In plain English
Source-Preserving Prompt Augmentation and Category-Adaptive Inference for Qwen3-1.7B: A Prospective Seed-Replication Study
Testing a compound inference intervention on a small language model without updating its parameters
Can a small AI model get meaningfully better without retraining?
On 96 benchmark tasks, a 1.7-billion-parameter Qwen model improved from 27.2% to 36.0% when the original prompt was preserved and its inference settings adapted to the task. That is a 32% relative improvement with no change to the model's weights. The strongest supported contribution came from changing how the model runs at inference time, not from rewriting the prompt. The trade-off was more tokens, compute and latency.