[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"item-2211":3},{"id":4,"title":5,"url":6,"summary":7,"summary_zh":8,"content":8,"source_name":9,"source_url":8,"published_at":10,"category":11,"cover_url":8,"hotness":12,"is_selected":13,"score":14,"score_detail":15,"sources":22,"tags":24,"view_count":30,"doi":8,"paper":31,"created_at":40},2211,"《视觉-语言模型对农业的了解比其表现出来的更多 基于评分标准的验证可弥补差距》","https:\u002F\u002Farcxiv.org\u002Fabs\u002F2609.09417","论文针对视觉-语言模型(VLM)在农作物病害、虫害、损伤、品质、物种识别等零样本任务中表现不佳的问题,构建包含116个数据集、834个类别、8324张图像的基准测试。线性探测显示VLM视觉编码器已能将与自监督DINOv3基线几乎同等可分离的农业特征编码,排除弱视觉表示为主要瓶颈。论文提出Probabilistic Pivot Tournament (PPT)验证器,通过基于固定诊断评分标准的K候选响应选择,使评判F1较未辅助基线近乎翻倍,在病害任务中将Gemma 4 E4B-it's病害F1提升至0.71,超过其自身上限0.60。",null,"arXiv 2609.09417 2026年9月","2026-09-10T00:00:00Z","论文",10,false,79,{"impact":16,"substance":17,"depth":16,"authority":18,"freshness":19,"relevant":20,"comment":21},18,22,13,8,1,"构建116数据集、8324图像的农业VLM基准，提出PPT评分标准验证器使病害F1近乎翻倍，方法新颖、数据扎实，对农业AI落地有实质参考价值。",[23],{"name":9,"url":6},[25,26,27,28,29],"智慧农业","农业人工智能","病害识别","多模态大模型","零样本学习",0,{"doi":8,"openalex_id":8,"authors":32,"venue":8,"cited_by_count":30,"oa_url":8,"card":33,"direction":37,"ingested_from":39},[],{"tldr":34,"method":35,"finding":36,"direction":37,"opportunity":38},"论文构建农业VLM基准，发现视觉编码器不弱，提出PPT验证器按评分标准选答案，大幅提升识别F1。","116数据集8324图基准，线性探测对比DINOv3，提出PPT概率枢轴锦标赛验","VLM农业知识被输出瓶颈掩盖，PPT使病害F1近翻倍，Gemma提升至0.71超其上限0.60。","农业人工智能与决策模型","可探索将评分标准验证器嵌入农业VLM推理链，或迁移到遥感、品质分级等更多农业零样本任务。","agent","2026-09-12T00:06:40.201845Z"]