4B-parameter model trained with DuoMem jumped from a 4.3% to a 77.9% task success rate. The 72B teacher it was distilled from scores 87.1%. The whole upgrade costs fewer than 10 million trainable parameters and a few megabytes of stored memories.
The speed gap is the more immediately practical finding: the 4B DuoMem model completes tasks more than three times faster in wall-clock time than the 72B teacher, which matters a great deal when the use case is a real-time assistant running without a cloud connection. Most efforts to shrink AI agents have focused on quantization or pruning — DuoMem's dual-axis distillation approach, targeting both context and parameters, is a less common combination and the benchmark numbers suggest it transfers procedural skill more efficiently than either axis alone.
The obvious question is how well this holds outside a controlled benchmark. ALFWorld is a text-based simulation, and the gap between simulated household tasks and the messier real world has tripped up embodied AI research before.