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

Evaluating Contextual Intelligence in Recyclability: A Comprehensive Study of Image-Based Reasoning Systems Eliot Park Harvard College Cambridge, MA 02138 eliot_park@college.harvard.edu Abhi Kumar Stanford University Stanford, CA 94305 abhi1@stanford.edu Pranav Rajpurkar Department of Biomedical Informatics Harvard Medical School Boston, MA 02115 pranav_rajpurkar@hms.harvard.edu 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 guide- lines; (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. 1 Introduction Effective waste management, particularly through recycling, is essential in promoting environmental sustainability. In 2018, the United States generated approximately 292.4 million tons of municipal solid waste, equating to 4.9 pounds per person per day [6]. Of this, 32.1 percent was either recycled or composted, a notable achievement but one that highlights the significant proportion of waste still destined for landfills. Within this waste stream, certain materials, such as paper and paperboard, achieved recycling rates as high as 68.2 percent, while others, such as plastics, lagged far behind at just 8.7 percent [6]. These disparities underscore the need for innovative approaches to improve recycling rates across all categories, including a better understanding by the general public in distinguishing which items should be recycled. arXiv:2601.00905v1 [cs.CV] 31 Dec 2025 - Analyze the materials present and their method of adhesion - Predict object recyclability based on materials and cleanliness - Some objects must be separated before recycling (e.g., a glass jar with a metal cap); if not, it is not recyclable. Separability is difficult to assess. - Analyze transformations (cleanliness and structure) - Predict recyclability based on location-specific guidelines - Difficult to find or generate images that show the same object’s transformation of contamination/damage - Guidelines often insufficient to determine truth labels - City- or country-specific guidelines vary, e.g., how to recycle glass and soiled items - Identify not only the object but its conditions - Use the provided guidelines to predict object recyclability - Guidelines often insufficient to determine truth labels 1. Matching with multiple types of disposal bins 2. Location-specific guidelines 3. Contamination or damage in the object 4. Multi-material objects - Analyze the available bins and the size of the openings - Analyze the object and its material, size, and cleanliness - Place the object in the correct bin - Often fails to recognize all the openings - Attempts to put large items (e.g., box) into small bins. GPT-4o Before: After: No Location Boston London San Francisco Expected: yes Result: yes Expected: “none” Result: “left” Without Location: With Boston Guidelines: Expected: yes Result: no Expected: yes Result: yes Expected: yes Result: yes GPT-4o Task Common Problems Example Claude 3.5 GPT-4o-mini Expected: yes Result: no Figure 1: Overview of our study. Four contextual predictions are tested for three models. 2 Related Works In a recent work [2], the potential of general vision-language models, specifically Contrastive Language-Image Pretraining (CLIP) [5], for automating the classification of waste materials for recycling was explored. The results were substantially better compared to previous approaches using simple convolution neural networks [7, 3, 4] with the model achieving an accuracy of 89% in zero-shot classification into a dozen different disposal methods. However, the approach had notable limitations. CLIP’s reliance on a predefined list of potential items meant it struggled with items outside of this list, reducing its effectiveness in real-world applications where waste items are highly varied. In particular, the model was not designed to handle common but challenging cases such as greasy, dirty, or broken items, w

This content is AI-processed based on open access ArXiv data.