Every Opinion Matters: Evaluating and Building Models with Pluralistic Views

Xiang LiAAAI 2025 · DOI

Introduction

Large Language Models (LLMs) have demonstrated remarkable capabilities in reasoning and knowledge retrieval. However, they often act as mirrors of the majority, reflecting dominant cultural norms and majority opinions while failing in "long-tail" or context-specific scenarios. This research addresses the critical need for AI systems that can represent a pluralistic range of human perspectives. By introducing probabilistic box embeddings and new generative evaluation metrics, this work moves beyond the "single ground truth" paradigm toward models that capture the full distribution of valid human responses.

The Limitations of Current Benchmarks

Most AI benchmarks are English-centric and designed around a single correct answer. While this approach is suitable for objective fields like mathematics or formal logic, it falters in common-sense reasoning and cultural knowledge where multiple valid perspectives exist. The research identifies two primary hurdles in the current landscape:

  1. Representation: Traditional probabilistic models do not scale effectively to millions of events, while LLMs—despite their scalability—function as "black boxes" that often produce inconsistent or biased probabilities.
  2. Evaluation: Standard multiple-choice tests are increasingly "brittle." Evidence suggests that models often learn to memorize answer patterns from their massive training sets rather than reasoning through the context. These tests fail to reward models that understand the inherent complexity of the real world.

Methodology: Representing and Evaluating Pluralism

To bridge the gap between rigid probabilistic models and flexible LLMs, the research proposes two primary strategies:

Probabilistic Box Embeddings

Instead of representing concepts as single points in a latent vector space, this method represents joint probability distributions as "boxes" (hyper-rectangles). This geometric approach allows the model to better handle transitive relations. For example, if "A entails B" and "B entails C," the geometric overlap and containment of the boxes naturally represent that relationship. This is particularly effective for capturing word and sentence-level entailment in a way that traditional vectors cannot.

Generative Distributional Evaluation

Moving away from the constraints of multiple-choice testing, the author developed a framework where models generate a ranked list of potential answers. To evaluate these outputs, the research employs a novel approach based on Kullback–Leibler (KL) divergence. This metric measures how closely the model’s predicted distribution of answers aligns with the actual distribution of diverse human responses. This shift ensures that models are penalized for ignoring minority views or over-representing a single perspective.

Key Insights and Results

The "Memorization" Trap

A systematic investigation revealed that LLM performance on traditional benchmarks is often inflated by memorization. Models frequently identify the correct choice based on the format of the question rather than an underlying understanding of the world. By treating common sense as a probability distribution rather than a binary fact, we can more accurately measure true reasoning capabilities.

Performance Gap and Metric Reliability

The research highlights a significant gap between current AI and human performance:

  • Human-AI Gap: On the proposed distributional evaluation, few-shot GPT-4 achieved a KL score of 0.68, while human performance reached 0.06 (where lower scores indicate closer alignment with the true distribution).
  • Metric Reliability: The new KL-based automatic evaluation showed a 0.73 correlation with human scoring, suggesting it is a robust and reliable proxy for human judgment in complex, pluralistic scenarios.

Conclusion and Future Work

While these frameworks provide a path toward more inclusive AI, the research acknowledges that underlying training data remains biased toward specific populations. Future work is focused on expanding the scope of datasets to include more global, non-Western perspectives and refining the geometry of box embeddings to better represent complex, multi-layered knowledge graphs. By treating every opinion as a data point in a broader distribution, we can build models that truly understand the nuances of the human experience.

Summary updated

BibTeX
@inproceedings{li2025every,
  author = {Xiang Li},
  title = {{Every Opinion Matters: Evaluating and Building Models with Pluralistic Views}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2025},
  doi = {10.1609/aaai.v39i27.35111},
}
Figure