A new research framework flips the script on video search by letting the AI ask you questions instead of just guessing what you want.
Researchers introduced ADEPT, a training-free agent designed to close what the paper calls the "Intent-Query Gap" — the mismatch between what a user types and what they actually mean. Rather than running a single search and hoping for the best, ADEPT uses an entropy-based engine to decide, at each turn, whether to ask the user a clarifying question ("ASK") or quietly refine the search results on its own ("REFINE"). The system was tested on two video datasets and beat non-interactive, heuristic, and Video-LLM baselines across the board. No fine-tuning required.
The bottleneck in video retrieval has always been ambiguity. Most systems treat a text query as a complete thought and return results accordingly — which works fine when the user knows exactly what clip they want and can describe it cleanly. ADEPT's entropy-driven approach acknowledges that users often can't, and routes around that limitation without expensive retraining.
The benchmark numbers look strong, but video retrieval in controlled lab conditions and in production on a massive archive are very different problems — and this paper doesn't test the latter.