The Right Inference Strategy Is All You Need: Nearly Training-Free Domain-Wise Inference for EgoCross Challenge Team Name: WFJ-KnowinEnvision

Leyi Wu, Yifan Zhao, Jinjie Zhang, Yinchuan Li, Yingcong Chen · 2026

Read The Right Inference Strategy Is All You Need: Nearly Training-Free Domain-Wise Inference for EgoCross Challenge Team Name: WFJ-KnowinEnvision by Leyi Wu, Yifan Zhao, Jinjie Zhang, Yinchuan Li, Yingcong Chen (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2606-00829.txt.

Description

Wu and co-authors win ground on a video challenge with almost no training in this 2026 paper. EgoCross tests models on questions about first-person video. The scenes are rare ones: surgery, factory assembly, extreme sports, and animal-mounted cameras. In the track they enter, the base model is fixed and only 20 training samples are given.

Their claim is that the frozen model knows more than its first scores show. It fails to carry its skill into the new format without the right interface. So they craft a strategy per domain. Inputs, prompts, and answer mapping shift to fit each scene type. Surgery and animal questions run on the base model alone. Sports and industry use a checkpoint fine-tuned for two epochs on the 20 samples. The system scores 66.98 percent overall, they report. The lesson: fit the interface to the domain before scaling the model.

Catalog record

CREATORLeyi Wu, Yifan Zhao, Jinjie Zhang, Yinchuan Li, Yingcong Chen
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE25 KB
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2606.00829

Browse the full library · About the Tower of Babel