Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Where Norms and References Collide: Evaluating LLMs on Normative Reasoning

Mitchell Abrams, Kaveh Eskandari Miandoab, Felix Gervits, Vasanth Sarathy, Matthias Scheutz

TL;DR

This work introduces SNIC, a diagnostic testbed for norm-based reference resolution in embodied contexts, to evaluate whether LLMs can extract and apply physically grounded social norms. Through a three-phase data-generation pipeline (seed elicitation, validation, and procedural augmentation), the authors create 9,000 SNIC instances from 51 validated seeds and test a suite of LLMs under varying inputs, including explicit norm lists and formal representations. Results show that LLMs struggle to consistently apply norms without explicit normative cues, but performance substantially improves when norms are provided, underscoring a gap in intrinsic normative grounding and the need for alignment for normative reasoning in agentic systems ($\rho$-based analyses reveal norm interactions). The SNIC dataset thus offers a controlled, interpretable benchmark for evaluating normative reasoning in socially situated NLP tasks with potential impact on assistive robotics and embodied AI.

Abstract

Embodied agents, such as robots, will need to interact in situated environments where successful communication often depends on reasoning over social norms: shared expectations that constrain what actions are appropriate in context. A key capability in such settings is norm-based reference resolution (NBRR), where interpreting referential expressions requires inferring implicit normative expectations grounded in physical and social context. Yet it remains unclear whether Large Language Models (LLMs) can support this kind of reasoning. In this work, we introduce SNIC (Situated Norms in Context), a human-validated diagnostic testbed designed to probe how well state-of-the-art LLMs can extract and utilize normative principles relevant to NBRR. SNIC emphasizes physically grounded norms that arise in everyday tasks such as cleaning, tidying, and serving. Across a range of controlled evaluations, we find that even the strongest LLMs struggle to consistently identify and apply social norms, particularly when norms are implicit, underspecified, or in conflict. These findings reveal a blind spot in current LLMs and highlight a key challenge for deploying language-based systems in socially situated, embodied settings.

Where Norms and References Collide: Evaluating LLMs on Normative Reasoning

TL;DR

This work introduces SNIC, a diagnostic testbed for norm-based reference resolution in embodied contexts, to evaluate whether LLMs can extract and apply physically grounded social norms. Through a three-phase data-generation pipeline (seed elicitation, validation, and procedural augmentation), the authors create 9,000 SNIC instances from 51 validated seeds and test a suite of LLMs under varying inputs, including explicit norm lists and formal representations. Results show that LLMs struggle to consistently apply norms without explicit normative cues, but performance substantially improves when norms are provided, underscoring a gap in intrinsic normative grounding and the need for alignment for normative reasoning in agentic systems (-based analyses reveal norm interactions). The SNIC dataset thus offers a controlled, interpretable benchmark for evaluating normative reasoning in socially situated NLP tasks with potential impact on assistive robotics and embodied AI.

Abstract

Embodied agents, such as robots, will need to interact in situated environments where successful communication often depends on reasoning over social norms: shared expectations that constrain what actions are appropriate in context. A key capability in such settings is norm-based reference resolution (NBRR), where interpreting referential expressions requires inferring implicit normative expectations grounded in physical and social context. Yet it remains unclear whether Large Language Models (LLMs) can support this kind of reasoning. In this work, we introduce SNIC (Situated Norms in Context), a human-validated diagnostic testbed designed to probe how well state-of-the-art LLMs can extract and utilize normative principles relevant to NBRR. SNIC emphasizes physically grounded norms that arise in everyday tasks such as cleaning, tidying, and serving. Across a range of controlled evaluations, we find that even the strongest LLMs struggle to consistently identify and apply social norms, particularly when norms are implicit, underspecified, or in conflict. These findings reveal a blind spot in current LLMs and highlight a key challenge for deploying language-based systems in socially situated, embodied settings.
Paper Structure (11 sections, 3 figures, 4 tables)

This paper contains 11 sections, 3 figures, 4 tables.

Table of Contents

  1. Introduction
  2. Background
  3. Dataset Generation and Augmentation Pipeline
  4. Step 1: Seed Dataset Elicitation (n=120)
  5. Step 2: Validation Analysis and Seed Set Filtering (n=51)
  6. Step 3: Procedural Augmentation of 9,000 Cases
  7. 3.1: Ground Truth and Quality Assurance
  8. Dataset
  9. Evaluations
  10. Discussion & Conclusion
  11. Limitations

Figures (3)

  • Figure 1: Sample text-vignette from human-subject experiment validating our dataset. This question is a reference task with the social norm In a cooking task, you should cook with tools that are clean.
  • Figure 2: Data augmentation pipeline to create our final dataset (SNIC) (n=9,000) for LLM evaluation.
  • Figure 3: Spearman Correlation of Social Norms