July 30, 2026 | Andrew Lawlor

Training 0.71% of a 7B model’s parameters beat Gemini 3.1 Pro zero-shot on the same benchmark — at a fraction of the inference cost.

A two-stage system for verifying whether news images match their captions. LLaVA-1.5-7B was fine-tuned with QLoRA on the hardest subset of the NewsCLIPpings benchmark, where falsified image-caption pairs are constructed specifically to defeat similarity-based detection. The fine-tuned model reached 78.96% accuracy, +11.98 points over the strongest reported baseline, while training only 0.71% of parameters. It also outperformed Gemini 3.1 Pro evaluated zero-shot on the same task (F1 0.764 vs 0.694). The model was then embedded in a Playwright pipeline that scrapes live articles from BBC, Reuters, NYT, WSJ, and AP News and produces real-time verdicts. The practical implication: for a specialized, well-defined task, a fine-tuned open-weight model you run yourself can beat a frontier API — which changes the economics of deployment considerably.

Read the full analysis (PDF) Related: Private Model Training

Conducted as part of graduate AI coursework at Stanford.