Large Language Models Are Challenged by Habitat-Centered Reasoning

Sadaf Ghaffari, Nikhil KrishnaswamyEMNLP 2024 · DOI

Abstract

Large Language Models (LLMs) have shown remarkable progress in identifying object affordances—the actions an object permits. However, they struggle with "habitats," the specific environmental configurations required for those actions to be realized. This paper introduces HabitatQA, a benchmark designed to evaluate reasoning about non-canonical object setups. Results show that even state-of-the-art models like LLaMA 3 and LLaVA plateau at near-random performance on habitat-centered tasks, often due to strong linguistic biases toward typical object orientations.

Introduction

In robotics and AI, affordances describe the actions an object allows (e.g., a chair is for sitting). However, this paper argues that researchers often overlook habitats: the locally-scoped environment or configuration that enables those actions. For instance, a screwdriver's affordance of "screwing" is only enabled if its tip is correctly oriented toward a screw head. If the tool is held by the wrong end or locked away, the affordance is effectively disabled.

Current LLMs and Multimodal LLMs (MLLMs) are typically trained on "Internet-scale" data where objects appear in their most typical, "canonical" orientations. This research investigates whether these models truly understand the underlying physics and logic of object use or if they are simply reciting statistical patterns from their training data.

Methodology and Dataset

The researchers conducted a zero-shot evaluation of 16 open-weight models, including text-only models (LLaMA 2/3, FLAN) and multimodal models (LLaVA, BLIP).

The HabitatQA Benchmark

The authors introduced HabitatQA, a curated dataset consisting of 210 questions and 617 image/question pairs. The dataset draws from three distinct image sources:

  • Natural: Crowdsourced photos of real-world, cluttered scenes.
  • Generated: Images from Stable Diffusion 2.1 to test rare or counterfactual scenarios.
  • Simulated: High-fidelity Unity renders for complex physical dynamics.

Questions were categorized into pairs: one focused on affordances (e.g., "Can this box contain something?") and one on habitats (e.g., "If a knife is in a glass, which way does the handle point?").

Analysis of Model Reasoning

The study revealed several critical insights into how models process physical information:

  • Canonical Bias: Models suffer from "stereotypical" training. They struggle with upside-down objects because their internal bias insists on standard orientations (e.g., chair legs pointing down).
  • Images as Noise: Surprisingly, providing visual context sometimes decreased performance for MLLMs. The visual input often introduced noise that conflicted with strong linguistic priors.
  • The Telic Role Gap: While LLMs learn the purpose (telic role) of an object through text, they fail to grasp the transformations required for that purpose (e.g., what happens when a jar is overturned).

Key Results

The experimental results highlight a significant performance gap between identifying what an object is for and understanding how to use it in context.

Accuracy of various models on affordance-centered questions, where LLaMA 3-70B reaches peak performance.

While LLaMA 3-70B achieved a respectable 77% accuracy on affordance questions, performance on habitat questions across all models hit a "low upper bound" of roughly 57%. In many habitat-centered scenarios, even the most advanced models performed only 7–9% better than random chance.

Accuracy of various models on habitat-centered questions, illustrating the performance plateau and the difficulty of situational reasoning.

Furthermore, visual grounding often failed. Attention maps (Grad-CAM) showed that even when a model correctly identified an object part, it often "grounded" the token to where it expected the part to be based on canonical orientations, rather than where it actually appeared in the image.

Conclusion and Future Work

HabitatQA serves as a challenge set revealing that current AI models lack a robust understanding of physical habitats. Future research will focus on augmenting LLMs with explicit counterfactual reasoning modules and using knowledge distillation to embed common-sense physical logic into smaller, more efficient models.

Summary updated

BibTeX
@inproceedings{ghaffari2024large,
  author = {Sadaf Ghaffari and Nikhil Krishnaswamy},
  title = {{Large Language Models Are Challenged by Habitat-Centered Reasoning}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {2024},
  doi = {10.18653/v1/2024.findings-emnlp.763},
}
Figure