Quantization beats pruning for language, loses to it for vision on edge devices

The paper's core contribution isn't that compression works differently on different hardware (practitioners already know that). The actual finding is that task structure, not just hardware constraints, inverts which optimization strategy succeeds: quantization dominates language tasks but pruning wins for vision, meaning no single compression recipe exists across domains.

This connects directly to the evaluation contamination risk exposed in the Edge-IIoTset audit from August 16th. That work showed how benchmark artifacts inflate reported performance; this paper reveals a complementary problem: even when benchmarks are clean, they don't predict which technique will work in the field. Together they suggest that practitioners deploying edge models face two validation gaps: contaminated benchmarks AND technique-task mismatches that benchmarks don't surface. The KV-Rescue work on memory optimization also shares this tension between lab constraints and production realities, though it focuses on inference rather than training-time compression.

If Qwen3.5 0.8B with Q5_K_M quantization maintains 93%+ F1 on SQuAD when deployed on actual Raspberry Pi hardware in a production inference pipeline (not just benchmark evaluation), that validates the transfer claim. If the same model's performance drops below 85% F1 in field conditions due to latency or memory pressures not captured in the benchmark, the finding's practical value collapses.