3B Unified Model Matches Specialists on Seven Counting Benchmarks

ABACUS adapts a 3B-parameter vision-language foundation model for counting and count-faithful image generation using density-aware adaptive zooming, a boundary-aware GRPO policy, and a cycle-consistent self-critique strategy requiring no external annotations. It achieves state-of-the-art across seven benchmarks, outperforming larger task-specific models without benchmark-specific fine-tuning.