Search Abstracts | Symposia | Slide Sessions | Poster Sessions
Age-related Behavioral Differences in Picture-Elicited Discourse and Their Simulation with Fine-Tuned Multimodal Language Models
Poster Session B, Wednesday, September 30, 4:30 - 6:30 pm, Wangari Maathai
This poster is part of the Sandbox Series.
Qimiao GAO1, Jixing LI2; 1City University of Hong Kong
Introduction. Picture-description tasks provide a naturalistic yet controlled way to examine language production across development and aging. Children, adults, and older adults may differ not only in how much they say, but also in how they organize utterances, select lexical items, use function words and connectors, and construct syntactic structures. In the present study, we first examined age-related behavioral differences in Chinese picture-elicited discourse. We then used these human behavioral profiles as targets for simulating age-specific language production with fine-tuned multimodal large language models. Finally, we examined whether adapting the models to different age groups produced systematic internal gradient changes in model weights, providing a model-internal diagnostic of how age-conditioned behavioral patterns are learned. Methods. The human corpus included 345 participants and 1,760 task-level picture-description records from four age groups: children aged 0–11 years, 127 participants and 495 records; adolescents aged 12–18 years, 62 participants and 239 records; adults aged 19–60 years, 133 participants and 850 records; and older adults aged 60 years and above, 23 participants and 176 records. For each response, we extracted behavioral and linguistic measures including response length, mean length of utterance, lexical diversity, function-word ratio, connector use, utterance count, and syntactic node count. Representational analyses were restricted to the three tasks shared by all four age groups: cat, umbrella, and window. For each matched task, 2,048-dimensional response embeddings were L2-normalized, projected using principal component analysis, and tested with permutational multivariate analysis of variance. Based on the observed human age-group profiles, we then generated 4,000 synthetic image-grounded descriptions from 200 line-art scenes and fine-tuned separate age-conditioned Qwen2.5-Omni models to simulate the language-production patterns of different age groups. Model outputs were compared with human behavioral profiles, and parameter-gradient heatmaps were used to examine which model components showed stronger training-related changes during age-specific fine-tuning. Results. Human picture-elicited discourse showed significant age-related behavioral and representational differences. Response embeddings revealed reliable but overlapping age-group structure across all three matched tasks: cat, R² = 0.119, p = 0.001; umbrella, R² = 0.162, p = 0.001; and window, R² = 0.240, p = 0.001. These results suggest that age differences in picture-description behavior are not strictly categorical, but reflect graded shifts in lexical composition, utterance organization, and representational structure. Fine-tuned multimodal models reproduced several human-like behavioral patterns, especially response length, lexical density, lexical diversity, and function-word ratio. However, the models showed larger discrepancies for mean length of utterance, utterance count, connector use, and syntax-linked measures, particularly for childhood and adolescence. These results indicate that surface-level lexical and length-related patterns are more readily simulated than age-specific utterance organization and syntactic structure. Internal gradient analyses further showed stronger relative training-related changes in attention key and value projections across middle-to-late layers, suggesting that age-conditioned fine-tuning altered specific model components involved in mapping visual input to language output.
Topic Areas: Computational Approaches,