Title: The Right Inference Strategy Is All You Need: Nearly Training-Free Domain-Wise Inference for EgoCross Challenge

Original: arXiv:2606.00829v1 Announce Type: new

Abstract: The EgoCross benchmark assesses multimodal large language models (MLLMs) on egocentric video question answering tasks characterized by significant domain shifts. Unlike typical daily-life scenarios, the test videos in this challenge originate from diverse and specialized contexts, including surgical procedures, industrial assembly lines, extreme sports, and cameras mounted on animals. In the source-limited track, participants are restricted to using a fixed base model, Qwen3-VL-4B, with access to only 20 training samples in the official task-specific support set. This constraint shifts the focus of the challenge from model scaling to the effective extraction of visual, temporal, and answer-selection cues by a constrained model.

Our central insight is that the frozen baseline model is not inherently incapable of handling these rare scenarios; instead, it struggles to transfer its pre-existing visual-language knowledge to the new task format without a suitable interface. To address this, we employ a domain-wise inference strategy that handles the four target domains individually. We design distinct input processing, prompting, and answer-mapping procedures tailored to the specific characteristics of each domain. These methods enhance the interpretability of rare egocentric scenes for the VLM by highlighting the most relevant cues for each context.

The resulting system is nearly training-free. Specifically, the base Qwen3-VL-4B model directly answers questions related to surgery and animals, while the XSports and industrial domains utilize only the official SFT checkpoint, which was trained for two epochs on the limited 20-sample dataset. During the final evaluation, this straightforward approach achieved an overall accuracy of 66.98%. This result demonstrates that careful, domain-aware inference can offset the limitations of the base model’s strength and recover a significant portion of the capability already embedded in the baseline.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC