Language Models Can Resolve Reference Compositionally, But It's Not Their Native Strength: The Case of the Personal Relation Task

2025-02-27 · Source: cs.CL updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Expert, extended

Summary

A study evaluated Large Language Models (LLMs) and humans on compositional reference resolution using the Personal Relation Task (PRT), which involves interpreting complex noun phrases within a defined universe of people and relationships. The task differentiates between Extensional (identifying a specific referent) and Intensional (providing a formal, symbolic representation) interpretations. Researchers found that humans performed better on Extensional tasks (82.8% accuracy) than Intensional (71.1%), while LLMs exhibited the opposite strength, achieving 95.0% accuracy on Intensional tasks versus 79.8% on Extensional. Both groups performed better with English representations than abstract ones. LLM performance decayed with increasing complexity, dropping from 93.0% at complexity 3 to 82.4% at complexity 6. Reasoning-enhanced models demonstrated a more stable performance across complexities, narrowing the gap between Extensional and Intensional tasks.

Key takeaway

For AI Scientists and Machine Learning Engineers designing or evaluating LLM applications requiring deep semantic understanding, you should recognize that current LLMs, despite high overall accuracy, fundamentally differ from humans in compositional understanding. Your models will likely excel at symbolic transformations (Intensional tasks) but struggle with concrete referential grounding (Extensional tasks), especially with increasing complexity or abstract representations. Prioritize integrating explicit grounding mechanisms or using reasoning-enhanced models for applications demanding robust real-world reference resolution.

Key insights

LLMs excel at symbolic compositional interpretation (Intensional) but struggle with referential grounding (Extensional), unlike humans.

Principles

• LLMs' text-only training limits their referential grounding capabilities.
• Compositional tasks can be distinguished by sense (Intensional) versus reference (Extensional).
• Reasoning-enhanced LLMs can mitigate weaknesses in extensional reasoning.

Method

The Personal Relation Task (PRT) evaluates compositional interpretation by asking models/humans to resolve complex noun phrases in a defined relational universe, distinguishing between identifying referents (Extensional) and formal representations (Intensional).

In practice

• Design LLM evaluations to explicitly contrast intensional vs. extensional understanding.
• Consider reasoning-enhanced LLMs for tasks requiring robust referential grounding.
• Use abstract representations to control for pretraining data contamination.

Topics

• Large Language Models
• Compositionality
• Semantic Interpretation
• Referential Grounding
• Extensional Semantics
• Intensional Semantics
• Reasoning Models

Code references

• avatars16/Personal-Relation-Task-on-LLMs

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, NLP Engineer

Related on AIssential

• Probing Semantic Alignment, Lexical Invariance, and Syntactic Influence in LLM Metaphor Processing — LLMs struggle with deep metaphor comprehension, often misinterpreting concepts and relying on lexical associations over context or full syntactic understanding.
• Forget LLM: MIT's New RLM (Phase Shift in AI) — MIT's Recursive Language Model (RLM) uses a neurosymbolic approach to overcome LLM context rot and enhance reasoning.
• LLM Parameters for Math Across Languages: Shared or Separate? — LLM mathematical reasoning involves partially shared and partially language-specific parameters, with English dominating parameter sets.
• Presupposition and Reasoning in Conditionals: A Theory-Based Study of Humans and LLMs — LLMs often match human linguistic judgments via surface patterns, not deep pragmatic reasoning.
• Your AI Probably Understands Your Marriage Joke Better Than Your SQL Parser — LLMs excel at interpreting human language's implied, non-literal meanings, unlike literal parsers.
• Grounding vs. Compositionality: On the Non-Complementarity of Reasoning in Neuro-Symbolic Systems — Symbol grounding alone is insufficient for compositional generalization; explicit reasoning objectives are necessary.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.CL updates on arXiv.org.