New 4B AI model uses reinforcement learning to precisely identify image parts

Multimodal large language models (MLLMs) excel at grounding whole objects from natural language queries, but they systematically fail when the query names a specific part (e.g., "car wheel" vs. "car"). The authors trace this limitation to the lack of an object-part hierarchy: models treat parts as independent objects, localizing them in a single step. To solve this, they introduce Object-Part Hierarchical Reflective Grounding (OP-HRG), a two-stage strategy that first identifies the parent object's bounding box, then refines it to locate the named part. A self-check module evaluates the result and, if needed, re-encodes a tighter crop for correction.

Training is done via a novel part-aware GRPO (Group Relative Policy Optimization) framework, a reinforcement learning approach that assigns separate rewards for object detection and part localization stages. Despite using only 4B parameters, the OP-HRG model outperforms larger 7B grounding LLMs and the widely used SAM3 across three benchmarks: PascalPart, PartImageNet, and InstructPart. The approach also transfers effectively to reasoning segmentation tasks, demonstrating its generalizability. The work, accepted at ECCV 2026, highlights how structured reasoning and RL can significantly improve fine-grained visual understanding without scaling model size.