Research map / Vision-language and multimodal / Multimodal LLMs

Deep Residual Injection for Full-Spectrum Forensic Signal Perception in Multimodal Large Language Models

ICML 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Benchmarks and analysis of 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 Deep-VRM, a Qwen-2.5-VL-7B-based MLLM for AI-generated image detection that freezes early LLM layers and injects LoRA-adapted visual-encoder features as a residual path into layer 16, achieving SOTA across GenImage, AIGI-Now, WildRF, AIGI-Bench and others.

Problem

Generic MLLMs underperform at AI-generated image detection because learning low-level generator artifacts via fine-tuning disrupts pre-trained semantic representations (catastrophic forgetting), and existing MLLM detectors rely on external expert detectors.

Approach

Layer-wise probing shows semantic separability forms in early-to-middle LLM layers, so the method freezes those layers and adds a residual path that injects LoRA-adapted visual-encoder artifact features into an intermediate layer (K=16), letting later trainable layers fuse semantic and forensic signals.

Key results

• Deep-VRM reaches 97.42% average ACC on GenImage and 89.51% on WildRF, 90.48% on AIGI-Bench, and 93.28% average across GenImage, AIGI-Now, AIGI-Bench, and WildRF versus 88.75% for Forensic-Chat.
• Full fine-tuning of Qwen-2.5-VL-7B on D2 collapses general benchmarks (BLINK 0.0373 vs 0.5481 backbone; MME 506 vs 1677), while Deep-VRM keeps them near backbone (MME 1636).
• Ablation shows injection depth K=16 gives best average (92.97/92.91), while shallow K=4 drops WildRF to 74.82.

Surprising findings

• Full fine-tuning of the LLM on artifact data causes catastrophic forgetting, with BLINK dropping from 0.548 to 0.037.
• Fine-tuning all layers without freezing early layers slightly improves GenImage and SynthBuster but hurts in-the-wild WildRF (83.49 vs 89.51).

Limitations

• Performance degrades on Qwen-2.5-VL-3B, attributed to smaller scale and fixed LoRA rank possibly underfitting.
• Weaker on some subsets such as SynthBuster DALLE3-like/ Community Forensics GALIP and EvalGen-style cases, and AIGI-Holmes Janus (77.40).

Experimental setup

Datasets: GenImage MS-COCO ImageNet Unsplash Flux-dev generated images Echo 4o images SD2.1 VAE reconstructions

Benchmarks: GenImage GenImage++ AIGI-Holmes EvalGen CommunityAI SynthBuster AIGI-NOW WildRF AIGI-Bench Community Forensics BLINK RealWorldVQA

Metrics: balanced accuracy fake-class accuracy ACC

Baselines: UnivFD AIDE NPR DRCT OMAT AIGI-Holmes SAFE C2P-CLIP FatFormer CO-SPY PatchShuffle Forensic-Chat

Builds on: Qwen2.5-VL LoRA DDA Forensic-Chat

Closest papers by meaning

• Unlocking the Capabilities of Large Vision-Language Models for Generalizable and Explainable Deepfake Detection (ICML 2025 · similarity 0.86)
• BusterX: MLLM-Powered AI-Generated Video Forgery Detection and Explanation (NeurIPS 2026 · similarity 0.85)
• The Coherence Trap: When MLLM-Crafted Narratives Exploit Manipulated Visual Contexts (CVPR 2026 · similarity 0.85)
• Rethinking Vision-Language Model in Face Forensics: Multi-Modal Interpretable Forged Face Detector (CVPR 2025 · similarity 0.85)
• Forensics-Bench: A Comprehensive Forgery Detection Benchmark Suite for Large Vision Language Models (CVPR 2025 · similarity 0.85)
• Identity-Aware Vision-Language Model for Explainable Face Forgery Detection (AAAI 2026 · similarity 0.85)