-
Notifications
You must be signed in to change notification settings - Fork 0
Description
在学校里做实验的时候,老师如何确定我们做了实验并且达到了预期效果呢?——最常见的做法是让学生写一份实验报告交上来。
现在,AI智能体拿到一个任务以后如何检验执行的效果有没有达到预期呢?我们也可以让AI在执行任务的同时主动提交一份证据链报告,边做边收集任务完成的证据,自我检查是否符合预期,不符合就继续做。
在LLM/VLM驱动的智能体(Agent)的强化学习(RL)研究中,一直面临一个巨大的挑战:
你交给智能体一个任务,它干完了,但你不知道完成度如何。
为了确认它是否真的准确完成了任务,我们不得不建立庞大的“监督系统”来复核它的每一步操作。这种“被动验证”往往需要:
这两种常见的先完成任务(task completion)再校验轨迹(outcome verification)的机制有以下缺点:
针对以上问题,我们提出了一种简单的RL训练方法,让智能体自己成为“质检员”,在尽可能减少校验器(Verifier)审核压力的同时,让智能体学会主动分解子目标并且留痕存证。
什么是SmartSnap?
SmartSnap的核心思想是将GUI智能体从“被动的执行者”转变为“主动的自证者”。
简单来说,智能体在完成任务的同时,还会主动收集、筛选并提交一份“证据快照集”。
这份证据就像是任务的“结项报告”,让验证者只需看一眼快照,就能确认任务是否成功。
三大核心突破:从“执行”到“自证”
- 角色升级:双重使命的“自证代理”
传统的智能体只负责“做(Execute)”,而SmartSnap提出了“自证智能体”(Self-Verifying Agent),赋予了它“自我验证(Verify)”的第二使命。
它在操作过程中会像人类一样思考:“为了证明我已经改好了设置,我需要把对开关状态截图并作为证据提交。”
- “3C原则”:高效率的证据美学
为了避免给验证者造成信息过载,SmartSnap提出了证据策展的3C原则
- 强化学习驱动:GRPO+内在奖励反馈
我们利用GRPO算法对智能体进行了训练。通过精心设计的奖励机制(Intrinsic Reward Shaping),引导智能体在保证任务成功率的同时,不断提升证据的质量,尽可能减少奖励黑客行为(reward hacking)。
战绩显赫:小模型也不错
wWw.7Ttq.coM/jiangshan/
wWw.7Ttq.coM/zhoushan/
wWw.7Ttq.coM/dinghai/
wWw.7Ttq.coM/putuo/
wWw.7Ttq.coM/daishan/
wWw.7Ttq.coM/chengsi/
wWw.7Ttq.coM/taizhoushi/
wWw.7Ttq.coM/jiaojiang/
wWw.7Ttq.coM/huangyan/
wWw.7Ttq.coM/luqiao/
wWw.7Ttq.coM/yuhuan/
wWw.7Ttq.coM/sanmen/
wWw.7Ttq.coM/tiantai/
wWw.7Ttq.coM/xianji/
wWw.7Ttq.coM/wenling/
wWw.7Ttq.coM/linhai/
wWw.7Ttq.coM/lishui/
wWw.7Ttq.coM/qingtian/
wWw.7Ttq.coM/jinyun/
wWw.7Ttq.coM/suichang/
wWw.7Ttq.coM/songyang/
wWw.7Ttq.coM/yunhe/
wWw.7Ttq.coM/lishuiqingyuan/
wWw.7Ttq.coM/lishuijingning/
wWw.7Ttq.coM/longquan/
wWw.7Ttq.coM/hefei/
wWw.7Ttq.coM/yaohai/
wWw.7Ttq.coM/luyang/
wWw.7Ttq.coM/shushan/
wWw.7Ttq.coM/baohe/
wWw.7Ttq.coM/changfeng/
wWw.7Ttq.coM/feidong/
wWw.7Ttq.coM/feixi/
wWw.7Ttq.coM/wuhu/
wWw.7Ttq.coM/jinghu/
wWw.7Ttq.coM/yijiang/
wWw.7Ttq.coM/sanshan/
wWw.7Ttq.coM/wuhuxian/
wWw.7Ttq.coM/fanchang/
wWw.7Ttq.coM/nanling/
wWw.7Ttq.coM/bangbu/
wWw.7Ttq.coM/longzihu/
wWw.7Ttq.coM/bangshan/
wWw.7Ttq.coM/yuhui/
wWw.7Ttq.coM/huaishang/
wWw.7Ttq.coM/huaiyuan/
wWw.7Ttq.coM/wuhe/
wWw.7Ttq.coM/guzhen/
wWw.7Ttq.coM/huainan/
wWw.7Ttq.coM/tianjiaan/
wWw.7Ttq.coM/xiejiaji/
wWw.7Ttq.coM/bagongshan/
wWw.7Ttq.coM/panji/
wWw.7Ttq.coM/fengtai/
wWw.7Ttq.coM/maanshan/
wWw.7Ttq.coM/huashan/
wWw.7Ttq.coM/dangtu/
wWw.7Ttq.coM/huaibei/
wWw.7Ttq.coM/duji/
wWw.7Ttq.coM/lieshan/
wWw.7Ttq.coM/huaibeisuixi/
wWw.7Ttq.coM/tongling/
wWw.7Ttq.coM/jiaoqu/
wWw.7Ttq.coM/anqing/
wWw.7Ttq.coM/yixiu/
wWw.7Ttq.coM/huaining/
wWw.7Ttq.coM/congyang/
wWw.7Ttq.coM/anqingqianshan/
wWw.7Ttq.coM/taihu/
wWw.7Ttq.coM/susong/
wWw.7Ttq.coM/wangjiang/
wWw.7Ttq.coM/anqingyuexi/
wWw.7Ttq.coM/huangshan/
wWw.7Ttq.coM/tunxi/
wWw.7Ttq.coM/huangshanqu/
wWw.7Ttq.coM/xiexian/
wWw.7Ttq.coM/xiuning/
wWw.7Ttq.coM/yixian/
wWw.7Ttq.coM/qimen/
wWw.7Ttq.coM/chuzhou/
wWw.7Ttq.coM/langya/
wWw.7Ttq.coM/nanqiao/
wWw.7Ttq.coM/laian/
wWw.7Ttq.coM/quanjiao/
wWw.7Ttq.coM/dingyuan/
wWw.7Ttq.coM/fengyang/
wWw.7Ttq.coM/tianchang/
wWw.7Ttq.coM/mingguang/
wWw.7Ttq.coM/fuyang/
wWw.7Ttq.coM/yingzhou/
wWw.7Ttq.coM/yingdong/
wWw.7Ttq.coM/yingquan/
wWw.7Ttq.coM/linquan/
wWw.7Ttq.coM/fuyangtaihe/
wWw.7Ttq.coM/funan/
wWw.7Ttq.coM/yingshang/
wWw.7Ttq.coM/jieshou/
wWw.7Ttq.coM/suzhoushi/
wWw.7Ttq.coM/yongqiao/
wWw.7Ttq.coM/dangshan/
wWw.7Ttq.coM/xiaoxian/
wWw.7Ttq.coM/lingbi/
wWw.7Ttq.coM/sixian/
wWw.7Ttq.coM/chaohu/
wWw.7Ttq.coM/lujiang/
wWw.7Ttq.coM/chaohuwuwei/
wWw.7Ttq.coM/hanshan/
wWw.7Ttq.coM/hexian/
wWw.7Ttq.coM/luan/
wWw.7Ttq.coM/yuan/
wWw.7Ttq.coM/shouxian/
wWw.7Ttq.coM/huoqiu/
wWw.7Ttq.coM/shucheng/
wWw.7Ttq.coM/jinzhai/
wWw.7Ttq.coM/huoshan/
wWw.7Ttq.coM/bozhou/
wWw.7Ttq.coM/qiaocheng/
wWw.7Ttq.coM/woyang/
wWw.7Ttq.coM/mengcheng/
wWw.7Ttq.coM/lixin/
wWw.7Ttq.coM/chizhou/
wWw.7Ttq.coM/guichi/
wWw.7Ttq.coM/dongzhi/
wWw.7Ttq.coM/shitai/
wWw.7Ttq.coM/chizhouqingyang/
wWw.7Ttq.coM/fuzhou/
wWw.7Ttq.coM/mawei/
wWw.7Ttq.coM/minhou/
wWw.7Ttq.coM/fuzhoulianjiang/
wWw.7Ttq.coM/luoyuan/
wWw.7Ttq.coM/minqing/
wWw.7Ttq.coM/yongtai/
wWw.7Ttq.coM/pingtan/
wWw.7Ttq.coM/fuqing/
wWw.7Ttq.coM/changleshi/
wWw.7Ttq.coM/xiamen/
wWw.7Ttq.coM/siming/
wWw.7Ttq.coM/haicang/
wWw.7Ttq.coM/huli/
wWw.7Ttq.coM/jimei/
wWw.7Ttq.coM/tongan/
wWw.7Ttq.coM/xiangan/
wWw.7Ttq.coM/putian/
wWw.7Ttq.coM/chengxiang/
wWw.7Ttq.coM/hanjiang/
wWw.7Ttq.coM/putianlicheng/
wWw.7Ttq.coM/xiuyu/
wWw.7Ttq.coM/xianyou/
wWw.7Ttq.coM/sanming/
wWw.7Ttq.coM/meilie/
wWw.7Ttq.coM/mingxi/
wWw.7Ttq.coM/qingliu/
wWw.7Ttq.coM/ninghua/
wWw.7Ttq.coM/datian/
wWw.7Ttq.coM/youxi/
wWw.7Ttq.coM/shaxian/
wWw.7Ttq.coM/jiangle/
wWw.7Ttq.coM/taining/
wWw.7Ttq.coM/jianning/
wWw.7Ttq.coM/yongan/
wWw.7Ttq.coM/quanzhou/
wWw.7Ttq.coM/fengze/
wWw.7Ttq.coM/quangang/
wWw.7Ttq.coM/huian/
wWw.7Ttq.coM/anxixian/
wWw.7Ttq.coM/yongchun/
wWw.7Ttq.coM/dehua/
wWw.7Ttq.coM/shishi/
wWw.7Ttq.coM/jinjiang/
wWw.7Ttq.coM/quanzhounanan/
wWw.7Ttq.coM/zhangzhou/
wWw.7Ttq.coM/yunxiao/
wWw.7Ttq.coM/zhangpu/
wWw.7Ttq.coM/zhaoan/
wWw.7Ttq.coM/changtai/
wWw.7Ttq.coM/zhangzhoudongshan/
wWw.7Ttq.coM/nanjingxian/
wWw.7Ttq.coM/pinghe/
wWw.7Ttq.coM/huaan/
wWw.7Ttq.coM/longhai/
wWw.7Ttq.coM/nanping/
wWw.7Ttq.coM/yanping/
wWw.7Ttq.coM/shunchang/
wWw.7Ttq.coM/nanpingpucheng/
wWw.7Ttq.coM/guangze/
wWw.7Ttq.coM/songxi/
wWw.7Ttq.coM/zhenghe/
wWw.7Ttq.coM/shaowu/
wWw.7Ttq.coM/wuyishan/
wWw.7Ttq.coM/jianou/
wWw.7Ttq.coM/jianyang/
wWw.7Ttq.coM/longyan/
wWw.7Ttq.coM/xinluo/
wWw.7Ttq.coM/changting/
wWw.7Ttq.coM/longyanyongding/
wWw.7Ttq.coM/shanghang/
wWw.7Ttq.coM/wuping/
wWw.7Ttq.coM/liancheng/
wWw.7Ttq.coM/zhangping/
wWw.7Ttq.coM/ningde/
wWw.7Ttq.coM/jiaocheng/
wWw.7Ttq.coM/xiapu/
wWw.7Ttq.coM/gutian/
wWw.7Ttq.coM/ningdepingnan/
wWw.7Ttq.coM/shouning/
wWw.7Ttq.coM/zhouning/
wWw.7Ttq.coM/zherong/
wWw.7Ttq.coM/fuan/
wWw.7Ttq.coM/fuding/
wWw.7Ttq.coM/nanchang/
wWw.7Ttq.coM/donghu/
wWw.7Ttq.coM/xihu/
wWw.7Ttq.coM/qingyunpu/
wWw.7Ttq.coM/wanli/
wWw.7Ttq.coM/qingshanhu/
wWw.7Ttq.coM/nanchangxian/
wWw.7Ttq.coM/xinjian/
wWw.7Ttq.coM/anyi/
wWw.7Ttq.coM/jinxian/
wWw.7Ttq.coM/jingdezhen/
wWw.7Ttq.coM/fuliang/
wWw.7Ttq.coM/leping/
wWw.7Ttq.coM/pingxiang/
wWw.7Ttq.coM/anyuan/
wWw.7Ttq.coM/xiangdong/
wWw.7Ttq.coM/lianhua/
wWw.7Ttq.coM/shangli/
wWw.7Ttq.coM/pingxiangluxi/
wWw.7Ttq.coM/jiujiang/
wWw.7Ttq.coM/jiujianglushan/
wWw.7Ttq.coM/wuning/
wWw.7Ttq.coM/xiushui/
wWw.7Ttq.coM/yongxiu/
wWw.7Ttq.coM/dean/
wWw.7Ttq.coM/xingzi/
wWw.7Ttq.coM/duchang/
wWw.7Ttq.coM/hukou/
wWw.7Ttq.coM/pengze/
wWw.7Ttq.coM/ruichang/
wWw.7Ttq.coM/gongqingcheng/
wWw.7Ttq.coM/xinyu/
wWw.7Ttq.coM/yushui/
wWw.7Ttq.coM/fenyi/
wWw.7Ttq.coM/yingtan/
wWw.7Ttq.coM/yuehu/
wWw.7Ttq.coM/yujiang/
wWw.7Ttq.coM/guixi/
wWw.7Ttq.coM/ganzhou/
wWw.7Ttq.coM/zhanggong/
wWw.7Ttq.coM/ganxian/
wWw.7Ttq.coM/ganzhouxinfeng/
wWw.7Ttq.coM/dayu/
wWw.7Ttq.coM/shangyou/
wWw.7Ttq.coM/chongyi/
wWw.7Ttq.coM/anyuanxian/
wWw.7Ttq.coM/ganzhoulongnan/
wWw.7Ttq.coM/dingnan/
wWw.7Ttq.coM/quannan/
wWw.7Ttq.coM/ningdu/
wWw.7Ttq.coM/yudu/
wWw.7Ttq.coM/xingguo/
wWw.7Ttq.coM/huichang/
wWw.7Ttq.coM/xunwu/
wWw.7Ttq.coM/shicheng/
wWw.7Ttq.coM/ruijin/
wWw.7Ttq.coM/nankang/
wWw.7Ttq.coM/jian/
wWw.7Ttq.coM/jizhou/
wWw.7Ttq.coM/jianxian/
wWw.7Ttq.coM/jishui/
wWw.7Ttq.coM/xiajiang/
wWw.7Ttq.coM/jianxingan/
wWw.7Ttq.coM/yongfeng/
wWw.7Ttq.coM/taihe/
wWw.7Ttq.coM/suichuan/
wWw.7Ttq.coM/wanan/
wWw.7Ttq.coM/anfu/
wWw.7Ttq.coM/jianyongxin/
wWw.7Ttq.coM/jinggangshan/
wWw.7Ttq.coM/yichun/
wWw.7Ttq.coM/fengxin/
wWw.7Ttq.coM/wanzai/
wWw.7Ttq.coM/shanggao/
wWw.7Ttq.coM/yifeng/
wWw.7Ttq.coM/jingan/
wWw.7Ttq.coM/tonggu/
wWw.7Ttq.coM/zhangshu/
wWw.7Ttq.coM/gaoan/
wWw.7Ttq.coM/fuzhoushi/
wWw.7Ttq.coM/linchuan/
wWw.7Ttq.coM/nancheng/
wWw.7Ttq.coM/fuzhoulichuan/
wWw.7Ttq.coM/nanfeng/
wWw.7Ttq.coM/chongren/
wWw.7Ttq.coM/lean/
wWw.7Ttq.coM/yihuang/
wWw.7Ttq.coM/jinxi/
wWw.7Ttq.coM/zixi/
wWw.7Ttq.coM/dongxiang/
wWw.7Ttq.coM/guangchang/
wWw.7Ttq.coM/shangrao/
wWw.7Ttq.coM/shangraoxian/
wWw.7Ttq.coM/guangfeng/
wWw.7Ttq.coM/yushan/
wWw.7Ttq.coM/qianshan/
wWw.7Ttq.coM/hengfeng/
wWw.7Ttq.coM/shangraoyiyang/
wWw.7Ttq.coM/yugan/
wWw.7Ttq.coM/poyang/
wWw.7Ttq.coM/wannian/
wWw.7Ttq.coM/mouyuan/
wWw.7Ttq.coM/dexing/
wWw.7Ttq.coM/jinan/
wWw.7Ttq.coM/lixia/
wWw.7Ttq.coM/huaiyin/
wWw.7Ttq.coM/tianqiao/
wWw.7Ttq.coM/licheng/
wWw.7Ttq.coM/changqing/
wWw.7Ttq.coM/pingyin/
wWw.7Ttq.coM/jiyang/
wWw.7Ttq.coM/shanghe/
wWw.7Ttq.coM/zhangqiu/
wWw.7Ttq.coM/qingdao/
wWw.7Ttq.coM/shinan/
wWw.7Ttq.coM/shibei/
wWw.7Ttq.coM/huangdao/
wWw.7Ttq.coM/laoshan/
wWw.7Ttq.coM/licang/
wWw.7Ttq.coM/chengyang/
wWw.7Ttq.coM/jiaozhou/
wWw.7Ttq.coM/jimo/
wWw.7Ttq.coM/pingdu/
wWw.7Ttq.coM/jiaonan/
wWw.7Ttq.coM/laixi/
wWw.7Ttq.coM/zibo/
wWw.7Ttq.coM/zichuan/
wWw.7Ttq.coM/zhangdian/
wWw.7Ttq.coM/boshan/
wWw.7Ttq.coM/linzi/
wWw.7Ttq.coM/zhoucun/
wWw.7Ttq.coM/huantai/
wWw.7Ttq.coM/gaoqing/
wWw.7Ttq.coM/yiyuan/
wWw.7Ttq.coM/zaozhuang/
wWw.7Ttq.coM/xuecheng/
wWw.7Ttq.coM/zaozhuangyicheng/
wWw.7Ttq.coM/taierzhuang/
wWw.7Ttq.coM/shanting/
wWw.7Ttq.coM/tengzhou/
wWw.7Ttq.coM/dongying/
wWw.7Ttq.coM/dongyinghekou/
wWw.7Ttq.coM/kenli/
wWw.7Ttq.coM/lijin/
wWw.7Ttq.coM/guangrao/
wWw.7Ttq.coM/yantai/
wWw.7Ttq.coM/zhifu/
wWw.7Ttq.coM/yantaifushan/
wWw.7Ttq.coM/mouping/
wWw.7Ttq.coM/laishan/
wWw.7Ttq.coM/changdao/
wWw.7Ttq.coM/longkou/
wWw.7Ttq.coM/laiyang/
wWw.7Ttq.coM/laizhou/
wWw.7Ttq.coM/penglai/
wWw.7Ttq.coM/zhaoyuan/
wWw.7Ttq.coM/qixia/
wWw.7Ttq.coM/haiyang/
wWw.7Ttq.coM/weifang/
wWw.7Ttq.coM/hanting/
wWw.7Ttq.coM/fangzi/
wWw.7Ttq.coM/kuiwen/
wWw.7Ttq.coM/linchun/
wWw.7Ttq.coM/changle/
wWw.7Ttq.coM/qingzhou/
wWw.7Ttq.coM/shouguang/
wWw.7Ttq.coM/anqiu/
wWw.7Ttq.coM/gaomi/
wWw.7Ttq.coM/changyi/
wWw.7Ttq.coM/jining/
wWw.7Ttq.coM/rencheng/
wWw.7Ttq.coM/jiningweishan/
wWw.7Ttq.coM/yutai/
wWw.7Ttq.coM/jinxiang/
wWw.7Ttq.coM/jiaxiang/
wWw.7Ttq.coM/wenshang/
wWw.7Ttq.coM/sishui/
wWw.7Ttq.coM/jiningliangshan/
wWw.7Ttq.coM/qufu/
wWw.7Ttq.coM/yanzhou/
wWw.7Ttq.coM/taian/
wWw.7Ttq.coM/daiyue/
wWw.7Ttq.coM/ningyang/
wWw.7Ttq.coM/dongping/
wWw.7Ttq.coM/xintai/
wWw.7Ttq.coM/weihai/
wWw.7Ttq.coM/huancui/