Researchers from CERTH-ITI describe their TRECVID 2025 runs in a new arXiv paper, covering both ad-hoc video search (AVS) and video question answering (VQA). The work combines foundation-model-based retrieval with a training-free reranking step, and applies inference-time ensembling to a fine-tuned audio-visual MLLM for QA.
For AVS, the team first uses several vision-language models to encode the query—augmented by LLM-generated rephrasings—and candidate video shots, producing weighted similarity scores. A multimodal LLM then reranks the top-N shots by assessing semantic alignment with the original query, improving contextual matching without any model training.
For VQA, they fine-tune an audio-visual MLLM on the TRECVID training set. At inference, they sample multiple random frame subsets and generate responses with greedy decoding; for open-ended answer generation, a majority vote aggregates these responses, while for multiple-choice questions they mean-pool the logits over candidate answers. This combination yielded the top score across three metrics in the open-ended task and the second-highest in multiple-choice.
Because the paper is a single system description, there are no conflicting findings to compare. Its main contribution is showing that training-free reranking and simple frame-subset ensembling can push already fine-tuned multimodal models to state-of-the-art results on standard video benchmarks.