LLM-based Detection of Manipulative Political Narratives

Technical Summary: LLM-based Detection of Manipulative Political Narratives

Problem Statement
The paper addresses the critical challenge of detecting Foreign Information Manipulation and Interference (FIMI) within large-scale, unfiltered social media datasets. The primary difficulty lies in distinguishing between legitimate, albeit controversial, political critique and coordinated, manipulative strategic narratives. Traditional approaches, such as standard topic modeling or fact-checking for veracity, often fail because manipulative content frequently relies on the malicious reframing of factual events (malinformation) rather than pure falsehoods. Furthermore, existing methods struggle to separate coordinated state-aligned campaigns from genuine populist or personal political opinions, often overlooking the specific rhetorical nuances and intent that define strategic narratives.

Methodology
The authors propose a Large Language Model (LLM) driven data-processing pipeline designed to detect, filter, and cluster manipulative narratives without relying on a predefined list of target categories. The methodology consists of four sequential stages:

1. Prompt-Based Filtering:

   • Objective: To isolate manipulative posts from legitimate criticism and unrelated content.
   • Mechanism: A reasoning model (Qwen3.5-122B) is guided by a detailed few-shot prompt. This prompt defines FIMI characteristics (e.g., blame-shifting, moral inversion, victimhood) and explicitly instructs the model to differentiate between coordinated manipulation and standard political skepticism.
   • Constraints: The prompt includes specific "known motifs" (e.g., Doppelgänger, Storm-1516) and explicitly excludes false positives such as normal government criticism or personal economic frustration. The model outputs a boolean classification and a reasoning string in JSON format.
   • Strategy: The system prioritizes high recall (0.92) over precision during this stage, accepting a higher rate of false positives to ensure no potential narrative fragments are discarded.

2. Intent-Driven Embedding:

   • Objective: To map the filtered posts into an embedding space where proximity reflects manipulative intent rather than just topical similarity.
   • Mechanism: The authors utilize the Qwen3-Embedding-8B model. Unlike standard sentence transformers, this model is prompted with specific instructions to "Identify the strategic narrative, manipulative intent, and underlying disinformation motive." This ensures that posts sharing similar rhetorical strategies and malicious intents are clustered closely together.

3. Dimensionality Reduction and Clustering:

   • Mechanism: The high-dimensional embeddings (4096 dimensions) are reduced using UMAP (Uniform Manifold Approximation and Projection) to five dimensions.
   • Clustering: HDBSCAN (Hierarchical Density-Based Spatial Clustering of Applications with Noise) is applied to the reduced space. This unsupervised algorithm determines the number of clusters dynamically based on density.
   • Optimization: Hyperparameters were tuned (specifically min_cluster_size) to balance noise reduction with semantic distinctness. A minimum cluster size of 400 was selected, resulting in 41 distinct clusters while keeping noise levels manageable (~29.7%).

4. Strategic Narrative Extraction:

   • Objective: To generate human-readable labels for each cluster that describe the underlying storyline.
   • Mechanism: A larger reasoning model (Qwen3.5-397B) is employed. It receives the cluster's documents and c-TF-IDF keywords. Using a few-shot prompt, the model synthesizes a complete narrative label including the core claim, the targeted adversary, and the manipulative angle.

Key Contributions
The paper outlines three primary technical contributions:

1. Prompt-Based Reasoning: Moving beyond traditional BERT-based classification, the authors introduce a reasoning approach guided by explicit FIMI characteristics and few-shot examples. This successfully isolates strategic manipulative content from legitimate critique based on rhetorical nuances.
2. Intent-Driven Embedding: The pipeline shifts the focus of the embedding model from general topics to specific manipulative intents. By explicitly configuring the embedding model via prompt, related storylines are positioned closer in the vector space, facilitating more accurate clustering.
3. Strategic Narrative Extraction: The authors replace standard topic extraction (keyword lists) with a specialized prompt that forces the model to output complete storylines. These narratives explicitly capture the core claim, the enemy, and the manipulative angle, providing a richer understanding of the campaign structure.

Results
Applied to a dataset of 1,255,895 social media posts (from X, Reddit, and Telegram) collected prior to the 2025 German federal elections, the pipeline successfully identified 41 distinct manipulative narrative clusters.

• Filtering Performance: The initial filtering achieved an F1 score of 0.77, characterized by high recall (0.92) and lower precision (0.66). The authors note that the model is stricter than human raters, often flagging individual populist opinions as manipulative; however, the downstream HDBSCAN clustering effectively filters these false positives as noise, as they do not form dense, coherent clusters.
• Narrative Themes: The extracted clusters were grouped into four main thematic pillars:
  1. The "Great Replacement": Narratives framing migration as a weapon to destroy the native population.
  2. The Proxy War: Storylines portraying Germany as a US colony sacrificing national peace for foreign interests (specifically regarding Russia).
  3. The Climate Dictatorship: Framing climate policy as a fabricated hoax by globalists to control and impoverish citizens.
  4. The AfD as Savior: Positioning the right-wing AfD party and external actors (e.g., Elon Musk) as the only saviors against a corrupt "deep state."

Significance and Claims
The paper claims that its approach effectively bridges the gap between detecting coordinated disinformation and understanding the underlying strategic narratives. By integrating prompt-based filtering with unsupervised clustering, the system can uncover new narrative groups without relying on a predefined taxonomy.

A significant finding is the synergy between the high-recall filtering stage and the density-based clustering stage. The pipeline leverages the LLM's ability to apply its internal knowledge (e.g., recognizing specific domains like rtde.media as proxies for state media) to evaluate posts, even when such specific entities are not explicitly listed in the prompt. The authors argue that this method is superior to traditional topic modeling because it captures the intent and storyline of manipulation rather than just the subject matter.

The paper maintains a modest stance regarding its limitations, acknowledging the "blurry line" between coordinated manipulation and personal radical political views. It emphasizes that the extracted narratives are unedited outputs intended for transparency and do not reflect the authors' political views. The work concludes that while the current approach is effective, future iterations could benefit from programmatic prompt optimization and the integration of visual language models to handle multimodal content.