Evaluating Contextual Intelligence in Recyclability: A Comprehensive Study of Image-Based Reasoning Systems

TL;DR

The paper tackles the problem of enabling accurate recyclability judgments from images in the face of variable local guidelines and item conditions. It leverages state-of-the-art vision-language models (GPT-4o, GPT-4o-mini, Claude 3.5) across four controlled experiments to test bin-type reasoning, location-specific guidelines, phase changes, and multi-material item classification. Key findings show GPT-4o generally provides the strongest performance, yet challenges remain with size reasoning (bin openings) and certain material combinations, underscoring the role of prompt design and contextual adaptation. The work demonstrates the potential of context-aware AI to support public recycling practices and environmental sustainability, while also outlining practical deployment considerations and gaps for future refinement.

Abstract

While the importance of efficient recycling is widely acknowledged, accurately determining the recyclability of items and their proper disposal remains a complex task for the general public. In this study, we explore the application of cutting-edge vision-language models (GPT-4o, GPT-4o-mini, and Claude 3.5) for predicting the recyclability of commonly disposed items. Utilizing a curated dataset of images, we evaluated the models' ability to match objects to appropriate recycling bins, including assessing whether the items could physically fit into the available bins. Additionally, we investigated the models' performance across several challenging scenarios: (i) adjusting predictions based on location-specific recycling guidelines; (ii) accounting for contamination or structural damage; and (iii) handling objects composed of multiple materials. Our findings highlight the significant advancements in contextual understanding offered by these models compared to previous iterations, while also identifying areas where they still fall short. The continued refinement of context-aware models is crucial for enhancing public recycling practices and advancing environmental sustainability.

Figures (3)

• Figure 1: Overview of our study. Four contextual predictions are tested for three models.
• Figure 2: Performance on Bin Testing (Experiment 1). Recyclability prediction was made in the context of the available bin type, requiring accurate assessment of size of the item and the opening.
• Figure 3: Three additional tests for contextual predictions. (a) City-specific guidelines were added to the prompt. The direction of the arrows indicates that the classification improved (or remained the same) as a result. (b) Pairs of images were tested for contamination (e.g., soiled item) and structural changes (e.g., broken glass). The changes in the pair predictions are illustrated with the blue and pink bands corresponding to correct and incorrect items, respectively. For instance, for GPT-4o, four pairs were Y-N pairs (recyclable prior to contamination / not recyclable after contamination); three items were predicted correctly (blue) but one item was predicted to be N-N. In general, contaminated items that are still recyclable were incorrectly predicted as no longer being recyclable.