The Normative Reasoning Gap in Language Models

20 concepts10 papers3 proposals

Large language models (LLMs) can explain quantum physics or write complex code with ease, yet they frequently stumble when navigating a simple dinner party conversation. This discrepancy reveals a "normative reasoning gap": a fundamental divide between a model’s ability to process structured, technical information and its ability to apply the unwritten social rules that govern human interaction.

Knowledge Without Experience

The primary challenge is not a lack of factual information, but a failure in application. As Bian et al. argue, current models act as "knowledgeable but inexperienced solvers." They may possess a vast library of commonsense facts, but they lack the "experience" to determine which specific social norm should be pulled "off the shelf" to solve a particular interpersonal problem.

This leads to a performance disparity across different domains of reasoning. While models handle scientific and physical concepts effectively, they struggle with social norms and temporal logic—areas where human knowledge is often implicit and rarely codified in the academic texts that dominate training sets (Bian et al.).

Physical Perception vs. Social Logic

One might assume that failures in social reasoning, especially in multimodal models, stem from a failure to perceive the environment. However, research using the EgoNormia benchmark suggests otherwise. Rezaei et al. and the authors found that even when Vision-Language Models (VLMs) accurately identify objects and actions in a scene, they fail to determine the socially appropriate response.

This indicates that the bottleneck is normative reasoning, not visual perception. A model might see a person crying and identify the "action" correctly, yet fail to grasp the social expectation of offering comfort versus continuing a technical explanation. Similarly, Ghosal et al. demonstrate that while models grasp physical causality (e.g., "if I drop a glass, it breaks"), they lack a deep understanding of the nuanced social rules that govern how humans interact.

The Professional Training Bias

The gap is exacerbated by how models are trained. Ou et al. suggest that training LLMs to be "helpful assistants" may actually make them worse at perceiving human emotion and personality. By prioritizing robotic helpfulness, models lose the grounding in common social experiences necessary for realistic interaction.

Furthermore, training data is often skewed toward academic and professional content. Xue et al. note that this creates a performance gap where models excel at professional knowledge but struggle with the informal logic of daily life. This imbalance leaves models ill-equipped to handle the nuances of casual human interaction, where social harmony often supersedes raw information exchange.

Formalizing the Unwritten

To bridge this gap, researchers have attempted to formalize social logic into structures models can process. Hwang et al. introduced SOCIAL CHEMISTRY, a framework that breaks down everyday situations into "Rules-of-Thumb" (RoTs) across dimensions like legality, moral foundations, and cultural pressure. By treating social norms as a structured "Neural Norm Transformer" task, the authors aim to teach AI to generate human-like social judgments.

Other approaches focus on categorization and context:

  • Semantic Tagging: Wang et al. found that simply tagging questions with categories like "Feelings" or "Social Norms" allows standard models to perform as well as complex graph-based architectures.
  • Social Context: Zhan et al. demonstrated that providing explicit information about the relationship between speakers can more than double a model's ability to detect rude or inappropriate behavior.

The Challenge of Cultural Specificity

Social norms are not universal; they are deeply rooted in specific cultural contexts. Zhan et al. argue that without culture-specific benchmarks, models cannot accurately evaluate social gaffes or provide relevant remediations. For instance, a model trained primarily on Western data may fail to recognize status-based etiquette central to other cultures.

Even when models can generate socially appropriate text, they often struggle to detect when a norm has been violated. Li et al. observed that models frequently misidentify neutral context-setting as social violations, highlighting a persistent lack of deep common-sense reasoning regarding social tension. Bridging the normative reasoning gap requires moving beyond fluent text generation toward a model that understands the "why" behind human behavior.

Key Foundational Work

CIDER: Commonsense Inference for Dialogue Explanation and Reasoning (2021) — contributed 2 concept elements to this direction.
Semantic Categorization of Social Knowledge for Commonsense Question Answering (2021) — contributed 2 concept elements to this direction.
Social Chemistry 101: Learning to Reason about Social and Moral Norms (2020) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_normative_reasoning_gap_aa9c98ee,
  title = {normative reasoning gap},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/normative-reasoning-gap/rdr_aa9c98ee81b14d9e83e023d3df8f78fb},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_normative_reasoning_gap_aa9c98ee_v1,
  title = {normative reasoning gap},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/normative-reasoning-gap/rdr_aa9c98ee81b14d9e83e023d3df8f78fb/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake