[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"detail-sidebar-cat-0-en-105":3,"doc-seo-273820-105":59,"doc-detail-273820-en":131},{"code":4,"msg":5,"data":6},0,"success",[7,13,18,23,28,33,38,43,48,51,55],{"id":8,"doc_module":4,"doc_module_name":9,"category_name":10,"show_sort_weight":11,"slug":12},1,"Document","Story & Novel",90,"story-novel",{"id":14,"doc_module":4,"doc_module_name":9,"category_name":15,"show_sort_weight":16,"slug":17},2,"Literature",80,"literature",{"id":19,"doc_module":4,"doc_module_name":9,"category_name":20,"show_sort_weight":21,"slug":22},4,"Exam",70,"exam",{"id":24,"doc_module":4,"doc_module_name":9,"category_name":25,"show_sort_weight":26,"slug":27},5,"Comic",60,"comic",{"id":29,"doc_module":4,"doc_module_name":9,"category_name":30,"show_sort_weight":31,"slug":32},6,"Technology",50,"technology",{"id":34,"doc_module":4,"doc_module_name":9,"category_name":35,"show_sort_weight":36,"slug":37},7,"Healthcare",40,"healthcare",{"id":39,"doc_module":4,"doc_module_name":9,"category_name":40,"show_sort_weight":41,"slug":42},8,"Research & Report",30,"research-report",{"id":44,"doc_module":4,"doc_module_name":9,"category_name":45,"show_sort_weight":46,"slug":47},9,"Religion & Spirituality",20,"religion-spirituality",{"id":46,"doc_module":4,"doc_module_name":9,"category_name":49,"show_sort_weight":46,"slug":50},"World Cup","world-cup",{"id":52,"doc_module":4,"doc_module_name":9,"category_name":53,"show_sort_weight":52,"slug":54},10,"Lifestyle","lifestyle",{"id":56,"doc_module":4,"doc_module_name":9,"category_name":57,"show_sort_weight":24,"slug":58},19,"General","general",{"code":4,"msg":60,"data":61},"ok",{"site_id":62,"language":63,"slug":64,"title":65,"keywords":66,"description":67,"schema_data":68,"social_meta":124,"head_meta":126,"extra_data":128,"updated_unix":130},105,"en","calibrating-criterion-revision-in-llm-agents-failure-modes-and-a-trace-anchored-protocol","Calibrating Criterion Revision in LLM Agents - Failure Modes and a Trace-Anchored Protocol","","Language-model agents can improve after failure or persist across episodes without revising what counts as success. The work targets criterion revision as an attribution problem: when an initial criterion accepts an outcome that violates a broader commitment, what observations justify claiming the system formed and persistently used a revised criterion? It proposes five non-compensatory evidence conditions, evaluates CMB-0.1 across cross-domain cases, and shows no model trial satisfies all conditions, motivating a trace-anchored CMB-0.4 protocol.",{"@graph":69,"@context":123},[70,84,106],{"@type":71,"itemListElement":72},"BreadcrumbList",[73,77,79,82],{"item":74,"name":75,"@type":76,"position":8},"https://docshare.wps.com","Home","ListItem",{"item":78,"name":9,"@type":76,"position":14},"https://docshare.wps.com/document/",{"item":80,"name":40,"@type":76,"position":81},"https://docshare.wps.com/document/research-report/",3,{"item":83,"name":65,"@type":76,"position":19},"https://docshare.wps.com/document/calibrating-criterion-revision-in-llm-agents-failure-modes-and-a-trace-anchored-protocol/273820/",{"url":83,"name":65,"@type":85,"image":86,"author":91,"headline":65,"publisher":94,"fileFormat":97,"inLanguage":63,"description":67,"dateModified":98,"datePublished":99,"encodingFormat":97,"isAccessibleForFree":100,"interactionStatistic":101},"DigitalDocument",{"url":87,"@type":88,"width":89,"height":90},"https://docshare.wps.com/thumbnails/calibrating-criterion-revision-in-llm-agents-failure-modes-and-a-trace-anchored-protocol/273820.png","ImageObject",300,407,{"name":92,"@type":93},"Genevieve","Person",{"url":74,"name":95,"@type":96},"DocShare","Organization","application/pdf","2026-10-17","2026-09-15",true,{"@type":102,"interactionType":103,"userInteractionCount":105},"InteractionCounter",{"@type":104},"ViewAction",11,{"@type":107,"mainEntity":108},"FAQPage",[109,115,119],{"name":110,"@type":111,"acceptedAnswer":112},"What problem does the paper focus on regarding LLM agents?","Question",{"text":113,"@type":114},"It focuses on criterion revision as an attribution problem: justifying when an agent should be said to have formed and persistently used a revised success criterion after violating a broader commitment.","Answer",{"name":116,"@type":111,"acceptedAnswer":117},"What are the five non-compensatory conditions proposed for criterion revision evidence?",{"text":118,"@type":114},"They are criterion-failure detection, a model-emitted proposal, new-episode transfer, intervention sensitivity on the claimed carrier, and preservation.",{"name":120,"@type":111,"acceptedAnswer":121},"What does the evaluation of CMB-0.1 conclude?",{"text":122,"@type":114},"No model trial satisfies all five conditions; the results indicate CMB-0.1 behaves as an instrument-calibration outcome rather than a reliable model ranking, motivating a successor protocol.","https://schema.org",{"og:url":83,"og:type":125,"og:title":65,"og:site_name":95,"og:description":67},"article",{"robots":127,"canonical":83},"index,follow",{"doc_id":129,"site_id":62},273820,1789472871,{"code":4,"msg":5,"data":132},{"doc_id":129,"user_id":133,"nickname":92,"user_avatar":134,"doc_module":4,"category_id":39,"category_name":40,"doc_title":65,"doc_description":67,"doc_content":135,"file_id":136,"file_url":137,"file_type":138,"file_size":139,"view_count":105,"is_deleted":4,"is_public":8,"is_downloadable":8,"audit_status":8,"page_count":140,"language":141,"language_code":63,"site_id":62,"html_lang":63,"table_of_contents":142,"faqs":143,"seo_title":144,"seo_description":67,"update_tm":130,"read_time":145,"preview_page_count":4},1374391974585,"https://ap-avatar.wpscdn.com/davatar_276721f389ce27ea32af1340a28f341c","arXiv : 2608 .20729v 1 [cs .AI] 2 1 Aug 2026  \nCalibrating Criterion Revision in LLM Agents: Failure Modes and a Trace-Anchored Protocol  \nGuodong Xu  \nQingdao Guodongxiansheng Network Technology Co., Ltd.  \n(Guˇodòng X¯ıansheng) [kzkz137806@gmail. com](kzkz137806@gmail. com)  \nAbstract  \nLanguage-model agents can improve after failure or carry text across episodes without revising what counts as success. We study the narrower attribution problem of criterion revision: when criterion K0 accepts an outcome violating a broader commitment B, what observations justify saying that the system formed and persistently used K1 ? We require five non-compensatory conditions: criterion-failure detection, a model-emitted proposal, new-episode transfer, intervention sensitivity on the claimed carrier, and preservation.  \nWe evaluate CMB-0 . 1 on twelve cross-domain cases and four arms: stateless inference, appendonly history, model-generated but harness-committed state, and evaluator-written oracle state. Seven mechanism fixtures yield 84 deterministic scorer trials; four local quantized artifacts yield 96 calls and 192 model–case–arm trials. No model trial satisfies all five conditions, but this zero does not establish general capability absence. Eleven calls remain invalid after one retry; several commitments disclose the target distinction; the harness performs commits; deletion reuses a stateless call; and conflict changes multiple factors.  \nQwen2.5-7B answers every transfer and preservation item without revision state, exposing  \nzero-state reconstruction.  \nThese failures make CMB-0.1 an instrument-calibration result rather than a model ranking.  \nWe derive a prospective, trace-anchored CMB-0.4 protocol requiring concealed transfer, explicit WRITE/NO-WRITE/ESCALATE actions, a separately logged policy-selected commit, matched interventions, repeated hidden items, and a frozen executable oracle. It is a successor design, not a completed confirmatory result. The paper contributes a measurement chain, an empirical diagnosis of its first implementation, and a more discriminating protocol for future tests of criterion revision.  \nKeywords: language-model agent systems; criterion revision; persistent memory; intervention; scorer validation; calibration study  \n1 Introduction  \nThe capability boundary of large language models is often drawn as a curve of task difficulty. Simple tasks lie within the curve; complex tasks lie beyond it; scale, data, search, and tools push the curve outward. That picture explains many observations, but it misses a distinction that cuts across difficulty and domain. Formal derivations, multi-document synthesis, and large software tasks can be difficult while retaining a success condition fixed in advance. Training, search, tool use, and error correction then have a direction. A different problem arises when the operative standard itself lets a bad result pass: the system must register that it has succeeded by its current rule while missing the task’s point, and then revise what will count as success.  \nWe call this narrower structure criterion metabolism. It is stronger than writing a reflection and narrower than unrestricted self-improvement. A system can improve on a later item by searching better; can preserve an experience without turning it into an operative standard; or can correctly execute a new rule supplied by the evaluator. Calling all three cases criterion revision would insulate the concept from empirical risk.  \n\n| Table 1: Evidence status of the instrument and its successor protocol. |  |  |\n| --- | --- | --- |\n| Version | Status in this manuscript | Permitted inference |\n| CMB-0.1 | Executed calibration: 84 | Instrument diagnosis and bounded observations about the four |\n| CMB-0.4 | deterministic scorer-fixture trials and 192 materialized local-model trials. Prospective protocol derived from the observed CMB-0 . 1 failure modes. No CMB-0.4 model-performance, write-selection, transfer, or\u003Cbr>causal","cbCaimNvmmcuV93L","https://ap.wps.com/l/cbCaimNvmmcuV93L","pdf",433904,18,"English","# Introduction\n## Criterion metabolism and criterion revision\n## Recent related systems\n## Contributions\n## Instrument calibration (CMB-0.1)\n## Trace-anchored successor protocol (CMB-0.4)","[{\"question\":\"What problem does the paper focus on regarding LLM agents?\",\"answer\":\"It focuses on criterion revision as an attribution problem: justifying when an agent should be said to have formed and persistently used a revised success criterion after violating a broader commitment.\"},{\"question\":\"What are the five non-compensatory conditions proposed for criterion revision evidence?\",\"answer\":\"They are criterion-failure detection, a model-emitted proposal, new-episode transfer, intervention sensitivity on the claimed carrier, and preservation.\"},{\"question\":\"What does the evaluation of CMB-0.1 conclude?\",\"answer\":\"No model trial satisfies all five conditions; the results indicate CMB-0.1 behaves as an instrument-calibration outcome rather than a reliable model ranking, motivating a successor protocol.\"}]","Calibrating Criterion Revision in LLM Agents - Failure Modes and a Trace-Anchored Protocol | PDF",45]