OmniVL-Guard: Towards Unified Vision-Language Forgery Detection and Grounding via Balanced RL

AI summary

Overview

Research area: Computer Vision / Multimodal Large Language Models, specifically unified forgery detection and tampering localization across text, images, and videos, trained with reinforcement learning.

Technical level: Advanced. The paper combines multimodal LLM training, Chain-of-Thought data synthesis, and a mathematical decomposition of policy-gradient dynamics in reinforcement learning, so readers will benefit from prior familiarity with RL fine-tuning and vision-language models.

Scope (1 sentence): The paper proposes OmniVL-Guard, a reinforcement-learning framework with a self-evolving Chain-of-Thought data pipeline and a balanced multi-task reward optimization method (ARSPO), evaluated on a new 73k-sample SFT plus 110k-sample RL corpus (FSFR) and a roughly 700k-sample in-domain test pool.

What This Paper Is About

Existing forgery detectors typically handle only one or two modalities (e.g., images alone, or text alone, or image-text pairs), while real misinformation mixes text, images, and videos together. The authors build a single model that both judges whether content is forged and localizes where it is forged (image regions, text spans, or video time segments), and they identify a problem they call "difficulty bias": the easy binary classification task dominates the training gradient, so the harder grounding tasks improve far less. The paper's goal is to fix that imbalance during reinforcement-learning training so all tasks improve together.

Key Contributions

1. A unified vision-language forgery framework. OmniVL-Guard processes images, videos, and text in one paradigm for both forgery detection and manipulation grounding, and the authors state they pioneer the integration of Reinforcement Learning into this problem, to improve forensic reasoning and cross-modal generalization.

2. The FSFR dataset (Full-Spectrum Forensic Reasoning). A multimodal corpus supporting the full SFT-RL-Test pipeline, built with a Self-Evolving CoT Generation pipeline that synthesizes high-quality CoT annotations. FSFR_sft contains 73k samples (with hard samples, D_s^h, accounting for approximately 15%), and FSFR_rl comprises 110k samples. FSFR_sft = D_s^h ∪ D_s^3 and FSFR_rl = D_r \ D̄_r.

3. ARSPO (described in the abstract as "Adaptive Reward Scaling Policy Optimization" and in the body as "Adaptive Reward Shaping Policy Optimization"). An adaptive reward-shaping mechanism with task-specific reward mapping functions and a dynamic coefficient adjustment algorithm, which the authors say resolves exploration imbalance in multi-task RL training and substantially boosts performance on high-difficulty tasks.

4. Self-Evolving CoT Generation. A four-stage strategy — source data collection, forensic reasoning seed priming, seed bootstrapping through self-evolution, and collaborative hard-CoT synthesis — that avoids the "Efficiency-Bias Dilemma" of low-quality vanilla generation and the "Hindsight Bias" of forcing ground truth into reasoning paths.

Main Findings

• General-purpose MLLMs are weak at localization. In Table 1(a), GPT5 scores 73.71 on binary classification but 19.01 (image loc.), 34.96 (text loc.), and 24.60 (video loc.); Gemini3 scores 74.54, 21.32, 32.77, 24.11; Seed1.6 scores 72.99, 20.97, 31.86, 26.29. The authors use this to argue these models cannot directly generate good CoT data.

• Conventional RL over-optimizes the easy task. In Table 1(b), starting from an SFT baseline of 53.38 / 51.08 / 44.67 / 33.08, GRPO raises binary classification to 89.42 (+36.0) while image localization drops to 50.96 (-0.1); SAPO reaches 90.75 (+37.4) with image localization at 51.24 (+0.2). ARSPO reaches 90.85 (+37.5), 54.26 (+3.2), 63.78 (+19.1), and 59.22 (+26.1).

• Strong in-domain results. In Table 2, OmniVL-Guard reaches 96.20 / 93.12 / 98.58 / 75.52 on binary classification (Text, Image, Video, Text-Image) and 54.26 (Image Loc. IoU), 63.78 (Text Loc. F1), 59.22 (Video Loc. tIoU). Improvements over the best comparison are +6.97, +2.73, -0.23, +3.44, +5.73, +22.92, and +37.79 respectively — the authors highlight +37.79 on video temporal localization and +22.92 on text localization.

• Robust zero-shot out-of-domain performance. In Table 3, OmniVL-Guard scores 93.69 on ISOT (text), 63.64 on CASIA2.0 (image), 79.38 on MMFakeBench (image-text), and 63.55 on FakeSV (video-text). The authors report it surpasses domain-specific models such as FKA-Owl (62.32 on MMFakeBench), which they attribute to overfitting in those models.

• Every ARSPO component helps. In Table 4, SFT alone gives 83.73 / 32.41 / 33.82 / 63.57 / 51.08 / 44.67 / 33.08. Adding SAPO yields +24.33 average improvement, adding Task-Based Reward Mapping Function (TBRMF) yields +26.42, adding Dynamic Coefficient Adjustment (DCA) yields +26.93, and the complete ARSPO yields +28.33, with Video Loc. rising to 59.22%.

• Reward mapping steepness has an optimum. Figures 4(a)–4(b) show performance peaks at a = 3; beyond that, the authors report performance declines because overly steep mappings amplify noise ("reward overfitting").

• ARSPO also helps in single-task settings. Figure 4(c) reports that exponential mapping reaches 0.56 (+4%) for image localization and 0.66 (+8%) for text localization without any inter-task competition, which the authors say shows ARSPO's benefit is "reshaping" gradient signals, not just balancing tasks.

• Three self-evolution iterations are enough. Table 5 compares stopping at D_s^3 ∪ D_s^h with extending to D_s^4 ∪ D_s^h: differences are small (Image Localization 54.26 to 54.41; Video Localization changes by -0.15), so the authors terminate the loop at k = 3.

• Theoretical explanation. Equation 4 decomposes the gradient change rate into a task-difficulty sensitivity term H_k' and a reward-function sensitivity term g_k', which the authors use to explain a "Matthew Effect" in multi-task RL where simple tasks hijack the update direction.

Methodology in Plain English

The authors first gather source data from public datasets: FakeNewsCorpus and MCFEND for text; FakeClue, LOKI, and ForgeryNet for images; GenVideo, DVF, and ForgeryNet for video; and SAMM, MDSM, DGM4, and NewsCLIPpings for image-text. Video-text data is deliberately withheld from training and reserved for out-of-domain zero-shot evaluation. All data is split into D_s (for SFT data generation), D_r (for RL), and D_t (for in-domain testing).

To build trustworthy reasoning data, they first have state-of-the-art MLLMs (Seed1.6-VL, Gemini3, ChatGPT5) reason over a small fraction of D_s, then filter and verify those reasoning paths against ground truth and by a separate MLLM, producing a seed set D_s^0 of 6.7k samples. They SFT Qwen3VL-8B on this seed, run RL on a subset D̄_r, and obtain a warm-up policy π0. Then they loop: the previous round's model generates reasoning for the remaining D_s data, the outputs are filtered and validated, and the good samples are merged into the seed to form D_s^k. Importantly, each round re-initializes from the base Qwen3VL-8B rather than the previous model, to avoid accumulating distributional bias. The loop stops after three iterations.

For long-tail samples the model still gets wrong, they use a multi-agent workflow where one MLLM writes a CoT using the known answer, a second rewrites it to read like a natural deduction made without knowing the answer, and a third evaluates it — producing D_s^h.

For the RL stage, they mathematically analyze how task difficulty affects gradients and conclude that a convex reward mapping amplifies gradient contribution from superior responses. So they assign identity mapping g_k(x) = x to binary classification and exponential mapping g_k(x) = e^{a_k x} to the three localization tasks. On top of that, they run a dynamic coefficient adjustment routine (Algorithm 1) that monitors each task's progress every T steps, identifies the "laggard" task with the smallest relative gain, boosts its weight, decays weights for mastered tasks (clamped to at least 1), rescues tasks whose performance is regressing, suspends updates for tasks in a rapid ascent phase, and finally rescales all coefficients by dividing by the minimum coefficient among tasks.

Why This Matters

Impact on research. The paper reframes multimodal forgery detection as a multi-task optimization problem rather than a collection of separate uni-modal or bi-modal detectors, and it provides a gradient-level analysis of why easy tasks dominate harder ones in multi-task RL. That analysis (the "difficulty bias" / "Matthew Effect" framing) may transfer to other multi-task RL settings beyond forensics. The FSFR dataset and the released code give the community a shared benchmark and training pipeline.

Real-world applications:

• Social-media content moderation, where posts interleave text, images, and videos and must be judged and annotated with the manipulated region or segment.
• Newsroom fact-checking and verification, checking both whether an article's text is fabricated and whether accompanying imagery has been altered.
• Image and video forensics for legal, insurance, or journalistic evidence review, where locating the tampered region matters as much as flagging it.
• Platform-level provenance and trust-and-safety pipelines that need one model instead of separate per-modality detectors.

Industry relevance. Organizations handling user-generated multimodal content currently maintain fragmented detector stacks; a single model that handles image, video, text, and image-text inputs — and generalizes zero-shot to unseen forgery types, as shown by the ISOT, CASIA2.0, MMFakeBench, and FakeSV results — is attractive for deployment cost and coverage. The released dataset and code lower the barrier to reproducing and extending this line of work.

Future Directions

• Extending beyond the three primary modalities. The current framework covers text, images, videos, and image-text; audio and richer audio-visual combinations are not addressed, and video-text is used only for zero-shot evaluation rather than training.

• Understanding the collaborative hard-CoT synthesis. The provided paper content is truncated in the middle of the ablation discussion of Collaborative Hard-CoT Synthesis, so its isolated contribution is not reported here; a fuller analysis of how multi-agent CoT rewriting affects downstream performance remains an open question.

• Generalizing the ARSPO balance mechanism. The dynamic coefficient adjustment requires hyper-parameters (ε_mom, ε_rescue, τ_high, α_decay, α_boost) and periodic monitoring every T steps; how sensitive results are to these settings, and whether the same mechanism helps other multi-task RL domains, is not established.

• Scaling and cost. The self-evolution loop uses multiple large MLLMs for validation and a 3-iteration termination chosen for compute/performance trade-off; whether similar quality can be achieved with cheaper annotation or fewer rounds for new modality domains is unexplored.

Target Audience

This paper is aimed at researchers and practitioners working on multimodal misinformation detection, deepfake and tampering localization, and reinforcement-learning-based fine-tuning of multimodal LLMs. It will also interest engineers building trust-and-safety or content-verification systems who need a single model covering text, image, video, and image-text inputs, and dataset builders interested in self-evolving Chain-of-Thought generation pipelines.

Authors’ abstract

Existing forgery detection methods are often limited to uni-modal or bi-modal settings, failing to handle the interleaved text, images, and videos prevalent in real-world misinformation. To bridge this gap, this paper targets to develop a unified framework for omnibus vision-language forgery detection and grounding. In this unified setting, the {interplay} between diverse modalities and the dual requirements of simultaneous detection and localization pose a critical ``difficulty bias`` problem: the simpler veracity classification task tends to dominate the gradients, leading to suboptimal performance in fine-grained grounding during multi-task optimization. To address this challenge, we propose \textbf{OmniVL-Guard}, a balanced reinforcement learning framework for omnibus vision-language forgery detection and grounding. Particularly, OmniVL-Guard comprises two core designs: Self-Evolving CoT Generatio and Adaptive Reward Scaling Policy Optimization (ARSPO). {Self-Evolving CoT Generation} synthesizes high-quality reasoning paths, effectively overcoming the cold-start challenge. Building upon this, {Adaptive Reward Scaling Policy Optimization (ARSPO)} dynamically modulates reward scales and task weights, ensuring a balanced joint optimization. Extensive experiments demonstrate that OmniVL-Guard significantly outperforms state-of-the-art methods and exhibits zero-shot robust generalization across out-of-domain scenarios. The dataset and code are publicly available at https://github.com/shen8424/OmniVL-Guard.