Research map / Vision-language and multimodal / Multimodal LLMs

Ground What You See: Hallucination-Resistant MLLMs via Caption Feedback, Diversity-Aware Sampling, and Conflict Regularization

AAAI 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Reinforcement learning for visual reasoning · 1 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.

Analyzes RL-induced hallucination in MLLMs and proposes caption-based visual-grounded reasoning with a caption reward, reward-variance sample selection, and NTK-similarity InfoNCE regularization, evaluated on Qwen-VL-2.5-7B.

Problem

Hallucinations in multimodal LLMs are amplified by reinforcement learning (GRPO-style) training, stemming from visual misinterpretation in chained reasoning, limited exploration diversity, and destructive conflicts between training samples.

Approach

Adds planning and caption stages before thinking/answer, rewards captions by whether an LLM can answer from the caption alone, keeps only high-reward-variance samples for RL, and regulates NTK similarity between samples with an InfoNCE loss using a threshold rather than minimizing it.

Key results

• Qwen-VL-2.5-7B-Ours reaches 65.6 on MMVU, 50.8 on VideoHallucer and 88.7 on POPE, the best among open-source and proposed models in the table.
• Standard GRPO on Qwen-VL-2.5-7B lowers VideoHallucer to 42.3 from 46.5 for the base model, while the proposed method raises it to 50.8.
• Ablation: removing InfoNCE loss drops MMVU from 65.6 to 63.8 and POPE from 88.7 to 86.8.

Surprising findings

• Standard GRPO training reduces hallucination-benchmark performance (VideoHallucer 46.5 to 42.3), supporting the claim that RL can induce hallucination.

Limitations

• Training data for the RL stage is not clearly described in the text, limiting reproducibility.
• NTK similarity is approximated by cosine similarity of final-layer log-prob gradients rather than computed exactly.
• MathVista performance is moderate relative to the best reasoning baseline (Vision-R1-7B at 73.5).

Experimental setup

Datasets: ChartQA

Benchmarks: MMVU VideoMMMU VideoHallucer MathVista POPE MMBench

Metrics: accuracy

Baselines: GPT-4o Gemini-1.5-Pro LLaVA-OneVision-7B VILA-1.5-8B InternVL2.5-8B Video-R1-7B Vision-R1-7B Qwen-VL-2.5-7B-GRPO

Builds on: GRPO DeepSeek-R1 InfoNCE Neural tangent kernel

Closest papers by meaning

• Understanding the Role of Hallucination in Reinforcement Post-Training of Multimodal Reasoning Models (CVPR 2026 · similarity 0.93)
• Look Carefully: Adaptive Visual Reinforcements in Multimodal Large Language Models for Hallucination Mitigation (ICLR 2026 · similarity 0.92)
• Robust Multimodal Large Language Models Against Modality Conflict (ICML 2025 · similarity 0.92)
• Combating Multimodal LLM Hallucination via Bottom-Up Holistic Reasoning (AAAI 2025 · similarity 0.92)
• Learning to Decode Against Compositional Hallucination in Video Multimodal Large Language Models (ICML 2026 · similarity 0.92)
• Detecting and Preventing Hallucinations in Large Vision Language Models (AAAI 2024 · similarity 0.92)