[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"doc-detail-85715-en":3,"doc-seo-85715-105":29,"detail-sidebar-cat-0-en-105":91},{"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":20,"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},85715,4398048950312,"Violet","https://ap-avatar.wpscdn.com/avatar/400002538284de19e3c?_k=1778320343897328908",8,"Research & Report","Length Penalties Make Chain-of-Thought Less Monitorable","Length-penalized reinforcement learning can shorten chain-of-thought reasoning while concealing the influence that drives a model’s answer. Experiments show that training with length penalties still allows misleading hint steering even when hints appear less frequently in the generated traces. Token–accuracy evaluation can mark such runs successful despite reduced faithfulness, because remaining traces may still not reveal what shaped the response. Qwen3-4B and Qwen3-14B trained under compression reduce reasoning tokens sharply, keep most multiple-choice accuracy, and leave hint influence near baseline while lowering monitor-detectable hint attribution and confidence.","arXiv :2607 .09786v 1 [ cs .AI] 8 Jul 2026  \nLength Penalties Make Chain-of-Thought Less Monitorable  \nBryce Little*  \nAbstract  \nLength-penalized reinforcement learning can shorten chain-of-thought reasoning while hiding an influence that drives the model’s answer. In our experiments, training with length penalties does not stop misleading hints from steering models, even though the models’ chains of thought mention the hint much less often. A token–accuracy evaluation would count these runs as successful because they use fewer reasoning tokens with little accuracy loss; it would miss whether the remaining trace still shows what drove the answer. We train Qwen3-4B and Qwen3-14B variants with different target chain lengths, then evaluate them with biasing-hint interventions on held-out MMLU-Pro-R and four transfer benchmarks. Compression sharply cuts reasoning tokens, preserves most multiple-choice accuracy, and leaves hint influence near baseline. At the strongest target, lower-bound faithfulness falls to 63.1% of baseline for Qwen3- 14B and 69.4% for Qwen3-4B; the raw rate at which a monitor catches hint use falls from 69% to 49% and from 60% to 48% . To separate length from content, we randomly delete sentences from uncompressed baseline chains until the remaining text matches the compressed length. Even after this length matching, compressed chains disclose the hint 7–35 percentage points less often than baseline chains that we shorten at random, for both Qwen3 sizes and all five evaluation distributions. Compression therefore does more than shorten reasoning, preferentially removing the cues a monitor needs to see what influenced the answer. Together, these results reveal a compression–monitorability frontier in which cheaper reasoning can preserve answers while making the influences behind them harder to detect.  \nMedian CoT tokens  \n3 , 000 2 , 000  \n1 , 000  \n0  \n3 , 000 2 , 000  \n1 , 000  \n0  \n Median CoT length  ∆ accuracy  ∆ faithfulness  \nQwen3-14B  \nQwen3-4B  \nBaseline 60% 50% 40% 30%  \n0  \n−10  \n−20  \n−30  \n−40  \n0  \n−10  \n−20  \n−30  \n−40  \nChange from baseline (pp)  \nCoT compression target  \nFigure 1: Length-penalized RL compresses CoT with little accuracy loss but lower faithfulness on the held-out MMLU-Pro-R test set. Bars show median CoT length; lines show baseline-relative changes in accuracy and faithfulness. Whiskers show 95% bootstrap confidence intervals; we set baselines to zero.  \n*[bryceplittle@gmail.com](bryceplittle@gmail.com)  \n1 Introduction  \nChain-of-thought (CoT) generation gives a language model extra test-time compute to improve its answers (Wei et al., 2022; Kojima et al., 2022), and reasoning models now generate long traces by default (OpenAI, 2024; DeepSeek-AI, 2025; Yang et al., 2025) . Because thinking tokens increase inference cost, recent work trains models to use fewer of them, often with substantial reductions and little or no loss in task accuracy (Arora and Zanette, 2025; Aggarwal and Welleck, 2025; Yi and Wang, 2025; Xiang et al., 2025) . This token–accuracy criterion overlooks a second role of the trace. For monitoring, a chain of thought helps only insofar as it exposes what drove the answer (Korbaket al., 2025; Baker et al., 2025) . Our paper studies this missing axis of evaluation. We show that compression can preserve task behavior while stripping away the trace evidence that would reveal a biasing influence.  \nTo measure the tradeoff between efficiency and monitorability, we train Qwen3-4B and Qwen3- 14B with calibrated length penalties targeting several compression regimes. We also train a reinforcement learning (RL) control with the same correctness and format rewards but no length penalty, isolating the effect of compression from RL itself. We adopt the hint-intervention setup of Turpin et al. (2023) and Chen et al. (2025), adding a hint pointing to a target answer, measuring whether the model switches toward that target, and counting the chain of thought as faithful only","cbCaiuL2qBPyqQQN","https://ap.wps.com/l/cbCaiuL2qBPyqQQN","pdf",430028,1,31,"English","en",105,"# Abstract\n# Introduction\n## Efficiency vs. monitorability in chain-of-thought\n## Experimental setup and hint intervention\n## Results: accuracy preservation with faithfulness loss","[{\"question\":\"What problem does the paper identify with length-penalized chain-of-thought training?\",\"answer\":\"Length-penalized training can compress reasoning traces while hiding the influence that drives answers, reducing what monitors can detect even when accuracy remains similar.\"},{\"question\":\"How do the authors evaluate both efficiency and monitorability?\",\"answer\":\"They compare token–accuracy measures with hint-intervention evaluations that assess whether a monitor can detect the hint’s role in the trace, reporting changes in faithfulness and monitor catch rates.\"},{\"question\":\"What do the results show about compression’s impact on accuracy and hint influence?\",\"answer\":\"Compression sharply cuts reasoning tokens and preserves most multiple-choice accuracy, while hint influence on answers stays near baseline; however, faithfulness and detectable hint use decline substantially.\"}]",1784205761,78,{"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":86,"head_meta":88,"extra_data":90,"updated_unix":27},"length-penalties-make-chain-of-thought-less-monitorable","",{"@graph":35,"@context":85},[36,53,68],{"@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/length-penalties-make-chain-of-thought-less-monitorable/85715/",4,{"url":51,"name":13,"@type":54,"author":55,"headline":13,"publisher":57,"fileFormat":60,"inLanguage":23,"description":14,"dateModified":61,"datePublished":62,"encodingFormat":60,"isAccessibleForFree":63,"interactionStatistic":64},"DigitalDocument",{"name":9,"@type":56},"Person",{"url":40,"name":58,"@type":59},"DocShare","Organization","application/pdf","2026-07-17","2026-07-16",true,{"@type":65,"interactionType":66,"userInteractionCount":20},"InteractionCounter",{"@type":67},"ViewAction",{"@type":69,"mainEntity":70},"FAQPage",[71,77,81],{"name":72,"@type":73,"acceptedAnswer":74},"What problem does the paper identify with length-penalized chain-of-thought training?","Question",{"text":75,"@type":76},"Length-penalized training can compress reasoning traces while hiding the influence that drives answers, reducing what monitors can detect even when accuracy remains similar.","Answer",{"name":78,"@type":73,"acceptedAnswer":79},"How do the authors evaluate both efficiency and monitorability?",{"text":80,"@type":76},"They compare token–accuracy measures with hint-intervention evaluations that assess whether a monitor can detect the hint’s role in the trace, reporting changes in faithfulness and monitor catch rates.",{"name":82,"@type":73,"acceptedAnswer":83},"What do the results show about compression’s impact on accuracy and hint influence?",{"text":84,"@type":76},"Compression sharply cuts reasoning tokens and preserves most multiple-choice accuracy, while hint influence on answers stays near baseline; however, faithfulness and detectable hint use decline substantially.","https://schema.org",{"og:url":51,"og:type":87,"og:title":13,"og:site_name":58,"og:description":14},"article",{"robots":89,"canonical":51},"index,follow",{"doc_id":7,"site_id":24},{"code":4,"msg":5,"data":92},[93,97,101,105,110,115,120,123,128,131,135],{"id":20,"doc_module":4,"doc_module_name":45,"category_name":94,"show_sort_weight":95,"slug":96},"Story & Novel",90,"story-novel",{"id":46,"doc_module":4,"doc_module_name":45,"category_name":98,"show_sort_weight":99,"slug":100},"Literature",80,"literature",{"id":52,"doc_module":4,"doc_module_name":45,"category_name":102,"show_sort_weight":103,"slug":104},"Exam",70,"exam",{"id":106,"doc_module":4,"doc_module_name":45,"category_name":107,"show_sort_weight":108,"slug":109},5,"Comic",60,"comic",{"id":111,"doc_module":4,"doc_module_name":45,"category_name":112,"show_sort_weight":113,"slug":114},6,"Technology",50,"technology",{"id":116,"doc_module":4,"doc_module_name":45,"category_name":117,"show_sort_weight":118,"slug":119},7,"Healthcare",40,"healthcare",{"id":11,"doc_module":4,"doc_module_name":45,"category_name":12,"show_sort_weight":121,"slug":122},30,"research-report",{"id":124,"doc_module":4,"doc_module_name":45,"category_name":125,"show_sort_weight":126,"slug":127},9,"Religion & Spirituality",20,"religion-spirituality",{"id":126,"doc_module":4,"doc_module_name":45,"category_name":129,"show_sort_weight":126,"slug":130},"World Cup","world-cup",{"id":132,"doc_module":4,"doc_module_name":45,"category_name":133,"show_sort_weight":132,"slug":134},10,"Lifestyle","lifestyle",{"id":136,"doc_module":4,"doc_module_name":45,"category_name":137,"show_sort_weight":106,"slug":138},19,"General","general"]