Training-Free AMD Framework Closes 27-Point Agent Gap for Small LLMs Using Hierarchical Teacher Memory

Small LLMs fail at long-horizon tool-use because they cannot generate enough successful trajectories to learn from. AMD sidesteps this entirely—no retraining required—by distilling a large teacher's successful-run memory into three tiers of structured knowledge that a 4B student can consume at inference time, closing a 27-point gap on a hard benchmark.