[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"doc-detail-82681-en":3,"doc-seo-82681-105":29,"detail-sidebar-cat-0-en-105":90},{"code":4,"msg":5,"data":6},0,"success",{"doc_id":7,"user_id":8,"nickname":9,"user_avatar":10,"doc_module":4,"category_id":11,"category_name":12,"doc_title":13,"doc_description":14,"doc_content":15,"file_id":16,"file_url":17,"file_type":18,"file_size":19,"view_count":4,"is_deleted":4,"is_public":20,"is_downloadable":20,"audit_status":20,"page_count":21,"language":22,"language_code":23,"site_id":24,"html_lang":23,"table_of_contents":25,"faqs":26,"seo_title":13,"seo_description":14,"update_tm":27,"read_time":28},82681,3848291630094,"Emma Wilson","https://eur-avatar.wpscdn.com/davatar_085a072bc5b1113ac321206ff7593b45",8,"Research & Report","Attributing Structured-Output Gains in Function Calling: Interface Alignment versus Procedural Transfer","Structured-output benchmarks reward both task decisions and interface compliance, so prompt-induced function-calling gains require attribution before they can be interpreted as transferable skill. A four-layer gain-attribution protocol is introduced for prompt-prepended skill injection: canonicalized rescoring, format-only controls, repaired/balanced induction, and portability checks. Experiments on BFCL with API-Bank, MATH-500, and MultiHop-RAG show gains often stem from interface alignment rather than procedural transfer.","Attributing Structured-Output Gains in Function Calling: Interface Alignment versus Procedural Transfer  \nWanyi Chen 1 and Daoyuan Chen2 and Fang Kong 1 *  \n1 Soochow University  \n2Tongyi Lab, Alibaba Group  \narXiv :2607 .02595v 1 [ cs . SE] 1 Jul 2026  \nAbstract  \nStructured-output benchmarks reward both task decisions and interface compliance, so prompt-induced function-calling gains require attribution before they can be interpreted as transferable skill. We introduce a four-layer gain-attribution protocol for prompt-prepended skill injection, combining canonicalized rescoring, format-only controls, repaired/balanced induction, and portability checks. Applied to the Berkeley Function Calling Leaderboard (BFCL) and scoped with API-Bank, MATH- 500, and MultiHop-RAG, the protocol shows that several apparent gains are better attributed to interface alignment than to procedural transfer: format-only prompts match or exceed full skills in key BFCL cells, repaired/balanced induction removes the largest sub-frontier gains, and API-Bank target-native gains are matched within 0.5 percentage points (pp) by lengthmatched generic procedural prompts. These findings treat format compliance as a useful engineering capability while clarifying what a structured-output score certifies. We release BFCL-CANONICAL and recommend canonicalized metrics, balanced induction, and formatonly baselines for function-calling skill-gain attribution. Code and data are available at  \n[https://github. com/couragec/skill-injecti](https://github. com/couragec/skill-injecti)[on-attribution.](on-attribution.)  \n1 Introduction  \nPrompt-based interventions often improve LLM benchmark scores, but a score alone does not say what capability improved. In function-calling evaluations, the same prepended skill can change both task-solving policy and evaluator-facing output form. This makes gain attribution necessary for deciding whether an observed improvement reflects reusable procedure, interface alignment, or both.  \nTrajectory-derived skill injection is a representative case: agents distill trajectories into reusable  \n* Corresponding author.  \nskills (Wang et al., 2023 ; Ni et al., 2026), while broader tool-use systems connect language models to external actions and APIs (Yao et al., 2023 ; Schick et al., 2023 ; Qin et al., 2024 ; Patil et al., 2024) or automate agent-design search and guided trajectory synthesis (Hu et al., 2025 ; Xu et al., 2025) . These uses assume that a stronger model can write reusable procedures. In function-calling benchmarks, however, an injected skill may improve procedural transfer (tool and argument choice), interface alignment (the schema expected by the evaluator), or both. Distinguishing these mechanisms is essential if benchmark gains are to support claims about transferable agentic skill.  \nWe therefore ask a simple but under-examined question: when prompt-prepended skill text improves function-calling scores, what attribution does the evidence justify?  \nA score-only answer is insufficient for this distinction. If a skill says how to wrap a function call, the measured score may improve even when tool and argument selection do not. Conversely, a semantically correct call can be marked wrong when it uses a wrapper key or API syntax that the scorer does not accept. We therefore separate two questions that are often merged: whether the modelselected the right action, and whether it expressed that action in the evaluator’s preferred interface.  \nInconsistent evidence. Prior negative-transfer, metric-sensitivity, and contract-skill results (Liet al., 2026 ; Xu et al., 2026 ; Lu et al., 2026) motivate direct attribution: when benchmark scores improve, what mechanism is responsible?  \nFunction-calling skill-gain attribution. We propose a four-layer attribution protocol for promptprepended function-calling skill gains in structuredoutput audits based on the Berkeley Function Calling Leaderboard (BFCL) and API-Bank. First, canonicalized resc","cbCaib8N3tsqs0hq","https://ap.wps.com/l/cbCaib8N3tsqs0hq","pdf",15757844,1,18,"English","en",105,"# Introduction\n# Related Work\n## Skill extraction, reuse, and negative transfer","[{\"question\":\"Why is gain attribution necessary for function-calling benchmarks?\",\"answer\":\"Because a single prompt-prepended skill can improve both the model’s task-solving policy and the evaluator-facing output format. A score alone cannot identify whether the improvement comes from reusable procedure, interface alignment, or both.\"},{\"question\":\"What is the four-layer attribution protocol proposed in the paper?\",\"answer\":\"It uses canonicalized rescoring to separate wrapper-key aliasing from policy changes, format-only controls to test interface-contract effects, repaired/balanced induction to check dependence on source style or demonstrations, and cross-family/cross-benchmark portability checks.\"},{\"question\":\"What do the experiments suggest about where the biggest gains come from?\",\"answer\":\"In the BFCL/API-Bank-style setting, several large gains align better with interface alignment than robust procedural transfer. Format-only prompts can match or exceed full-skill gains, the largest gains disappear under repaired/balanced induction, and length-matched generic procedural prompts nearly reproduce extracted-skill gains.\"}]",1784182257,45,{"code":4,"msg":30,"data":31},"ok",{"site_id":24,"language":23,"slug":32,"title":13,"keywords":33,"description":14,"schema_data":34,"social_meta":85,"head_meta":87,"extra_data":89,"updated_unix":27},"attributing-structured-output-gains-in-function-calling-interface-alignment-versus-procedural-transfer","",{"@graph":35,"@context":84},[36,53,67],{"@type":37,"itemListElement":38},"BreadcrumbList",[39,43,47,50],{"item":40,"name":41,"@type":42,"position":20},"https://docshare.wps.com","Home","ListItem",{"item":44,"name":45,"@type":42,"position":46},"https://docshare.wps.com/document/","Document",2,{"item":48,"name":12,"@type":42,"position":49},"https://docshare.wps.com/document/research-report/",3,{"item":51,"name":13,"@type":42,"position":52},"https://docshare.wps.com/document/attributing-structured-output-gains-in-function-calling-interface-alignment-versus-procedural-transfer/82681/",4,{"url":51,"name":13,"@type":54,"author":55,"headline":13,"publisher":57,"fileFormat":60,"inLanguage":23,"description":14,"dateModified":61,"datePublished":61,"encodingFormat":60,"isAccessibleForFree":62,"interactionStatistic":63},"DigitalDocument",{"name":9,"@type":56},"Person",{"url":40,"name":58,"@type":59},"DocShare","Organization","application/pdf","2026-07-16",true,{"@type":64,"interactionType":65,"userInteractionCount":4},"InteractionCounter",{"@type":66},"ViewAction",{"@type":68,"mainEntity":69},"FAQPage",[70,76,80],{"name":71,"@type":72,"acceptedAnswer":73},"Why is gain attribution necessary for function-calling benchmarks?","Question",{"text":74,"@type":75},"Because a single prompt-prepended skill can improve both the model’s task-solving policy and the evaluator-facing output format. A score alone cannot identify whether the improvement comes from reusable procedure, interface alignment, or both.","Answer",{"name":77,"@type":72,"acceptedAnswer":78},"What is the four-layer attribution protocol proposed in the paper?",{"text":79,"@type":75},"It uses canonicalized rescoring to separate wrapper-key aliasing from policy changes, format-only controls to test interface-contract effects, repaired/balanced induction to check dependence on source style or demonstrations, and cross-family/cross-benchmark portability checks.",{"name":81,"@type":72,"acceptedAnswer":82},"What do the experiments suggest about where the biggest gains come from?",{"text":83,"@type":75},"In the BFCL/API-Bank-style setting, several large gains align better with interface alignment than robust procedural transfer. Format-only prompts can match or exceed full-skill gains, the largest gains disappear under repaired/balanced induction, and length-matched generic procedural prompts nearly reproduce extracted-skill gains.","https://schema.org",{"og:url":51,"og:type":86,"og:title":13,"og:site_name":58,"og:description":14},"article",{"robots":88,"canonical":51},"index,follow",{"doc_id":7,"site_id":24},{"code":4,"msg":5,"data":91},[92,96,100,104,109,114,119,122,127,130,134],{"id":20,"doc_module":4,"doc_module_name":45,"category_name":93,"show_sort_weight":94,"slug":95},"Story & Novel",90,"story-novel",{"id":46,"doc_module":4,"doc_module_name":45,"category_name":97,"show_sort_weight":98,"slug":99},"Literature",80,"literature",{"id":52,"doc_module":4,"doc_module_name":45,"category_name":101,"show_sort_weight":102,"slug":103},"Exam",70,"exam",{"id":105,"doc_module":4,"doc_module_name":45,"category_name":106,"show_sort_weight":107,"slug":108},5,"Comic",60,"comic",{"id":110,"doc_module":4,"doc_module_name":45,"category_name":111,"show_sort_weight":112,"slug":113},6,"Technology",50,"technology",{"id":115,"doc_module":4,"doc_module_name":45,"category_name":116,"show_sort_weight":117,"slug":118},7,"Healthcare",40,"healthcare",{"id":11,"doc_module":4,"doc_module_name":45,"category_name":12,"show_sort_weight":120,"slug":121},30,"research-report",{"id":123,"doc_module":4,"doc_module_name":45,"category_name":124,"show_sort_weight":125,"slug":126},9,"Religion & Spirituality",20,"religion-spirituality",{"id":125,"doc_module":4,"doc_module_name":45,"category_name":128,"show_sort_weight":125,"slug":129},"World Cup","world-cup",{"id":131,"doc_module":4,"doc_module_name":45,"category_name":132,"show_sort_weight":131,"slug":133},10,"Lifestyle","lifestyle",{"id":135,"doc_module":4,"doc_module_name":45,"category_name":136,"show_sort_weight":105,"slug":137},19,"General","general"]