Dense Coordinate-List Fine-Tuning Induces a Controllable Interference Surface in Vision-Language Models

Fine-tuning vision-language models to produce dense coordinate lists sharpens visual grounding but also alters how models serialize, repeat, and terminate structured outputs, according to a preprint posted to arXiv on June 12, 2026 [1]. The study examines this behavior as a generation and control surface [1]. In experiments with the Gemma 4 12B model, applying high-capacity q/k/v/o LoRA adapters lifted class-aware [email protected] from 0.007 to 0.448 while introducing repeated-tail pressure, with a duplicate rate of 0.080 and a maximum repeat of 23 [1]. A sweep of q/v ranks kept the maximum repeat at 21–22 across ranks 4–64, indicating that the interference persists across adapter capacities [1]. The target signal proved separable. An object-level repeat-stop intervention eliminated exact repeated records, driving the duplicate rate to 0.000 and the maximum repeat to 1, while preserving F1 scores of 0.494 to 0.490 and stricter [email protected] scores of 0.381 to 0.385 [1]. Structure-axis probes localized the effect to bounding-box coordinate object lists; dense non-bbox outputs and spatial or count JSON remained repeat-clean, even under high-capacity adapters [1]. Cross-model reproduction reinforced the findings. The Qwen3-VL-8B model reached a clean controlled endpoint with an [email protected] of 0.318 and a duplicate rate of 0.000 [1]. The COCO 2017 dataset, a widely used benchmark in computer vision, reproduced both the acquisition of grounding capability and the duplicate pressure [1]. The work was posted on arXiv, an open-access repository that hosts preprints across physics, computer science, and related fields and has grown to receive roughly 24,000 submissions per month as of late 2024 [8]. The authors frame dense coordinate-list adaptation as creating a structure-bound, cross-family interference surface that can be measured and controlled [1].