Evaluating Large Language Models for Movie Prediction with Particular Scenes

View at publisher Summarise with VIRAT
Authors: B K.J.S. Rodrigo, Saleem Adeeba
Cited by: 0
References: 19

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

Movie discovery using traditional metadata-based systems has its limitations when the user is only able to recall particular scenes of a movie instead of movie titles or genres. Large Language Models (LLMs) have emergent semantic reasoning abilities which may address this “Tip-of-the-Tongue” problem, but how well they perform in scene-to-movie prediction is systematically unexamined. This research performs a thorough benchmark assessment of four leading LLMs, namely, GPT-5, Perplexity Sonar Pro, Gemini 2.5 Flash and DeepSeek V3.2 Experimental on scene-to-movie prediction tasks on a curated 30 scene description queries with both unambiguous and ambiguous scenarios. Employing consistent prompt engineering, standard parameter settings, and ranking-based evaluation metrics, the research results show that LLMs are significantly more capable than conventional text retrieval baselines as semantic search engines. GPT-5 and Perplexity Sonar Pro had the best performance - Mean Reciprocal Rank (MRR) 0.9667 & Precision@1 96.67, Gemini 2.5 Flash and DeepSeek V3.2 had competitive but lower performance metrics. Qualitative error analysis found some common failure modes such as temporal knowledge cutoffs for recent releases (those post-2023), confusing gender and character attributes, and misinterpreting the context. The research sets up evaluation protocols, and proves the potential of LLM-based scene retrieval systems, and confirms the move to transformative movie discovery paradigms.

Fields of Study & Subjects

Information Retrieval and Search Behavior Topic Modeling Recommender Systems and Techniques

Publication details

Journal / source
International Conference on Advanced Research in Computing (ICARC)
Pages
1–6
Publication date
18 February 2026
Language
EN