From We to Me: Theory Informed Narrative Shift with Abductive Reasoning

This paper proposes a neurosymbolic approach that leverages social science theory and abductive reasoning to automatically extract rules for guiding Large Language Models in effectively shifting narratives between individualistic and collectivistic frameworks while preserving the original message's core meaning.

The Gist

Imagine you are a master translator, but instead of translating words from French to English, you are translating stories from one culture's mindset to another's.

Specifically, this paper tackles the challenge of taking a story written with a Collectivist mindset (where the group, family, and community are everything) and rewriting it to sound Individualistic (where the hero, personal choice, and self-reliance are everything), without losing the original plot.

Here is the breakdown of the paper using simple analogies:

1. The Problem: The "Lazy" Translator

Think of a standard AI (like a basic version of ChatGPT) as a lazy tourist who is asked to translate a story.

• The Request: "Rewrite this story about a village working together to build a dam, but make it sound like it's about one brave hero doing it alone."
• The Result: The lazy tourist just changes a few words here and there. They might say, "The hero built the dam," but they forget to change the feeling of the story. The AI often misses the subtle clues that make a story feel "group-oriented" (like phrases such as "all hands on deck" or "we did this").
• The Paper's Finding: Current AI is bad at this. It either fails to change the story's "soul" or it changes the plot so much that it's no longer the same story.

2. The Solution: The "Detective + Architect" Team

The authors propose a new method called Neurosymbolic Abductive Reasoning. Let's break that scary name down into a team of two:

• The Detective (Social Science Theory): Before the AI writes anything, a "Detective" (based on real psychology rules about individualism vs. collectivism) reads the story. The Detective asks: "What specific parts of this story make it feel like a group effort?"
  • Example: "Ah, the phrase 'young, old, weak, strong' is a clue that this is about the community. We need to change this."
• The Architect (Abductive Reasoning): This is the logic engine. It doesn't just guess; it works backward. It asks: "If we want the story to feel like a solo hero, what specific ingredients must we swap out?"
  • It creates a "shopping list" of changes: "Change 'all hands' to 'one determined soul'."

The Magic: The AI (the writer) only gets to work after the Detective and Architect have handed it a precise list of what to change. This prevents the AI from hallucinating or messing up the story.

3. The Analogy: Renovating a House

Imagine you have a house built for a large, multi-generational family (Collectivist). It has one giant dining room, shared bedrooms, and a communal kitchen.

• The Goal: You want to renovate it to suit a single, independent person (Individualist).
• The Bad Way (Zero-Shot AI): You just put a "For Sale" sign on the house and hope the new owner likes it. Or, you paint the walls red but leave the giant dining room. It feels weird and mismatched.
• The Paper's Way:
  1. Inspect: You walk through and identify exactly what needs to change (e.g., "Divide the dining room into a private study," "Lock the communal kitchen").
  2. Plan: You draw a blueprint (the logic rules) that ensures the house still functions as a home, just for one person.
  3. Build: You hire a contractor (the LLM) to make only those specific changes.
  4. Result: The house is now perfect for a solo dweller, but the foundation and the view (the core story) are exactly the same.

4. The Results: Why It Matters

The researchers tested this on many different AI models (GPT-4, Grok, Llama, etc.).

• Success Rate: Their method was 55% better at successfully shifting the story's mindset than just asking the AI to "do it."
• Fidelity: Crucially, their method kept the story's meaning intact. The "Zero-Shot" AI often changed the plot so much it became a different story. The new method kept the story recognizable.
• Efficiency: The AI didn't need to rewrite the whole story. It only changed about 32% of the words (the specific "clues" the Detective found), leaving the rest untouched.

5. The Big Picture

This isn't just about rewriting fairy tales. This is about communication.

• If you are a diplomat trying to explain a policy to a culture that values the group, you need to speak their language.
• If you are a journalist reporting on a tragedy, you might need to frame it differently for different audiences to ensure they understand the cause and effect correctly.

In short: The paper teaches AI how to be a cultural chameleon. It gives the AI a "rulebook" from social scientists so it can change the vibe of a story without breaking the plot, ensuring the message lands perfectly with the intended audience.

Technical Summary

Here is a detailed technical summary of the paper "From We to Me: Theory Informed Narrative Shift with Abductive Reasoning."

1. Problem Statement

The paper addresses the challenge of narrative shift: transforming a text to reflect a different cultural worldview (specifically between Individualism and Collectivism) while preserving the original story's core events and semantic fidelity.

• The Gap: Current Large Language Models (LLMs) struggle with this task. When prompted with a zero-shot instruction (e.g., "Make this story more individualistic"), LLMs often fail to identify the specific narrative markers that signal cultural orientation (e.g., changing "all hands dig" to "one determined soul"). Instead, they either fail to transform the text or generate incoherent outputs that lose the original story's meaning.
• The Goal: To develop a system that systematically alters the cultural framing of a narrative (e.g., shifting from a group-focused to an individual-focused perspective) without distorting the underlying plot or facts.

2. Methodology: Neurosymbolic Abductive Reasoning

The authors propose a neurosymbolic approach that combines social science theory, logical reasoning, and LLM generation. The system operates in two phases:

A. Theoretical Foundation: Narrative Diagnosis

The team developed a 20-item diagnostic survey grounded in cross-cultural psychology (Hofstede, Triandis, etc.).

• The survey evaluates narratives across 40 dimensions (20 individualistic, 20 collectivistic).
• It assesses specific elements: conflict, actors, goals, relationships, and moral evaluation.
• An LLM acts as a "diagnostician," scoring a story on these dimensions to generate a quantitative profile of its current narrative orientation.

B. Formalization as an Abduction Problem

The transformation is formalized as an Abductive Reasoning problem within a logical framework:

• Observations ($O$): The current state of the story, represented as annotated facts (e.g., c_feat(chunk, collectivism) = 0.8).
• Hypothesis ($H$): A set of potential modifications (changing specific chunks to have different feature ratings).
• Logic Program ($\Pi$): Rules derived from the training corpus that define how specific story chunks contribute to the overall narrative orientation.
• Goal: Find an Explanation ($E$)—a minimal set of chunks to modify—that, when transformed, maximizes the similarity to the target narrative corpus (e.g., shifting the score from 0.2 to 0.8) while maintaining consistency.

C. Two-Phase Architecture

1. Phase 1: Rule Learning:
   • The system analyzes a training corpus of stories with known orientations.
   • It learns logical rules (e.g., "If a story contains a chunk emphasizing 'shared responsibility,' it contributes to a collectivistic score").
   • These rules are encoded as a logic program ($\Pi$).
2. Phase 2: Iterative Abductive Transformation:
   • Diagnosis: The system diagnoses the input story to identify its current narrative features.
   • Abduction: It solves the abduction problem to identify which specific chunks need modification to achieve the target orientation with maximum parsimony (minimal changes).
   • Transformation: The identified chunks are sent to an LLM with a specific prompt to rewrite them according to the target orientation.
   • Iteration: This process repeats for a set number of iterations ($T_{max}$), selecting the version that best balances narrative shift and semantic fidelity.

3. Key Contributions

• Novel Framework: The first approach to treat narrative cultural shift as an abductive reasoning problem, moving beyond simple prompt engineering or fine-tuning.
• Theory-Grounded: Unlike generic style transfer, this method is explicitly grounded in social science theory (Individualism vs. Collectivism), ensuring the transformations are culturally coherent rather than just stylistic.
• Neurosymbolic Integration: It effectively bridges symbolic logic (for identifying what to change) and neural generation (for how to change it), preventing the "hallucinations" common in zero-shot LLM attempts.
• Efficiency: The method targets only essential narrative segments (chunks), preserving the majority of the original text.

4. Experimental Results

The authors evaluated their method against a Zero-Shot LLM baseline across multiple models (GPT-4o, Grok-4, Llama-4, Deepseek-R1) and two directions: Collectivistic $\to$ Individualistic (C$\to$I) and Individualistic $\to$ Collectivistic (I$\to$C).

Key Metrics:

• Narrative Shift: Measured by the improvement in the diagnostic survey score (how well the new story aligns with the target culture).
• Semantic Fidelity: Measured by KL Divergence (lower is better; indicates how close the transformed story is to the original).

Findings:

• Superior Performance: The abduction-guided approach significantly outperformed the zero-shot baseline.
  • C$\to$I Shift (GPT-4o): Achieved a 55.88% improvement over the baseline in narrative shift while maintaining 40.4% better semantic similarity (lower KL divergence).
  • I$\to$C Shift: Showed comparable improvements, with GPT-4o outperforming the baseline by 15% in shift while preserving 36% more similarity.
• Model Agnosticism: The method improved performance across both "reasoning" models (GPT-4o, Deepseek-R1) and "non-reasoning" models (Llama-4, Grok-4).
• Efficiency: The system modified only a median of 32.11% of the tokens in a story, demonstrating that targeted abductive reasoning is more efficient than rewriting the entire text.
• Deepseek-R1 Nuance: While Deepseek-R1's zero-shot baseline performed competitively for I$\to$C shifts (likely due to inherent biases in the model), it suffered from a 39.10% loss in content similarity compared to the authors' method.

5. Significance

• Cross-Cultural Communication: This technology enables the creation of messages that align with specific cultural worldviews, which is critical for diplomacy, journalism, and mitigating political polarization.
• Intelligence & Information Operations: The ability to tailor narratives for different audiences (domestic, allied, adversary) while maintaining factual consistency has significant applications in information operations.
• Advancing LLM Reasoning: The paper demonstrates that constraining LLMs with formal logic and domain-specific theory yields more reliable and coherent results than prompting alone, offering a blueprint for future "Theory-Grounded" AI systems.

In summary, the paper presents a robust, theory-driven framework that solves the "narrative shift" problem by using abductive reasoning to pinpoint exactly what needs to change in a story, ensuring the result is both culturally distinct and semantically faithful to the original.