Research map / Vision-language and multimodal / Multimodal LLMs

From Failure to Feedback: Group Revision Unlocks Hard Cases in Object-Level Grounding

CVPR 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Reinforcement learning for visual 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.

Proposes group-revision GRPO for object-level grounding: revises a sampled failed response with a group of follow-up candidates and converts the improvement (Hungarian-matched alignment cost) into a reward shaping term and advantage post-scaling.

Problem

GRPO with sparse response-level rewards gives little learning signal on hard grounding cases where all sampled responses fail the IoU/point criterion, leaving hard object-level grounding examples unlearned.

Approach

Sample one initial response, then condition a group of revised responses on it; compute a Hungarian-matched alignment cost for initial and revised outputs, and use the relative improvement as a shaping term added to the reward and as a multiplicative scale on the GRPO advantage.

Key results

• Ours (multi-object) reaches 66.73 gIoU on ReasonSeg test and 79.97% average counting accuracy, versus VisionReasoner's 75.7% average.
• Ours (multi-object) achieves 85.80% average Acc@0.5 on REC, surpassing VisionReasoner's 84.8%.
• Ablation: adding revision raises ReasonSeg val from 62.54 to 64.97 and consolidation further raises it to 66.99 in the single-object setting.

Surprising findings

• Initial responses are not degraded to inflate the shaping signal; the group-mean improvement is concentrated near 0-0.2, suggesting limited reward hacking.

Limitations

• Revision group requires extra rollouts per hard sample, increasing sampling cost (not quantified in the text).
• Single-object setting shows more segmentation mismatch hallucinations than the multi-object setting.
• Relies on Hungarian matching against ground-truth boxes and points, so it needs annotated objects during training.

Experimental setup

Datasets: RefCOCOg VisionReasoner7K ReasonSeg RefCOCO RefCOCO+ PixMoCount CountBench

Benchmarks: ReasonSeg RefCOCO RefCOCO+ RefCOCOg ReasonG PixMoCount CountBench SeedBenchV2+ ChartQA OCRBench POPE MMMU-Pro

Metrics: gIoU cIoU Acc@0.5 counting accuracy VQA accuracy

Baselines: LISA PixelLM PerceptionGPT SEGLLM Read Qwen2-VL-7B Qwen2.5-VL-7B Seg-R1 Seg-Zero VisionReasoner Groma VistaLLM

Builds on: GRPO Seg-Zero VisionReasoner potential-based reward shaping SAM2

Closest papers by meaning

• Can Large Vision-Language Models Correct Semantic Grounding Errors By Themselves? (CVPR 2025 · similarity 0.87)
• Connecting the Dots: Training-Free Visual Grounding via Agentic Reasoning (AAAI 2026 · similarity 0.86)
• First Logit Boosting: Visual Grounding Method to Mitigate Object Hallucination in Large Vision-Language Models (CVPR 2026 · similarity 0.86)
• GUI-G1: Understanding R1-Zero-Like Training for Visual Grounding in GUI Agents (NeurIPS 2025 · similarity 0.86)
• Dr. Seg: Revisiting GRPO Training for Visual Large Language Models through Perception-Oriented Design (CVPR 2026 · similarity 0.86)
• GeM-VG: Towards Generalized Multi-image Visual Grounding with Multimodal Large Language Models (AAAI 2026 · similarity 0.85)