Research map / Large language models / Language, safety and interpretability
Asymmetric Anchoring: Opening the Black Box of MLLMs for Forgery Detection
Zhiqiang Yang, Renshuai Tao, Chunjie Zhang, Zhaoxiang Liu, Xiaolong Zheng, Yao Zhao
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add anchoring representation loss L_arl on real images (ablation) | LLaVA-v1.5-7B + CLIP · SID-Set detection | accuracy: 91.78% -> 96.60% | Improved |
| add L_arl with SigLIPv2 encoder (ablation) | Vicuna-1.5-7B + SigLIPv2 · SID-Set detection | accuracy: 93.21% -> 97.04% | Improved |
| add L_arl (ablation) | LISA-7B-v1 · OpenSDID SD1.5 (in-domain) | accuracy: 91.17% -> 90.25% | Hurt |
| add L_arl (ablation) | LISA-7B-v1 · OpenSDID SDXL | accuracy: 77.65% -> 85.00% | Improved |
| add L_arl to a 13B LLM (ablation) | Vicuna-1.5-13B + CLIP · SID-Set detection | accuracy: 94.13% -> 97.24% | Improved |
| apply L_arl asymmetrically (real only) instead of to real and tampered (ablation) | AAP-lisa · SID-Set | localization IoU: 40.2 -> 56.8 | Improved |
| apply L_arl asymmetrically (real only) instead of symmetric (ablation) | AAP-lisa · SID-Set | detection accuracy: 94.8 -> 97.1 | Improved |
| inject alignment error map into the localization decoder (ablation) | AAP-lisa · Localization (avg of OpenSDID and SID-Set) | IoU: 41.52 -> 50.25 | Improved |
| inject error map on SID-Set only (ablation) | AAP-lisa · SID-Set | localization IoU: 43.80 -> 56.85 | Improved |
| add anchoring (AAP) to reduce representation drift | MLLM fine-tuning · Real vs Tampered features | inter-class distance: 5.4550 -> 13.2344 | Improved |
| use AAP versus SIDA for real image recognition | AAP-lisa · SID-Set (Real) | accuracy: 89.1 -> 97.8 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- LISA-7B-v1 — LISA: Reasoning Segmentation via Large Language Model (CVPR 2024) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Forensics-Bench: A Comprehensive Forgery Detection Benchmark Suite for Large Vision Language Models (CVPR 2025 · similarity 0.89)
- FakeShield: Explainable Image Forgery Detection and Localization via Multi-modal Large Language Models (ICLR 2025 · similarity 0.88)
- OmniVL-Guard: Towards Unified Vision-Language Forgery Detection and Grounding via Balanced RL (ICML 2026 · similarity 0.88)
- Toward Real-world Text Image Forgery Localization: Structured and Interpretable Data Synthesis (NeurIPS 2025 · similarity 0.88)
- TextShield-R1: Reinforced Reasoning for Tampered Text Detection (AAAI 2026 · similarity 0.87)
- Deep Residual Injection for Full-Spectrum Forensic Signal Perception in Multimodal Large Language Models (ICML 2026 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
