Research map / Vision-language and multimodal / Multimodal LLMs

CORE: Conflict-Oriented Reasoning for General Multimodal Manipulation Detection

ICML 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Interleaved image-text generation with MLLMs · 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 CORE, which builds a 14k-sample Conflict Attribution Corpus and trains MLLMs (Qwen2.5-VL-3B, Gemma3-4B) with modality-bridging pretraining and conflict-perception training to detect multimodal manipulations with few samples or zero-shot.

Problem

Multimodal manipulation/fake news detectors rely on manipulation-specific models and large labeled datasets, generalizing poorly to new manipulation types; MLLMs lack clear conceptual boundaries for detecting conflicts.

Approach

Frame manipulation detection as capturing intrinsic conflicts (cross-modal or with world knowledge); construct a conflict-attributed corpus with conflict factors and sources, then train MLLMs with a conflict-aware contrastive loss and conflict reasoning loss so concept representations separate, enabling rapid adaptation with few samples.

Key results

• With 100-750 training samples, CORE_Qwen and CORE_Gemma outperform the second-best baseline by an average of 9.94% and 10.50% in ACC.
• In zero-shot cross-manipulation detection on MDSM, CORE_Qwen and CORE_Gemma surpass the second-best method by over 14% and 12%.
• Removing the conflict-aware contrastive loss in CPT drops average accuracy by 11.75%, the largest ablation effect.

Surprising findings

• Directly prompting larger MLLMs (Llama-3.2-Vision-90B, Seed-VL-1.6) for conflict reasoning performs worse than CORE trained on only 100 samples.
• Scaling CAC from 14k to 38k slightly degrades performance, attributed to overfitting on SAMM-specific artifacts.

Limitations

• Cross-modal aligner and CAC depend on expert MLLM-generated annotations (GPT-4o, Gemini-2.5-Pro, Qwen3-VL-Plus) with only 1k human-verified samples.
• Possible overfitting to SAMM-specific manipulation artifacts when CAC is scaled up.
• Evaluated only with 3B and 4B backbones on a small set of benchmarks.

Experimental setup

Datasets: SAMM FineHARD DGM^4 MDSM MMFakeBench NewsCLIPpings CAC

Benchmarks: DGM^4 MDSM MMFakeBench NewsCLIPpings SAMM

Metrics: ACC

Baselines: HAMMER HAMMER++ RamDG FKA-Owl AMD Qwen2.5VL-3B Gemma3-4B Qwen3VL-235B

Builds on: SigLIP LoRA HAMMER SNIFFER

Closest papers by meaning

• CrossCheck-Bench: Diagnosing Compositional Failures in Multimodal Conflict Resolution (AAAI 2026 · similarity 0.88)
• Probabilistic Concept Graph Reasoning for Multimodal Misinformation Detection (CVPR 2026 · similarity 0.88)
• Debate with Images: Detecting Deceptive Behaviors in Multimodal Large Language Models (ICML 2026 · similarity 0.87)
• Benchmarking Multimodal Knowledge Conflict for Large Multimodal Models (AAAI 2026 · similarity 0.87)
• The Coherence Trap: When MLLM-Crafted Narratives Exploit Manipulated Visual Contexts (CVPR 2026 · similarity 0.87)
• ForgerySleuth: Empowering Multimodal Large Language Models for Image Manipulation Detection (NeurIPS 2025 · similarity 0.86)