Source-linked AI summary
MM-VeriAgent: Learning to Use Extensive Tools to Verify Multimodal Misinformation with Reinforcement Learning
Peipei Li, Shuhan Xia, Shengyang Liu, Zekun Li, Ran He
TL;DR
Mixed-source multimodal misinformation calls for sample-specific verification, but existing tool-using agents either follow fixed workflows or plan and search at inference time. MM-VeriAgent combines an extensive verification toolkit with a reinforcement-learned tool-use policy and cached tool outputs for training. On MMFakeBench, it raises multi-class accuracy from 34.3% to 70.4% and achieves 85.1% overall binary F1 without test-time search.
Problem
Mixed-source misinformation requires different verification mechanisms, while existing agents use limited tools and either fixed workflows or inference-time planning that constrains adaptability or adds inference cost.
Source: §2 · pp. 1, 3Method
MM-VeriAgent combines MM-VeriTools, a unified toolkit of selected textual, visual, and cross-modal tools, with reinforcement learning that trains an LVLM to choose tools and when to stop.
Source: §2 · p. 1Results
70.4% multi-class accuracy, up from 34.3%, and 85.1% overall binary F1; MM-VeriAgent outperforms MMD-Agent and SCPO on both tasks.
Source: §2 · pp. 2, 7Takeaways & Limitations
On MMFakeBench, the learned policy operates without test-time search, and Tool-Execution Cache reduces online tool executions during RL training.
Source: §2 · pp. 2, 7Evaluation is confined to MMFakeBench, so generalization to unseen forgery generators, news domains, and distortion types beyond its three source categories remains to be established.
Source: §2 · p. 7
Abstract
from arXiv · showhide
Real-world multimodal misinformation often involves mixed forgery sources, requiring sample-specific detection strategies. Existing tool-augmented methods rely on predefined workflows or inference-time planning, limiting adaptability or increasing inference cost. To address this issue, we introduce \textbf{MM-VeriAgent}, which learns to verify mixed-source multimodal misinformation with tools. We first build \textbf{MM-VeriTools}, a specialized toolkit for misinformation detection agents. By benchmarking various candidate models and methods on the sub-tasks required by mixed-source detection, we select the strongest for textual, visual, and cross-modal forgery analysis and encapsulate them as callable tools with a unified interface. On top of this toolkit, we train the LVLM agent with reinforcement learning to teach it how to use these tools to better solve mixed-source detection. Since many of the tools are specialized models whose online execution at every rollout severely limits RL efficiency, we further introduce \textbf{Tool-Execution Cache}, which pre-executes candidate tool calls and reuses their cached outputs during training. This preserves multi-step rollouts while reducing online tool execution, largely improving the training efficiency.Experiments on MMFakeBench demonstrate substantial accuracy gains over the base model without explicit tool search at inference time. Ablation and efficiency analyses further validate the learned tool-use policy and show that Tool-Execution Cache reduces online tool executions during training.
2 Minzu University of China
This section presents MM-VeriAgent’s learned tool-use approach for mixed-source misinformation verification and evaluates it on MMFakeBench without explicit test-time tool search. It also examines how tool selection, interaction-based training, and evidence-grounded reward affect performance.
Introduction
- MM-VeriAgent targets mixed-source posts because fixed tool workflows cannot adapt to sample-specific forgeries, while test-time planning adds inference computation.Textual, visual, and cross-modal distortions call for different verification mechanisms, and their sources are unknown at test time.Source: §2 · p. 1
Task Definition
- The task predicts both binary veracity and a four-way primary source label, with supporting evidence drawn from collected tool observations.The source classes are Real, Textual Veracity Distortion, Visual Veracity Distortion, and Cross-modal Consistency Distortion; distortions may co-occur, but each sample has a unique primary label.Source: §2 · p. 3
MM-VeriTools: Extensive Toolkit for Verification
- MM-VeriTools benchmarks and prunes candidate tools, retaining nine tools—four textual, four visual, and one cross-modal—while the learned policy chooses per-sample calls.Candidates are retained only if they outperform the base model on balanced, subtask-specific benchmarks disjoint from the test set.Source: §2 · pp. 3–4
Reinforcement Learning for Verification
- The agent is trained with SFT followed by GRPO, rewarding veracity, source attribution, evidence overlap, and valid output, with greater weight on source and evidence.Evidence overlap is rewarded only when source attribution is correct.Source: §2 · p. 4
Tool-Execution Cache
- At five teacher rollouts per sample, cache hit rate reaches 100%, cutting rollout time from 40.3 to 6.5 seconds per sample, a 6.2× speedup.At that hit rate, training requires no online tool execution.Source: Figure 3 · p. 5, §2 · p. 5
- Tool-Execution Cache uses teacher rollouts before training to record tool outputs, then matches policy calls by sample and tool to preserve multi-turn interaction.The substitution is faithful because tools are stateless and take fixed arguments; cache misses fall back to online execution.Source: §2 · p. 5
Main Results
- Multi-class accuracy rises from 34.3% to 70.4% when MM-VeriAgent augments Qwen2.5-VL-7B, without explicit tool search at inference time.The paper reports the result on MMFakeBench and says it is on par with substantially larger proprietary models.Source: §2 · p. 2, Table 3 · p. 6
- Even the strongest proprietary models remain around 70% on multi-class source attribution: Gemini-2.5-Pro scores 69.8% and GPT-5.5 scores 71.8%.The paper presents this as evidence that mixed-source verification remains challenging for single-forward-pass prediction.Source: Table 3 · p. 6, §2 · p. 6
Effect of Toolset and Training
- With the toolset fixed to MM-VeriTools, GRPO optimization on the remaining 2k rollouts reached 70.4% multi-class accuracy, versus 59.8% when those samples were added as supervised data (SFT-10k).Source: Figure 4 · p. 7, §2 · p. 7
Results
- MM-VeriAgent raises multi-class accuracy from 34.3% to 70.4% and achieves the best binary F1 (85.1%) overall, outperforming MMD-Agent and SCPO on both tasks.T2Agent reports higher multi-class accuracy but uses proprietary GPT-4o with test-time Monte Carlo tree search; MM-VeriAgent uses a single learned policy without test-time search.Source: §2 · p. 7
Effect of Toolset and Training
- Tools alone did not consistently help: untrained Qwen2.5-VL-7B reached 41.3% with MMD-Agent tools versus 26.0% with MM-VeriTools, while GPT-5-mini reached 65.2% versus 57.1%.The passage attributes the small model’s decline to an action space that overwhelms an untrained policy and says prompting cannot fully exploit the extensive toolkit.Source: §2 · p. 7
Verification tools
- Removing CCD tools dropped multi-class accuracy from 70.4% to 58.3% and macro-F1 from 71.5% to 58.6%; removing any tool category degraded both tasks.Textual, visual, and cross-modal tools provide complementary evidence for per-sample policy composition; without cross-modal evidence, the agent often confuses misinformation sources.Source: Table 5 · p. 7, §2 · p. 7
Evidence-grounded reward
- Removing the evidence-grounded reward lowered multi-class accuracy from 70.4% to 66.1% and macro-F1 from 71.5% to 66.7%.Rewarding evidence overlap steers the agent toward source-consistent evidence for fine-grained attribution.Source: Table 5 · p. 7, §2 · p. 7
Limitations
- The evaluation is confined to MMFakeBench, and generalization to unseen forgery generators, news domains, and distortion types remains unestablished.The cache assumes stateless tools with sample-determined inputs; extending the toolkit requires re-benchmarking and retraining, and tool budgets should scale with risk because verification costs more than answer-only classification.Source: §2 · p. 7