Publications

Compared three open-weight models (Gemma 4 E4B, Qwen 2.5 7B, Llama 3.1 8B) fine-tuned on 1,467 USCIS appeal decisions. Found that LoRA produces architecture-specific minority-class specialization rather than uniform improvement — Gemma LoRA achieved the strongest macro-F1 (0.341), the only configuration to beat the always-dismiss baseline.