[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"doc-seo-273812-105":3,"detail-sidebar-cat-0-en-105":81,"doc-detail-273812-en":130},{"code":4,"msg":5,"data":6},0,"ok",{"site_id":7,"language":8,"slug":9,"title":10,"keywords":11,"description":12,"schema_data":13,"social_meta":74,"head_meta":76,"extra_data":78,"updated_unix":80},105,"en","why2speak-faithful-reasoning-for-abstaining-action-policies","Why2Speak - Faithful Reasoning for Abstaining Action Policies","","Many agentic systems must choose between acting and abstaining, making faithful reasoning essential for oversight: an explanation is useful only when it matches the computation that produced the action, including decisions to do nothing. The work studies intervention timing in multi-party conversation, where an assistant decides whether to speak now or remain silent. Experiments with Qwen3-8B compare direct, reasoning, supervised fine-tuning, and reinforcement-learning policies and reveal a capability–auditability tradeoff and limits of standard faithfulness metrics.",{"@graph":14,"@context":73},[15,34,56],{"@type":16,"itemListElement":17},"BreadcrumbList",[18,23,27,31],{"item":19,"name":20,"@type":21,"position":22},"https://docshare.wps.com","Home","ListItem",1,{"item":24,"name":25,"@type":21,"position":26},"https://docshare.wps.com/document/","Document",2,{"item":28,"name":29,"@type":21,"position":30},"https://docshare.wps.com/document/research-report/","Research & Report",3,{"item":32,"name":10,"@type":21,"position":33},"https://docshare.wps.com/document/why2speak-faithful-reasoning-for-abstaining-action-policies/273812/",4,{"url":32,"name":10,"@type":35,"image":36,"author":41,"headline":10,"publisher":44,"fileFormat":47,"inLanguage":8,"description":12,"dateModified":48,"datePublished":49,"encodingFormat":47,"isAccessibleForFree":50,"interactionStatistic":51},"DigitalDocument",{"url":37,"@type":38,"width":39,"height":40},"https://docshare.wps.com/thumbnails/why2speak-faithful-reasoning-for-abstaining-action-policies/273812.png","ImageObject",300,407,{"name":42,"@type":43},"Genevieve","Person",{"url":19,"name":45,"@type":46},"DocShare","Organization","application/pdf","2026-10-26","2026-09-15",true,{"@type":52,"interactionType":53,"userInteractionCount":55},"InteractionCounter",{"@type":54},"ViewAction",10,{"@type":57,"mainEntity":58},"FAQPage",[59,65,69],{"name":60,"@type":61,"acceptedAnswer":62},"What does “faithful reasoning” mean in abstaining action policies?","Question",{"text":63,"@type":64},"An explanation is faithful only if it reflects the computation that actually influenced the chosen action, including when the correct behavior is to remain silent.","Answer",{"name":66,"@type":61,"acceptedAnswer":67},"Why is intervention timing in multi-party conversation used as a testbed?",{"text":68,"@type":64},"It forces an assistant to decide whether to intervene now or not, exposing issues like asymmetric action costs and severe class imbalance that are often missing in question-answering faithfulness benchmarks.",{"name":70,"@type":61,"acceptedAnswer":71},"What overall tradeoff does the study find when exposing reasoning?",{"text":72,"@type":64},"Stronger direct policies improve decision quality but provide no inspectable reasoning, while reasoning policies offer an auditable trace at the cost of lower performance, especially reduced recall of true intervention opportunities.","https://schema.org",{"og:url":32,"og:type":75,"og:title":10,"og:site_name":45,"og:description":12},"article",{"robots":77,"canonical":32},"index,follow",{"doc_id":79,"site_id":7},273812,1789472853,{"code":4,"msg":82,"data":83},"success",[84,88,92,96,101,106,111,115,120,123,126],{"id":22,"doc_module":4,"doc_module_name":25,"category_name":85,"show_sort_weight":86,"slug":87},"Story & Novel",90,"story-novel",{"id":26,"doc_module":4,"doc_module_name":25,"category_name":89,"show_sort_weight":90,"slug":91},"Literature",80,"literature",{"id":33,"doc_module":4,"doc_module_name":25,"category_name":93,"show_sort_weight":94,"slug":95},"Exam",70,"exam",{"id":97,"doc_module":4,"doc_module_name":25,"category_name":98,"show_sort_weight":99,"slug":100},5,"Comic",60,"comic",{"id":102,"doc_module":4,"doc_module_name":25,"category_name":103,"show_sort_weight":104,"slug":105},6,"Technology",50,"technology",{"id":107,"doc_module":4,"doc_module_name":25,"category_name":108,"show_sort_weight":109,"slug":110},7,"Healthcare",40,"healthcare",{"id":112,"doc_module":4,"doc_module_name":25,"category_name":29,"show_sort_weight":113,"slug":114},8,30,"research-report",{"id":116,"doc_module":4,"doc_module_name":25,"category_name":117,"show_sort_weight":118,"slug":119},9,"Religion & Spirituality",20,"religion-spirituality",{"id":118,"doc_module":4,"doc_module_name":25,"category_name":121,"show_sort_weight":118,"slug":122},"World Cup","world-cup",{"id":55,"doc_module":4,"doc_module_name":25,"category_name":124,"show_sort_weight":55,"slug":125},"Lifestyle","lifestyle",{"id":127,"doc_module":4,"doc_module_name":25,"category_name":128,"show_sort_weight":97,"slug":129},19,"General","general",{"code":4,"msg":82,"data":131},{"doc_id":79,"user_id":132,"nickname":42,"user_avatar":133,"doc_module":4,"category_id":112,"category_name":29,"doc_title":10,"doc_description":12,"doc_content":134,"file_id":135,"file_url":136,"file_type":137,"file_size":138,"view_count":55,"is_deleted":4,"is_public":22,"is_downloadable":22,"audit_status":22,"page_count":139,"language":140,"language_code":8,"site_id":7,"html_lang":8,"table_of_contents":141,"faqs":142,"seo_title":143,"seo_description":12,"update_tm":80,"read_time":144,"preview_page_count":4},1374391974585,"https://ap-avatar.wpscdn.com/davatar_276721f389ce27ea32af1340a28f341c","Why2Speak: Faithful Reasoning for Abstaining Action Policies  \nShreya Mendi, Brinnae Bent  \nPratt School of Engineering, Duke University  \narXiv : 2608 .20670v 1 [cs .AI] 2 1 Aug 2026  \nAbstract  \nMany agentic systems must repeatedly choose between acting and abstaining, making faithful reasoning important for oversight: an explanation is useful only if it reflects the computation that produced the action, including the decision todo nothing. We study this problem through intervention timing in multi-party conversation, a controlled action-policy setting in which an assistant must decide whether to speak now or remain silent. This setting exposes challenges that are largely absent from question-answering faithfulness benchmarks, including severe class imbalance, asymmetric action costs, and the possibility that exposing reasoning changes the policy being audited. Using Qwen3-8B as a hybrid model that can be decoded with or without chain-of-thought reasoning, we compare direct decision policies, reasoning policies, supervised fine-tuning, and reinforcement learning. We find a consistent capability–auditability tradeoff: the strongest direct policy achieves higher decision quality but exposes no reasoning to inspect, while the reasoning policy providesan auditable trace at the cost of lower performance, particularly in recall of true intervention opportunities. Attempts to close this gap with supervised fine-tuning either suppress reasoning or preserve reasoning without improving decision quality. Reinforcement learning also fails to improve the reasoning policy; we identify a mechanism underlying this failure: group-relative objectives provide no learning signal on confidently wrong prompts when sampled rollouts all select the same action. We further audit the reasoning policy using controlled activation probes and behavioral ablations and show that standard faithfulness methods can overstate evidence that exposed reasoning reflects the underlying decision process. We show that standard faithfulness evaluations can be misleading in abstaining action policies: probability-based metrics saturate under confident decisions, probe-based analyses are vulnerable to class imbalance and textual leakage, and reasoning ablations can confound reasoning content with changes in inference mode. Together, these results show that exposing reasoning can change an agent’s action policy rather than simply make it observable. We characterize this capability–auditability tradeoff and provide methodological controls for evaluating reasoning-based oversight of agents that can actor abstain.  \nIntroduction  \nAI agents are increasingly deployed in settings where action is optional. A clinical assistant may flag a concern or remain  \nsilent; a tutoring system may interrupt a student or allow productive struggle; a meeting assistant may add missing information or let the conversation continue. In these domains, in addition to producing a response, the agent must also decide whether acting at all is appropriate. Unnecessary action can disrupt tasks and erode trust, while inappropriate abstention can leave errors or missed opportunities unaddressed.  \nThis paper studies faithful reasoning for such abstaining action policies. We focus on intervention timing in multiparty conversation as a concrete testbed: given a transcript prefix, an assistant must decide whether to intervene now or remain silent. Concurrent work under review shows that a learned classifier can identify when to make a useful intervention significantly better than zero-shot prompting (Nama et al. 2026) . However, that policy is opaque: it emits an action token but provides no account of why that moment warrants speaking, or why silence is preferable.  \nOne solution is to require the policy to reason before acting. A conversational agent could produce a short explanation such as “the group has accepted an incorrect deadline, so the agent should correct it,” and then decide whether to interve","cbCaimxmFFpT68oN","https://ap.wps.com/l/cbCaimxmFFpT68oN","pdf",408590,18,"English","# Abstract\n# Introduction\n## Background: optional actions in AI systems\n## Faithfulness and unfaithful explanations\n## Why abstaining differs from question answering\n## Controlled testbed with Qwen3-8B\n## Contributions","[{\"question\":\"What does “faithful reasoning” mean in abstaining action policies?\",\"answer\":\"An explanation is faithful only if it reflects the computation that actually influenced the chosen action, including when the correct behavior is to remain silent.\"},{\"question\":\"Why is intervention timing in multi-party conversation used as a testbed?\",\"answer\":\"It forces an assistant to decide whether to intervene now or not, exposing issues like asymmetric action costs and severe class imbalance that are often missing in question-answering faithfulness benchmarks.\"},{\"question\":\"What overall tradeoff does the study find when exposing reasoning?\",\"answer\":\"Stronger direct policies improve decision quality but provide no inspectable reasoning, while reasoning policies offer an auditable trace at the cost of lower performance, especially reduced recall of true intervention opportunities.\"}]","Why2Speak - Faithful Reasoning for Abstaining Action Policies | PDF",45]