The authors constructed 327 handwritten queries grouped into 86 scenes.
Omni-modal models fail at resolving three-way ambiguity, benchmark finds
The TRIAD benchmark tests whether systems can jointly disambiguate text, image, and audio; even top models trail humans by 20 points and collapse when answer options are removed.
Chinese Tech
Zhaolu Kang · Yidi Wang · Yile Li · Siqi Zeng · Yingjie He · Richeng Xuan · +1 more
Peking University · Hunyuan, Tencent
Research Digest··3 min read
The authors introduce TRIAD, a benchmark of 327 text–image–audio queries each designed so that the correct answer is unique only when all three modalities are present.
§