{"id":14337,"date":"2026-10-21T19:04:50","date_gmt":"2026-10-21T11:04:50","guid":{"rendered":"https:\/\/ivstskl.changan.com.cn\/?p=14337"},"modified":"2026-10-21T21:00:51","modified_gmt":"2026-10-21T13:00:51","slug":"proact-bench-%e4%b8%ad%e8%8b%b1%e5%8f%8c%e8%af%ad%e6%99%ba%e8%83%bd%e4%bd%93%e8%bf%90%e8%a1%8c%e6%97%b6%e5%ae%89%e5%85%a8%e6%95%b0%e6%8d%ae%e9%9b%86","status":"publish","type":"post","link":"https:\/\/ivstskl.changan.com.cn\/?p=14337","title":{"rendered":"PROACT-Bench \u4e2d\u82f1\u53cc\u8bed\u667a\u80fd\u4f53\u8fd0\u884c\u65f6\u5b89\u5168\u6570\u636e\u96c6"},"content":{"rendered":"\n
\n
PROACT-Bench: Bilingual Dataset for Runtime Agent Safety<\/h5>\n\n\n\n
\n
\u63cf\u8ff0\uff08Description\uff09<\/h5>\n\n\n\n
\n\n\n\n
    \u672c\u6570\u636e\u96c6\u57fa\u4e8e\u667a\u80fd\u4f53\u4ea4\u4e92\u8f68\u8ff9\u6784\u5efa\uff0c\u53ef\u7528\u4e8e\u667a\u80fd\u4f53\u5b89\u5168\u8bc4\u4f30\u3001\u9996\u6b21\u4e0d\u5b89\u5168\u8fb9\u754c\u68c0\u6d4b\u548c\u8fd0\u884c\u65f6\u5b89\u5168\u62a4\u680f\u5f00\u53d1\u3002\u6570\u636e\u96c6\u4ec5\u9650\u7528\u4e8e\u5b89\u5168\u8bc4\u4f30\u3001\u9632\u5fa1\u6027\u57fa\u51c6\u6d4b\u8bd5\u548c\u62a4\u680f\u5f00\u53d1\u3002\u4ee5\u4e0b\u662f\u672c\u6570\u636e\u96c6\u7684\u8be6\u7ec6\u4fe1\u606f\uff1a<\/pre>\n\n\n\n
    (1)\u6570\u636e\u6765\u6e90\uff1a\u539f\u59cb\u4ea4\u4e92\u8f68\u8ff9\u6765\u81ea R-Judge\u3001ToolSafety\u3001ATBench \u548c AgentAlign\uff0c\u901a\u8fc7\u9010\u6b65\u8f68\u8ff9\u5c55\u5f00\u3001\u63a8\u7406\u589e\u5f3a\u56e0\u679c\u4fee\u6b63\u548c\u6587\u5316\u611f\u77e5\u672c\u5730\u5316\uff0c\u5f62\u6210\u4e2d\u82f1\u53cc\u8bed\u524d\u7f00\u7ea7\u5b89\u5168\u76d1\u7763\u6570\u636e\u3002\u6bcf\u6761\u524d\u7f00\u72b6\u6001\u5bf9\u5e94\u667a\u80fd\u4f53\u4e0b\u4e00\u8f6e\u6a21\u578b\u63a8\u7406\u524d\u5df2\u83b7\u5f97\u7684\u7d2f\u8ba1\u4e0a\u4e0b\u6587\u3002<\/pre>\n\n\n\n
    (2)\u6570\u636e\u5b57\u6bb5\uff1a\u5305\u62ec\u5bf9\u8bdd\u8bb0\u5f55\uff08conversations\uff09\u3001\u89d2\u8272\uff08from\uff09\u3001\u6587\u672c\u5185\u5bb9\uff08value\uff09\u3002system \u89d2\u8272\u63d0\u4f9b\u5b89\u5168\u8bc4\u4f30\u6307\u4ee4\u53ca\u53ef\u7528\u5de5\u5177\u4fe1\u606f\uff1bhuman \u89d2\u8272\u5305\u542b\u4efb\u52a1\u76ee\u6807\u3001\u5386\u53f2\u8f68\u8ff9\u3001\u5f53\u524d\u4e0a\u4e0b\u6587\u53ca\u5df2\u6709\u6216\u79bb\u7ebf\u63a8\u65ad\u7684\u8f85\u52a9\u7406\u7531\uff1bgpt \u89d2\u8272\u63d0\u4f9b\u5b89\u5168\u6807\u7b7e\uff08label\uff09\u53ca\u4e0d\u5b89\u5168\u72b6\u6001\u7684\u62d2\u7edd\u7406\u7531\uff08reason\uff09\u3002\u5176\u4e2d\uff0clabel \u4e3a 0 \u8868\u793a\u53ef\u7ee7\u7eed\u63a8\u7406\uff0c\u4e3a 1 \u8868\u793a\u5e94\u963b\u6b62\u4e0b\u4e00\u8f6e\u63a8\u7406\u3002<\/pre>\n\n\n\n
    (3)\u6570\u636e\u7edf\u8ba1\uff1a\u5171\u8ba1 155,780 \u6761\u524d\u7f00\u7ea7\u72b6\u6001\uff0c\u5176\u4e2d\u82f1\u6587 77,890 \u6761\u3001\u4e2d\u6587 77,890 \u6761\uff0c\u6db5\u76d6\u516d\u7c7b\u5b89\u5168\u98ce\u9669\u53ca\u7f51\u9875\u901a\u4fe1\u3001\u7535\u5b50\u5546\u52a1\u3001\u8f6f\u4ef6\u5f00\u53d1\u7b49\u591a\u79cd\u5de5\u5177\u4f7f\u7528\u573a\u666f\u3002<\/pre>\n\n\n\n
This dataset is constructed from agent interaction trajectories for safety evaluation, first-unsafe boundary detection, and runtime guardrail development. Use is limited to safety evaluation, defensive benchmarking, and guardrail development. Detailed information is provided below:<\/pre>\n\n\n\n
(1) Source: Trajectories from R-Judge, ToolSafety, ATBench, and AgentAlign are processed through Progressive Trajectory Unrolling, Reasoning-Augmented Causal Rectification, and Culturally-Aware Data Localization. Each prefix-level state captures the accumulated context available before the next LLM inference.<\/pre>\n\n\n\n
(2) Fields: Conversation records (conversations), role (from), and text content (value). The system role provides evaluation instructions and available tool context; human contains the task, history, current context, and available or offline-inferred auxiliary rationales; gpt provides a safety label (label) and a refusal reason (reason) for unsafe states. Label 0 allows continued inference; label 1 indicates that the next inference should be blocked.<\/pre>\n\n\n\n
(3) Statistics: 155,780 prefix-level states in total, with 77,890 each in English and Chinese, covering six risk categories and diverse tool-use scenarios.<\/pre>\n<\/div>\n\n\n\n
\n
\u6570\u636e\u96c6\u4e0b\u8f7d\uff08Download\uff09<\/h5>\n\n\n\n
\n\n\n\n
    \u6570\u636e\u4ee5 ShareGPT \u98ce\u683c\u7684 JSON \u7ed3\u6784\u5b58\u50a8\u3002\u6570\u636e\u96c6\u91c7\u7528\u53d7\u63a7\u8bbf\u95ee\u65b9\u5f0f\u63d0\u4f9b\uff0c\u7533\u8bf7\u8005\u987b\u63d0\u4ea4\u8bbf\u95ee\u7533\u8bf7\u3001\u6838\u9a8c\u6240\u5c5e\u673a\u6784\u8eab\u4efd\u5e76\u7b7e\u7f72\u6570\u636e\u4f7f\u7528\u534f\u8bae\u540e\uff0c\u83b7\u53d6\u5168\u91cf\u6570\u636e\u3002<\/pre>\n\n\n\n
    \u8bf7\u9075\u5b88\u6570\u636e\u4f7f\u7528\u534f\u8bae\u53ca\u9002\u7528\u7684\u539f\u59cb\u6570\u636e\u8bb8\u53ef\uff0c\u7981\u6b62\u4efb\u4f55\u6076\u610f\u5229\u7528\u3002\u5982\u679c\u6570\u636e\u96c6\u6709\u52a9\u4e8e\u60a8\u7684\u7814\u7a76\u5de5\u4f5c\uff0c\u8bf7\u5728\u60a8\u51fa\u7248\u7269\u7684\u9002\u5f53\u4f4d\u7f6e\u5f15\u7528\u8be5\u6570\u636e\u96c6\u7684\u51fa\u5904\u3002\u53c2\u8003\u793a\u4f8b\u5982\u4e0b\uff1a<\/pre>\n\n\n\n
[1] Jia, D., Liu, W., Du, X. et al. PROACT-Agent: Progressive Runtime Oversight and Active Circuit-breaking for Real-Time Safety. arXiv preprint arXiv:2609.34415 (2026). https:\/\/arxiv.org\/abs\/2609.34415<\/pre>\n\n\n\n
Data are stored in a ShareGPT-style JSON structure. Access is controlled: applicants must submit an access request, verify their organizational affiliation, and digitally sign a data usage agreement before receiving the full dataset.<\/pre>\n\n\n\n
Please comply with the data usage agreement and applicable source-data terms. Malicious exploitation is prohibited. If this dataset contributes to your research, please cite its provenance in your publication. The reference is as follows:<\/pre>\n\n\n\n
[1] Jia, D., Liu, W., Du, X. et al. PROACT-Agent: Progressive Runtime Oversight and Active Circuit-breaking for Real-Time Safety. arXiv preprint arXiv:2609.34415 (2026). https:\/\/arxiv.org\/abs\/2609.34415<\/pre>\n\n\n\n
\n
\n
\n
\u6570\u636e\u7533\u8bf7(Dataset Access Request)<\/p>\n\n\n\n
\"\"<\/figure>\n\n\n\n
Apply<\/a><\/p>\n<\/div>\n<\/div>\n<\/div>\n<\/div>\n\n\n\n
\n
\u7c7b\u522b\uff08Categories\uff09<\/h5>\n\n\n\n
\n\n\n\n
    \u667a\u80fd\u4f53\u5b89\u5168\uff08Agent Safety\uff09
    \u4e2d\u82f1\u53cc\u8bed\uff08Bilingual English and Chinese\uff09
    \u8fd0\u884c\u65f6\u5b89\u5168\u8bc4\u4f30\uff08Runtime Safety Evaluation\uff09
    \u9996\u6b21\u4e0d\u5b89\u5168\u8fb9\u754c\u68c0\u6d4b\uff08First-unsafe Boundary Detection\uff09<\/pre>\n<\/div>\n\n\n\n
\n
\u53d1\u5e03\u5355\u4f4d\uff08Institutions\uff09<\/h5>\n\n\n\n
\n\n\n\n
    \u667a\u80fd\u6c7d\u8f66\u5b89\u5168\u6280\u672f\u5168\u56fd\u91cd\u70b9\u5b9e\u9a8c\u5ba4\uff08State Key Laboratory of Intelligent Vehicle Safety Technology (IVST)\uff09<\/pre>\n<\/div>\n\n\n\n
\n
\n
\u8bb8\u53ef\u534f\u8bae\uff08Licence\uff09<\/h5>\n<\/div>\n\n\n\n
\n\n\n\n
\n
PROACT-Bench \u6570\u636e\u4f7f\u7528\u534f\u8bae\uff08PROACT-Bench Data Usage Agreement\uff09<\/a><\/p>\n<\/div>\n<\/div>\n<\/div>\n\n\n\n
<\/p>\n","protected":false},"excerpt":{"rendered":"
PROACT-Bench: B […]<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"wp-custom-template-25","format":"standard","meta":{"footnotes":""},"categories":[67],"tags":[],"class_list":["post-14337","post","type-post","status-publish","format-standard","hentry","category-67"],"_links":{"self":[{"href":"https:\/\/ivstskl.changan.com.cn\/index.php?rest_route=\/wp\/v2\/posts\/14337","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/ivstskl.changan.com.cn\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/ivstskl.changan.com.cn\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/ivstskl.changan.com.cn\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/ivstskl.changan.com.cn\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=14337"}],"version-history":[{"count":10,"href":"https:\/\/ivstskl.changan.com.cn\/index.php?rest_route=\/wp\/v2\/posts\/14337\/revisions"}],"predecessor-version":[{"id":14363,"href":"https:\/\/ivstskl.changan.com.cn\/index.php?rest_route=\/wp\/v2\/posts\/14337\/revisions\/14363"}],"wp:attachment":[{"href":"https:\/\/ivstskl.changan.com.cn\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=14337"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/ivstskl.changan.com.cn\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=14337"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/ivstskl.changan.com.cn\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=14337"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}