Grounded Chain-of-Thought for Multimodal Large Language Models

1. Introduction

• The paper frames visual hallucination as a reliability problem: an MLLM may answer correctly by relying on data-distribution bias rather than relevant image information.Such answers may not reflect what the model actually sees, despite not affecting performance on most vision-language benchmarks.Source: §1 · p. 1
• GCoT trains MLLMs to reason step by step about relevant visual cues and ground key steps, producing coordinates as an intuitive basis for the answer.Unlike visual CoT focused more on knowledge reasoning, GCoT targets both visual and spatial perception.Source: §1 · p. 2
• MM-GCoT provides 23,028 training examples across Attribute, Judgment, and Object tasks, plus about 994 evaluation examples assessed with answer accuracy, grounding accuracy, and answer-grounding consistency.Examples contain multiple reasoning steps with grounding information, including spatial coordinates.Source: §1 · p. 2, Figure 3 · p. 3
• Most evaluated MLLMs show low answer-grounding consistency, and training with GCoT improves visual-spatial reasoning while reducing answer–grounding inconsistency.The findings also report that consistency does not track grounding or reasoning capability, and that larger models can perform poorly on MM-GCoT despite stronger common-benchmark QA accuracy.Source: §1 · p. 2
• The paper presents MM-GCoT as a dataset for GCoT research and a benchmark for evaluating MLLM visual hallucination.It also reports that experiments reveal notable findings about existing MLLMs and validate the merits of GCoT capability.Source: §1 · p. 3

2. Related Work

• Existing hallucination benchmarks often rely on binary verification, which the authors say makes reasoning reliability hard to judge and motivates more granular examination of language prior.The paper discusses POPE and HallusionBench as examples of existing benchmarks.Source: §2 · p. 3
• Visual CoT approaches focus on explanations or knowledge reasoning, but do not verify whether each reasoning step is correct visual evidence.ScienceQA is described as grounding scientific reasoning in pretrained memory, while LLaVA-CoT outputs a detailed solution before its answer.Source: §2 · p. 3
• Unlike phrase grounding, which locates caption-mentioned objects, GCoT analyzes a question into grounded reasoning steps before answering; unlike GroundedQA, it includes annotated reasoning steps.The paper also contrasts MM-GCoT with VoCoT in dataset construction and evaluation metrics for visual hallucination.Source: §2 · pp. 3–4
• Grounded spatial reasoning in chain-of-thought is presented as relevant to vision-language-action models for embodied navigation and manipulation.The paper links this integration to more accurate and interpretable decision-making in complex embodied tasks.Source: §2 · p. 4

3. Grounded Chain-of-Thought

• GCoT decomposes an image-based question into steps, grounds task-related visual elements with spatial information, then uses those steps to predict an answer with coordinates.The coordinates serve as an intuitive basis for the answer.Source: §3 · p. 4
• The formulation replaces direct image-text-to-answer mapping with a multi-step decision process conditioned on prior visual evidence and textual reasoning states.The prose identifies I and T as the input image and text, and G_t and R_t as visual evidence and textual reasoning state at step t.Source: §3 · p. 4
• Figure 4 contrasts answer-first prompting, which requests text before a box, with grounding-first prompting, which requests the box before text.These are two prompting settings for evaluating LLaVA on MM-GCoT.Source: Figure 4 · p. 4
• The paper uses SFT, while noting that GCoT could also be cultivated with reinforcement-learning schemes and may inform vision-language-action research.The cited example of an RL scheme is GRPO; the text also identifies embodied scenarios as a relevant setting.Source: §3 · p. 4

Dataset Statistics

• MM-GCoT contains 23,028 training/validation samples and 994 hallucination-evaluation test samples across Attribute, Judgment, and Object tasks.Attribute asks about a specified object's attributes, Judgment checks a description, and Object identifies the category at a specified location.Source: §4 · p. 4

Dataset Construction

• A four-stage pipeline aligns Visual Genome region descriptions with object annotations, builds spatial-relation graphs, samples reasoning paths, and converts structured templates into natural-language questions.LLM verification checks training answers and grounding consistency, while test samples receive manual validation.Source: §4 · pp. 4–5

Evaluation Metrics

• Evaluation separates answer accuracy, grounding accuracy, and answer-grounding consistency; grounding accuracy uses Acc@0.5, requiring IoU > 0.5.Answer accuracy uses text matching against ground-truth answers.Source: §4 · p. 5

Evaluation Prompt Settings

• Answer-first prompting asks models to answer before giving coordinates, whereas grounding-first asks them to locate a potential answer before answering.Grounding-first is more challenging and tests whether answers are based on identified visual elements; answer-first tests whether the model sees the answer cue.Source: §4 · pp. 5–6

5.1. Implement Details

• The evaluation covers 12 MLLMs across LLaVA, LLaVA-OneVision, Qwen2.5-VL, and InternVL2.5; LLaVA GCoT is trained from scratch after replacing instruction-tuning samples with MM-GCoT examples.The replacement uses an equivalent number of examples and follows LLaVA's default settings.Source: §5.1 · p. 6, §5.2 · p. 6

5.2. Quantitative Analysis

• LLaVA-OneVision-72B reaches 75.7% average accuracy but only 11.1% answer-grounding consistency, and InternVL2.5-78B averages just 35.4% consistency.The reported results indicate that visual hallucination does not directly correlate with model scale.Source: §5.2 · p. 6
• Task performance varies: Attribute has strong grounding, Judgment has the highest answer accuracy, and Object has the strongest answer-grounding consistency.Examples include 82.6% grounding accuracy for Qwen2.5-VL-7B on Attribute under grounding-first, 86.1% average Judgment answer accuracy under answer-first, and 39.4% Object consistency for InternVL2.5-8B.Source: §5.2 · p. 6
• Grounding-first prompting can improve grounding while reducing answer accuracy: Qwen2.5-VL-7B gains 7.9% in Acc@0.5 but loses 21.2% in average answer accuracy.LLaVA-7B's answer accuracy is 8.9% higher with answer-first prompting, and Qwen2.5-VL-3B has a 45.4% accuracy gap between prompting strategies.Source: §5.2 · p. 6
• Qwen2.5-VL-7B exceeds its 72B counterpart in answer-grounding consistency by 18.2%, despite the larger model's stronger grounding performance on Attribute.InternVL2.5-38B improves answer accuracy over its 8B counterpart by 5.5% but has 0.3% lower Object-task grounding performance.Source: §5.2 · pp. 6–7
• LLaVA-7B GCoT gains 4.5% in accuracy and 55.7% in consistency over original LLaVA-7B, while LLaVA-13B GCoT raises consistency by 61.2% with a 2.2% grounding increase.LLaVA-13B GCoT performs comparably or better than Qwen2.5-VL-7B-Instruct across multiple metrics.Source: §5.2 · p. 7, Table 5 · p. 7

5.3. Qualitative Analysis

• Qualitative comparisons show Qwen2.5-VL-72B and InternVL2.5 can give correct answers while grounding them in irrelevant or mismatched regions; LLaVA GCoT uses stepwise visual reasoning to identify targets.The Figure 5 visualizations compare model scales under answer-first and grounding-first prompts, and Figure 6 includes cases with similar or undescribed objects.Source: §5.3 · pp. 7–8, Figure 5 · p. 7, Figure 6 · p. 8
• LLaVA-1.5 13B GCoT generalizes to open-world QA and referring expression comprehension, providing visual evidence and reasoning on tasks outside its GCoT training datasets.Figure 7 shows examples on common VQA and visual grounding tasks.Source: §5.3 · pp. 8–9, Figure 7 · p. 8

• GCoT trains MLLMs to use discrete visual reasoning steps with grounded information to address visual hallucination.Source: §6 · p. 9
• MM-GCoT supports training GCoT capability and evaluating visual hallucination in existing experiments.Source: §6 · p. 9
• Experiments report notable findings about hallucination in existing MLLMs and benefits of GCoT for trustworthy visual reasoning.Source: §6 · p. 9