Research map / Vision-language and multimodal / Multimodal LLMs

RegionReasoner: Region-Grounded Multi-Round Visual Reasoning

ICLR 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Chain-of-thought multimodal reasoning · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Introduces RegionReasoner, a GRPO-trained Qwen2.5-VL-7B that produces scene/focus/think/answer traces with reference-box citation and global-local keyword-consistency rewards, plus the RegionDial-Bench multi-round detection/segmentation benchmark built from RefCOCO+ and RefCOCOg.

Problem

Single-step visual reasoning and single-turn grounding models do not iteratively reference regions localized in earlier turns, so reference propagation and global-local semantic coherence degrade over multi-round dialogues.

Approach

Requires each reasoning trace to explicitly cite reference bounding boxes and adds a keyword-overlap reward aligning global scene and region captions with the reasoning text, optimized with GRPO on multi-turn rollouts.

Key results

• RegionReasoner-7B averages 80.7 AP on RefCOCO+ Multi-turn vs 74.8 for VisionReasoner-7B and 78.2 vs 73.6 on RefCOCOg Multi-turn.
• RegionReasoner-7B averages 69.6 gIoU on RefCOCO+ Multi-turn segmentation vs 64.3 for VisionReasoner-7B, and 66.5 vs 59.9 on RefCOCOg.
• On V*, RegionReasoner-7B reaches 72.77 overall vs 69.63 for VisionReasoner-7B and 67.53 for Qwen2.5-VL-7B.

Surprising findings

• Single-round ablation shows the consistency and logic rewards slightly lower single-round scores than ref-cite alone (e.g., 88.1 vs 88.6 on RefCOCO+ detection), gains appear mainly in multi-round settings.

Limitations

• Consistency reward relies on lightweight keyword extraction and a hand-crafted logic prior that may miss paraphrases.
• Grounding is limited to boxes and points rather than full masks; segmentation is evaluated via point outputs.
• Evaluation limited to RefCOCO+/RefCOCOg-derived dialogues up to 7 rounds; no longer dialogues or video.

Experimental setup

Datasets: RefCOCO+ RefCOCOg VisionReasoner training data V* benchmark

Benchmarks: RegionDial-Bench (RefCOCO+ Multi-turn, RefCOCOg Multi-turn) V* RefCOCO+ RefCOCOg

Metrics: AP50 gIoU accuracy

Baselines: Qwen2.5-VL-7B Qwen2-VL-7B Seg-Zero-7B VisionReasoner-7B SegLLM-7B

Builds on: VisionReasoner SegLLM GRPO Qwen2.5-VL

Closest papers by meaning

• VLM-R³: Region Recognition, Reasoning, and Refinement for Enhanced Multimodal Chain-of-Thought (NeurIPS 2025 · similarity 0.91)
• No Labels, No Problem: Training Visual Reasoners with Multimodal Verifiers (ICLR 2026 · similarity 0.91)
• SIFThinker: Spatially-Aware Image Focus for Visual Reasoning (AAAI 2026 · similarity 0.90)
• AdaReasoner: Dynamic Tool Orchestration for Iterative Visual Reasoning (ICLR 2026 · similarity 0.90)
• Multi-step Visual Reasoning with Visual Tokens Scaling and Verification (NeurIPS 2025 · similarity 0.90)
• VisionReasoner: Unified Reasoning-Integrated Visual Perception via Reinforcement Learning (ICLR 2026 · similarity 0.90)