Researchers have identified a significant readout failure in Video Large Language Models (LLMs) that hinders their ability to accurately count events, such as objects crossing a screen. While the models possess the necessary information, they struggle to process and output it correctly. The study proposes two readout-directed repairs, episode-localization supervision and a discrete event-symbol side channel, which improve same-object counting performance from 46.7% to 65.0% on a challenging benchmark. Further analysis revealed data integrity issues and highlighted that while the core reader capacity is present, the deployed output head remains a bottleneck. AI
IMPACT Identifies a specific failure mode in video LLMs, potentially guiding future model development and evaluation for improved event comprehension.
RANK_REASON Academic paper detailing a specific technical finding and proposed solution for LLM limitations. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
