Generating Place-Based Compromises Between Two Points of View

1. Problem Statement

Large Language Models (LLMs) excel in academic reasoning but struggle with social intelligence, particularly in generating empathically neutral compromises between opposing viewpoints. While LLMs can produce consensus statements or empathetic responses, they often fail to act as neutral mediators that propose resolutions acceptable to both parties without favoring one side.

• The Gap: Existing methods often focus on broad consensus or simple empathy, but generating a compromise requires proposing a solution where both parties are better off than without the compromise, requiring a balance of empathy toward conflicting stances.
• The Challenge: LLMs tend to anchor on one viewpoint or fail to balance the emotional weight of opposing arguments, leading to biased or unacceptable outcomes.

2. Methodology

The authors propose a two-stage framework: (1) Prompt Engineering to generate high-quality synthetic data using a large proprietary model, and (2) Model Alignment to distill these capabilities into smaller, open-source models without requiring external feedback during inference.

A. Dataset: ContrastingViews

The study utilizes a subset of the EmpathyFromPerspectives dataset (Chen et al., 2025), containing 2,400 pairs of contrasting views regarding public spaces (e.g., parks, trails).

• View A: Positive perspective (feeling safe/welcome).
• View B: Negative perspective (feeling unsafe/excluded).
• Each view includes a place description, reasoning, and suggested improvements.

B. Stage 1: Prompt Engineering & Data Generation

The authors evaluated four prompting strategies using Claude 3 Opus to generate compromises:

1. Single Prompt: Basic instruction to generate a compromise.
2. Chain of Thought (CoT): Explicitly asking the model to identify similarities before generating.
3. CoT + LLM Self-Evaluation: The model generates a compromise and then self-evaluates its empathy similarity.
4. CoT + Feedback (Iterative): The most effective method. It uses an external Empathic Similarity Model (based on the e5-large embedding model) to score the generated compromise against both View A and View B.
   • Mechanism: The system calculates the absolute difference between the similarity scores ($|score_A - score_B|$). If the difference is high, the model is prompted to refine the response to minimize this gap, iterating until the compromise is empathically neutral.

C. Stage 2: Human Evaluation

A study with 50 participants evaluated the acceptability of compromises generated by the different strategies.

• Method: Participants were assigned a specific viewpoint (either the positive or negative one) and asked to rate the acceptability of compromises from that perspective.
• Metric: Acceptability score (1-100) and preference ranking (First/Second choice).

D. Stage 3: Alignment of Smaller Models

To make the system efficient and remove the need for the proprietary model and external similarity scoring during inference, the authors trained smaller open-source models (Llama-3.1 8B and Mistral-7B).

• Approach: Instead of standard fine-tuning (imitation), they used Noise Contrastive Estimation (NCE) based alignment.
• Rationale: Inspired by how humans learn social norms by distinguishing between conforming and violating behaviors, the model is trained to discriminate between "good" (human-validated) compromises and "bad" (base model) outputs.
• Loss Function: The NCE loss maximizes the likelihood of the target (good) compromise while minimizing the likelihood of the base model's output. This preserves the model's broad world knowledge (avoiding catastrophic forgetting) while internalizing the preference for neutrality.
• Alternative: They also explored a Task-Loss based alignment that incorporates ROUGE scores to ensure semantic fidelity to the target compromise.

3. Key Contributions

1. Empathic Neutrality Framework: Introduced a novel method for generating compromises that are balanced not just logically, but emotionally, by minimizing the empathy gap between opposing viewpoints.
2. Iterative Prompt Engineering: Demonstrated that CoT + External Feedback (using an empathy similarity model) significantly outperforms standard CoT and self-evaluation methods in generating neutral compromises.
3. NCE-Based Alignment: Proposed a method to align smaller LLMs to generate empathically neutral compromises without requiring external reward models or similarity estimators at inference time. This approach outperforms standard fine-tuning and avoids the catastrophic forgetting issues associated with Direct Preference Optimization (DPO) in this context.
4. Synthetic Dataset Creation: Generated a high-quality dataset of 2,400+ place-based compromises, validated by human preference, to train and evaluate social intelligence in LLMs.

4. Results

Human Evaluation (Prompt Engineering)

• Preference: The CoT+Feedback method was the clear winner.
  • First Preference: CoT+Feedback 1 (37%) and CoT+Feedback 2 (40%).
  • Comparison: Significantly outperformed CoT (18%) and Single Prompt (5%).
  • Statistical Significance: The preference for CoT+Feedback over Single Prompt was statistically significant ($p \le 0.002$).
• Acceptability: Compromises generated with empathic neutrality feedback were rated as highly acceptable by both positive and negative viewpoint holders.

Model Alignment (Smaller Models)

• Generation Quality: Aligned models (Llama+NCE and Mistral+NCE) achieved the highest ROUGE scores (e.g., Llama NCE: 0.315 ROUGE-1 vs. Base: 0.164), indicating better semantic alignment with human-validated compromises.
• Empathic Neutrality: The aligned models produced compromises with a significantly smaller "neutrality gap" ($|score_A - score_B|$), approaching the performance of the iterative prompt-engineered outputs.
• Knowledge Retention: Unlike DPO, which caused catastrophic forgetting in initial experiments, the NCE and Task-Loss approaches preserved the models' pre-trained knowledge (verified via WikiText log-likelihood tests).

5. Significance and Future Work

• Practical Application: This work provides a scalable pipeline for deploying socially intelligent AI in conflict resolution, collaborative decision-making, and mediation scenarios where impartiality is critical.
• Efficiency: By distilling the capabilities of a large proprietary model into smaller open-source models via NCE alignment, the system becomes deployable without the computational cost of iterative external feedback loops.
• Limitations & Future Directions:
  • The current dataset is limited to place-based conflicts (safety/welcomingness). Future work aims to expand to abstract domains like policy debates and ethical dilemmas.
  • The study acknowledges potential biases in the training data and the need to address power asymmetries in future iterations.
  • The authors plan to explore multi-turn dialogues to simulate ongoing negotiations rather than single-shot compromises.

In conclusion, the paper establishes that external empathic feedback is crucial for training LLMs to generate neutral compromises and that NCE-based alignment is an effective, efficient method for transferring these social intelligence capabilities to smaller, open-source models.