💬 NLP

Towards Automated Community Notes Generation with Large Vision Language Models for Combating Contextual Deception

This paper addresses the challenge of automated Community Notes generation for image-based contextual deception by introducing the XCheck dataset, the ACCNote multi-agent framework built on large vision-language models, and the Context Helpfulness Score (CHS) evaluation metric, demonstrating superior performance over baselines and commercial tools.

Original authors: Jin Ma, Jingwen Yan, Mohammed Aldeen, Ethan Anderson, Taran Kavuru, Jinkyung Katie Park, Feng Luo, Long Cheng

Published 2026-03-25

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Jin Ma, Jingwen Yan, Mohammed Aldeen, Ethan Anderson, Taran Kavuru, Jinkyung Katie Park, Feng Luo, Long Cheng

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine social media as a giant, bustling town square. In this square, people are constantly shouting out stories, sharing photos, and spreading news. But sometimes, a troublemaker takes a real photo of a beautiful sunset and claims, "Look! This is a giant laser beam shooting down from the sky!" The photo is real, but the story attached to it is a lie. This is called contextual deception.

For a long time, the town square has relied on a group of volunteer neighbors (called Community Notes) to stand up, read the lies, and write little sticky notes correcting the story. They say, "Actually, that's just a sunset from 2019."

But here's the problem: There are too many lies, and not enough volunteers. The volunteers get tired, the notes take too long to appear, and the lies spread faster than the corrections.

This paper introduces a robot helper designed to do the job of those volunteers, but faster and smarter. Here is how it works, broken down into simple parts:

1. The Problem: The "Fake News" Puzzle

The authors realized that while computers are good at spotting fake photos (like a cat with six legs), they are bad at spotting fake stories about real photos.

The Old Way: Computers just look at the picture and say, "True" or "False."
The New Goal: The computer needs to act like a detective. It needs to say, "This photo is real, but the story is wrong. Here is the real story, and here is the proof."

2. The New Tool: A Detective Team (ACCNOTE)

The researchers built a system called ACCNOTE. Think of it not as one robot, but as a team of specialized detectives working together to solve a mystery.

Detective #1 (The Filter): When a suspicious post appears, this detective goes to the internet (like a massive library) to find similar pictures. But the internet is messy. This detective throws away broken links, ads, and useless pages, keeping only the good, trustworthy sources.
Detective #2 (The Sorter): This detective takes the good sources and sorts them into three piles:
- Pile A: Evidence that supports the lie.
- Pile B: Evidence that proves the lie is false.
- Pile C: Evidence that doesn't matter.
Detective #3 (The Thinkers): A group of robots looks at each pile separately. One tries to write a note based on the "Pro-Lie" evidence, another on the "Anti-Lie" evidence. This prevents them from getting confused by conflicting stories.
Detective #4 (The Judge): This is the boss. It reads all the notes written by the Thinkers and picks the best one. It checks: Is it clear? Is it neutral? Does it have proof? Then, it posts the final correction.

3. The Training Ground: The "XCHECK" Dataset

To teach these robot detectives, the researchers needed a practice field. They couldn't just use fake examples because real lies are tricky.

They went to the actual social media platform (X/Twitter) and collected thousands of real posts where people had already been tricked.
They gathered the original posts, the real corrections written by humans, and the internet evidence that proved the corrections right.
They called this collection XCHECK. It's like a giant textbook of real-life mysteries for the robots to study.

4. The Report Card: The "Helpfulness Score" (CHS)

How do you know if a robot's correction is actually good?

The Old Way: Computers usually check if the robot's words match the human's words (like checking if two essays use the same vocabulary). But that's silly! A robot can say the same thing in a totally different, better way.
The New Way: The researchers asked real humans to grade the notes. They created a new score called CHS (Context Helpfulness Score).
- Instead of asking, "Do these words match?" they asked, "Did this note help you understand the truth?"
- They checked if the note was clear, neutral, and had good sources.
- They found that the old "word-matching" scores were terrible at predicting if a human would actually find the note helpful. The new CHS score was much better.

5. The Results: Robots vs. Humans

The researchers tested their robot team against:

Standard AI models (that just guess without looking up facts).
A popular commercial AI tool (GPT-5-mini).
The human volunteers.

The Winner: The ACCNOTE robot team won.

It was better at spotting the lies than the standard AI.
It wrote better, more helpful notes than the commercial AI tool.
Most importantly, it wrote notes that were verifiable (it included the links to the proof), which made it much more trustworthy.

The Big Picture

This paper is about building a digital fact-checking army. By using a team of AI agents that search for evidence, sort it out, and write clear, neutral corrections, we can fight the spread of misleading stories on social media. It's not just about saying "That's a lie"; it's about saying, "Here is the truth, and here is why."

The ultimate goal is a social media world where lies don't have a chance to spread because the truth is corrected instantly, clearly, and with proof.

1. Problem Definition

The paper addresses the challenge of automated Community Notes generation for image-based contextual deception on social media.

Contextual Deception: Unlike content manipulation (deepfakes/Photoshop), contextual deception involves pairing an authentic, unaltered image with misleading text regarding its time, location, entities, or event.
The Gap: While existing research focuses on binary deception detection (True/False), platforms like X (Twitter) require Community Notes: concise, grounded, and neutral explanations that help users recover the missing or corrected context.
Challenges:
1. Data Scarcity: Lack of real-world datasets containing posts, ground-truth notes, and external evidence.
2. Dynamic Nature: Contextual deception requires retrieving external evidence to verify claims, which is difficult for static pre-trained models.
3. Evaluation Mismatch: Standard text metrics (BLEU, ROUGE) measure lexical overlap but fail to capture whether a note actually improves user understanding or meets criteria like credibility and neutrality.

2. Methodology

The authors propose a "Dataset → Method → Evaluation" framework.

A. Dataset: XCHECK

The authors curated XCHECK, a real-world multimodal dataset derived from the X Community Notes program.

Composition: 2,176 items (50% deceptive, 50% non-deceptive).
- Deceptive: 1,088 posts with ground-truth Community Notes.
- Non-deceptive: 1,088 posts from VisualNews (news articles) serving as controls.
Augmentation: Each item includes external context retrieved via reverse image search (Google Cloud Vision API), resulting in source URLs and summaries.
Annotation: Includes topic taxonomy (14 categories) and deceptive factor taxonomy (10 factors, e.g., Event, Time, Location, Entity).
Statistics: Deceptive posts are heavily concentrated in Politics, War/Conflict, and Conspiracy Narratives.

B. Proposed Method: ACCNOTE

ACCNOTE (Automated Context-Corrective Note generation) is a retrieval-augmented, multi-agent collaboration framework built on Large Vision-Language Models (LVLMs). It consists of three specialized agents:

Data Organizer Agent:
- Filtering: Removes noisy, broken, or irrelevant search results.
- Clustering: Groups remaining evidence into three stances: Supporting, Refuting, and Irrelevant. This prevents conflicting evidence from confusing the generation model.
Reasoner Agents (Pool):
- A pool of agents processes each evidence cluster in parallel.
- Uses Chain-of-Thought (CoT) prompting to:
  1. Identify the post's main claim.
  2. Verify the claim against visual details and internal knowledge.
  3. Verify against the specific evidence cluster.
  4. Determine if the post is deceptive and draft a candidate note with citations.
Judge Agent:
- Evaluates all candidate notes against the five Community Notes criteria: Credibility, Clarity, Relevance, Veracity, and Neutrality.
- Selects the best note as the final output.

C. Evaluation Metric: Context Helpfulness Score (CHS)

To address the failure of standard metrics, the authors developed CHS, a composite metric validated against a pilot user study.

Components:
- $CHS_1$ (Credibility): Cosine similarity of cited URLs vs. ground truth.
- $CHS_2$ (Clarity) & $CHS_4$ (Veracity): ROUGE-L similarity to ground truth text.
- $CHS_3$ (Relevance): Cosine similarity between the generated explanation and (Post Text + Ground Truth Explanation).
- $CHS_5$ (Neutrality): Penalizes strong sentiment polarity (using TextBlob).
Validation: CHS shows a strong correlation ( $\rho = 0.68$ ) with human ratings, significantly outperforming BLEU, ROUGE-L, and METEOR.

3. Key Contributions

XCHECK Dataset: The first real-world multimodal dataset specifically for automated Community Notes generation, featuring paired posts, ground-truth notes, and reverse-searched external context.
ACCNOTE Framework: A novel multi-agent system that effectively filters noise, handles conflicting evidence via stance-aware clustering, and generates grounded notes with explicit citations.
CHS Metric: A new evaluation metric aligned with human perceptions of helpfulness, moving beyond simple lexical overlap to measure credibility, relevance, and neutrality.
Comprehensive Evaluation: Demonstrated that the proposed method outperforms both closed-book baselines and naive RAG approaches.

4. Experimental Results

The authors evaluated ACCNOTE on XCHECK using four open-source LVLM backbones (LLaVA-NeXT, Qwen2.5-VL, InternVL3, NVILA) and compared them against SNIFFER and GPT5-mini.

Deception Detection:
- ACCNOTE significantly improved detection performance over closed-book and naive RAG baselines.
- Best Result: NVILA@15B with ACCNOTE achieved an F1 score of 0.8744 and Accuracy of 0.8745, surpassing the commercial baseline GPT5-mini (F1 ~0.72).
- Naive RAG often degraded performance due to noise; ACCNOTE's filtering and clustering were crucial.
Note Generation:
- ACCNOTE substantially improved the Context Helpfulness Score (CHS) across all models.
- Credibility ( $CHS_1$ ): ACCNOTE drastically improved credibility scores (e.g., from ~0.004 to ~0.33) by explicitly attaching source URLs, a capability missing in naive RAG and closed-book models.
- Qualitative Analysis: Case studies showed ACCNOTE correctly identified misattributed entities and provided verifiable citations, whereas baselines often hallucinated entities or failed to cite sources.

5. Significance

Scalability: Offers a path to automate the labor-intensive Community Notes workflow, reducing latency and human effort while maintaining quality.
Robustness: The multi-agent approach effectively handles the dynamic and noisy nature of real-world web search results, a common failure point for standard RAG.
Responsible AI: By prioritizing neutrality, veracity, and citation, the system aligns with the goals of responsible social networking.
Benchmarking: Establishes XCHECK and CHS as new standards for evaluating automated fact-checking and context-corrective systems.

The paper concludes that while challenges remain regarding precision-recall trade-offs and system modularity, ACCNOTE represents a significant step forward in automating the detection and explanation of image-based contextual deception.