Research map / Vision-language and multimodal / Multimodal LLMs

ChainMPQ: Interleaved Text-Image Reasoning Chains for Mitigating Relation Hallucinations

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.

Proposes ChainMPQ, a training-free method that decomposes relation questions into subject/object/relation sub-questions and chains answers and attention-based visual masks to reduce relation hallucinations in LVLMs.

Problem

Relation hallucinations in large vision-language models, where the model recognizes subject and object but misjudges the relation linking them, are underexplored relative to object and attribute hallucinations.

Approach

Decompose a relational yes/no question into five complementary sub-questions (locate subject, locate object, and three masked-component queries), answer them sequentially, and carry forward textual answers and top-k attention-derived visual masks as a bias on subsequent attention, forming an interleaved text-image reasoning chain.

Key results

• On MMRel with LLaVA-1.5-7B, ChainMPQ reaches 65.20% accuracy vs 59.02% vanilla and 63.50% for the best baseline (Calibrate).
• On R-Bench with Qwen2.5-VL-7B, accuracy rises from 79.85% vanilla to 83.92%.
• Light1 variant gives the best accuracy-latency trade-off (63.78% on MMRel at 1.5 s/sample vs 3.3 s for full method).

Surprising findings

• Full ChainMPQ roughly triples per-sample latency (0.9 s to 3.3 s) relative to vanilla decoding on LLaVA-1.5-7B.

Limitations

• Relies on attention distributions as a proxy for visual evidence, which may not capture the true reasoning process.
• Spatial-category performance is lower, attributed to misalignment between visual token granularity and object boundaries.
• Adds substantial inference latency from multiple sequential decoding steps.

Experimental setup

Datasets: MMRel R-Bench

Benchmarks: MMRel R-Bench MMBench MME

Metrics: accuracy precision F1 latency

Baselines: Vanilla Constraint-Aware Prompting Detect-then-Calibrate Chain-of-Thought prompting

Builds on: Interleaved Modal Chain-of-Thought Chain-of-Thought prompting Detect-then-Calibrate Constraint-Aware Prompting

Closest papers by meaning

• Evaluating and Analyzing Relationship Hallucinations in Large Vision-Language Models (ICML 2024 · similarity 0.92)
• MIRAGE: Assessing Hallucination in Multimodal Reasoning Chains of MLLM (NeurIPS 2025 · similarity 0.89)
• When Semantics Mislead Vision: Mitigating Large Multimodal Models Hallucinations in Scene Text Spotting and Understanding (NeurIPS 2025 · similarity 0.89)
• Grounding Language with Vision: A Conditional Mutual Information Calibrated Decoding Strategy for Reducing Hallucinations in LVLMs (NeurIPS 2025 · similarity 0.89)
• Visual Description Grounding Reduces Hallucinations and Boosts Reasoning in LVLMs (ICLR 2025 · similarity 0.89)
• Combating Multimodal LLM Hallucination via Bottom-Up Holistic Reasoning (AAAI 2025 · similarity 0.89)