Research map / Vision-language and multimodal / Multimodal LLMs

PRPO: Paragraph-level Policy Optimization for Vision-Language Deepfake Detection

ICML 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Multimodal LLM benchmarks · 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.

Introduces DF-R5, a ~115k-image deepfake reasoning dataset distilled from Gemini-2.5, a DX-LLaVA detector (LLaVA with CLIP ConvNeXT encoder), and PRPO, a label-free paragraph-level test-time RL method with visual-consistency and prediction-consistency rewards.

Problem

Multimodal LLMs perform poorly at deepfake detection and produce explanations that are hallucinated or misaligned with visual evidence, and existing datasets lack detailed reasoning annotations and generalize poorly to unseen generative domains.

Approach

Paragraph-level advantages in GRPO-style RL: each reasoning paragraph is rewarded for CLIP ConvNeXT image-text alignment of its YAKE keywords and the final answer is rewarded for agreeing with the majority of paragraph-level labels, with no ground-truth labels needed at test time.

Key results

• PRPO reaches 89.91% average F1 across five unseen domains versus 78.08% for DX-LLaVA and 75.26% (SIDA) reported in the main table.
• PRPO achieves an overall reasoning score of 4.55/5.0 by GPT-4o, above Gemini-2.5 at 4.20.
• On the DDIM transfer task, PRPO improves F1 from 92.34 (DX-LLaVA) to 95.88 and outperforms TTRL-style PPO and GRPO (93.95 and 92.09).

Surprising findings

• Unsupervised (label-free) PRPO outperformed supervised PRPO using ground-truth rewards on DDIM (95.88 vs 78.75 F1), which the authors attribute to overfitting of supervised rewards.
• Gemini-2.5 is the strongest MLLM tested on classification (77.60% accuracy on a 1,000-image sample), but the detector trained on its distilled labels still lags it in accuracy by about 2% before PRPO.

Limitations

• Focuses on face-centric deepfakes and does not address natural-scene deepfakes.
• Reasoning annotations are distilled from a single MLLM (Gemini), which may limit robustness and diversity.
• Inconsistent results: the average F1 for SIDA differs between the main table (75.26) and the appendix (75.56), and the PRPO-S vs PRPO-U comparison is on only two domains.

Experimental setup

Datasets: DF40 FaceForensics++ CelebDF DF-R5

Benchmarks: DF-R5 leave-one-domain-out (DDIM, PixArt, SD, SiT, StyleGAN3)

Metrics: accuracy precision recall F1 reasoning score (CAC, EGIA, RQ, CC, CU on 0-5 scale) n-gram overlap unique visual features

Baselines: LLaVA DE-FAKE FakeShield UnivCLIP SIDA LLaMA-4 Maverick Pixtral-12B Qwen2.5-VL-32B GPT-4o Gemma-3-27B-IT Gemini-2.5 PPO (TTRL)

Builds on: GRPO LLaVA TTRL DF40 YAKE

Closest papers by meaning

• Pixels Don't Lie (But Your Detector Might): Bootstrapping MLLM-as-a-Judge for Trustworthy Deepfake Detection and Reasoning Supervision (CVPR 2026 · similarity 0.89)
• CFPO: Counterfactual Policy Optimization for Multimodal Reasoning (ICML 2026 · similarity 0.88)
• Standing on the Shoulders of Giants: Reprogramming Visual-Language Model for General Deepfake Detection (AAAI 2025 · similarity 0.88)
• Veritas: Generalizable Deepfake Detection via Pattern-Aware Reasoning (ICLR 2026 · similarity 0.88)
• Dr. Seg: Revisiting GRPO Training for Visual Large Language Models through Perception-Oriented Design (CVPR 2026 · similarity 0.87)
• Unlocking the Capabilities of Large Vision-Language Models for Generalizable and Explainable Deepfake Detection (ICML 2025 · similarity 0.87)