[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"doc-detail-85225-en":3,"doc-seo-85225-105":30,"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":21,"is_downloadable":21,"audit_status":21,"page_count":22,"language":23,"language_code":24,"site_id":25,"html_lang":24,"table_of_contents":26,"faqs":27,"seo_title":13,"seo_description":14,"update_tm":28,"read_time":29},85225,13056703019662,"Evangeline","https://ap-avatar.wpscdn.com/avatar/be000253a8e92610077?_k=1778726343310543188",8,"Research & Report","Tool Adaptive LLM Reranker TALRanker","Generative Large Language Models (LLMs) improve information retrieval but often produce severe factual hallucinations when queries exceed their parametric knowledge. External tool-calling can reduce hallucinations; however, using tools for every document during reranking introduces prohibitive latency, creating an accuracy–efficiency tradeoff. This work proposes TALRanker, a framework that models pointwise relevance scoring as an agentic Markov decision process and trains it with a two-stage procedure for confidence-aware tool bypass and selective evidence retrieval. Results on standard and reasoning-intensive benchmarks show state-of-the-art ranking quality with throughput comparable to discriminative rerankers.","Tool-Adaptive LLM Reranker  \nZichuan Liu 1∗†, Ruijin Hua2∗  \n1 Carnegie Mellon University, Pittsburgh, PA 15213, USA  \n2Huazhong University of Science and Technology, Wuhan, China  \n[zichuanl@andrew.cmu.edu](zichuanl@andrew.cmu.edu), [ruijinhua07@gmail.com](ruijinhua07@gmail.com)  \narXiv :2607 . 10555v 1 [ cs .IR] 12 Jul 2026  \nAbstract  \nGenerative Large Language Models (LLMs) have revolutionized information retrieval, yet their strictly parametric nature frequently leads to severe factual hallucinations when confronted with complex queries beyond their epistemic boundaries. While external tool-calling can mitigate this, indiscriminately invoking search tools for every document during reranking incurs prohibitive latency overheads, creating an intractable accuracy-efficiency dilemma. To address this challenge, we propose TALRanker, a novel framework that formalizes pointwise relevance scoring as an agentic Markov decision process. We optimize it via a two-stage training paradigm. An initial warm-up utilizes a language-preserving hybrid loss to prevent the catastrophic forgetting of native generative capacities. Subsequently, an asymmetric cost-aware reward equipped in reinforcement learning forces the policy to autonomously bypass tools for maximum efficiency when confident, while selectively retrieving external evidence to avert severe hallucination penalties when uncertain. Extensive evaluations demonstrate that TALRanker achieves state-of-the-art performance across standard and reasoning-intensive retrieval benchmarks, matching throughput with pointwise rerankers while outperforming parameter-heavy reasoning models.  \nIntroduction  \nDriven by the escalating complexity of user needs, information retrieval systems are undergoing a fundamental shift: from executing simple keyword searches to resolving complex, inferential questions that demand deep reasoning (Weller et al. 2025a; Mao et al. 2024; Guo et al. 2025) . In advanced applications such as Retrieval-Augmented Generation (RAG) and autonomous agents (Yu et al. 2024; Liet al. 2026; Zeng et al. 2026; Wang et al. 2024), accurately assessing relevance under these inferential scenarios is critical. Consequently, traditional scalar models are largely superseded by Generative Relevance Models (GRMs) powered by Large Language Models (LLMs) (Li et al. 2025) . Specifically, reasoning-based GRMs that leverage chain-ofthought (Wei et al. 2022) have demonstrated remarkable capabilities in analyzing query-document relationships (Zhang et al. 2025a; Weller et al. 2025b) . However, these models are  \n∗These authors contributed equally.  \n†Corresponding author. Preprint.  \nBM25 RankGPT-4 TFRank  \nDirect-Point Reason-Point Rank1  \n Rank-R1  \n REARANK  \n TALReranker (Ours)  \nNDCG@10  \n\n|  |  |  |  | High | Effic |  | iency & |  |  | Accura | cy |  |  |\n| --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- |\n|  |  |  |  |  |  |  |  |  |  |  |  |  |  |\n| 30 |  |  |  | 8B |  |  |  |  |  | \u003Cbr>4B |  |  |  |\n|  |  |  |  |  |  |  | 4B |  |  |  |  |  |  |\n| 25 | 14B |  | \u003Cbr>8B |  |  |  | \u003Cbr>8B |  | 4B |  |  |  |  |\n| GPT | \u003Cbr>7 14B\u003Cbr>\u003Cbr>\u003Cbr>-4 7B |  |  B | 4B |  |  | \u003Cbr>7B |  |  | 1.7B\u003Cbr>1.7B\u003Cbr> |  |  | M25 |\n|  |  |  |  |  |  |  |  |  |  |  |  |  |  |\n| 15 | 3B |  |  |  |  |  |  |  |  |  | 0.6B |  |  |\n| 10\u003Cbr>5  |  |  |  |  |  |  |  |  |  |  |  | \u003Cbr>0 | \u003Cbr>.6B |\n|  |  |  |  |  |  |  |  |  |  |  |  |  |  |\n\n20  \nB  \n8B  \n0 200 400 600 800 1000 1200 1400 1600 1800  \nThroughput (Queries/Hour)  \nFigure 1: The tradeoff between precision (NDCG@10) and single-GPU throughput on the BRIGHT benchmark. TALRanker dominates the top-right quadrant, achieving state-ofthe-art accuracy while maintaining extreme inference speeds comparable to non-reasoning discriminative baselines.  \nfundamentally constrained by their parametric boundaries, operating in a “closed-book\" setting, and the internal reasoning pathways inevitably break down when confronted without-of-distributi","cbCaiifdvRjNk1f9","https://ap.wps.com/l/cbCaiifdvRjNk1f9","pdf",862667,3,1,12,"English","en",105,"# Introduction\n## Problem of accuracy-efficiency in tool-based reranking\n## Proposed solution: TALRanker framework\n# Method overview (tool-adaptive agentic scoring)\n## Two-stage training paradigm","[{\"question\":\"What problem does TALRanker address in LLM-based reranking?\",\"answer\":\"TALRanker targets the accuracy–efficiency dilemma: LLM rerankers can hallucinate on complex queries, but calling tools for every querydocument pair causes unacceptable latency in practical systems.\"},{\"question\":\"How does TALRanker decide when to use external tools?\",\"answer\":\"TALRanker estimates relevance as an agentic Markov decision process and learns an adaptive routing policy that bypasses tools when internal confidence is high, while selectively retrieving evidence when uncertainty is detected.\"},{\"question\":\"What is the purpose of TALRanker’s two-stage training?\",\"answer\":\"The warm-up stage uses a language-preserving hybrid loss to prevent catastrophic forgetting. The subsequent stage uses an asymmetric, cost-aware reinforcement learning reward to trade off inference cost and hallucination risk.\"}]",1784201853,30,{"code":4,"msg":31,"data":32},"ok",{"site_id":25,"language":24,"slug":33,"title":13,"keywords":34,"description":14,"schema_data":35,"social_meta":86,"head_meta":88,"extra_data":90,"updated_unix":28},"tool-adaptive-llm-reranker-talranker","",{"@graph":36,"@context":85},[37,53,68],{"@type":38,"itemListElement":39},"BreadcrumbList",[40,44,48,50],{"item":41,"name":42,"@type":43,"position":21},"https://docshare.wps.com","Home","ListItem",{"item":45,"name":46,"@type":43,"position":47},"https://docshare.wps.com/document/","Document",2,{"item":49,"name":12,"@type":43,"position":20},"https://docshare.wps.com/document/research-report/",{"item":51,"name":13,"@type":43,"position":52},"https://docshare.wps.com/document/tool-adaptive-llm-reranker-talranker/85225/",4,{"url":51,"name":13,"@type":54,"author":55,"headline":13,"publisher":57,"fileFormat":60,"inLanguage":24,"description":14,"dateModified":61,"datePublished":62,"encodingFormat":60,"isAccessibleForFree":63,"interactionStatistic":64},"DigitalDocument",{"name":9,"@type":56},"Person",{"url":41,"name":58,"@type":59},"DocShare","Organization","application/pdf","2026-07-24","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 TALRanker address in LLM-based reranking?","Question",{"text":75,"@type":76},"TALRanker targets the accuracy–efficiency dilemma: LLM rerankers can hallucinate on complex queries, but calling tools for every querydocument pair causes unacceptable latency in practical systems.","Answer",{"name":78,"@type":73,"acceptedAnswer":79},"How does TALRanker decide when to use external tools?",{"text":80,"@type":76},"TALRanker estimates relevance as an agentic Markov decision process and learns an adaptive routing policy that bypasses tools when internal confidence is high, while selectively retrieving evidence when uncertainty is detected.",{"name":82,"@type":73,"acceptedAnswer":83},"What is the purpose of TALRanker’s two-stage training?",{"text":84,"@type":76},"The warm-up stage uses a language-preserving hybrid loss to prevent catastrophic forgetting. The subsequent stage uses an asymmetric, cost-aware reinforcement learning reward to trade off inference cost and hallucination risk.","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":25},{"code":4,"msg":5,"data":92},[93,97,101,105,110,115,120,122,127,130,134],{"id":21,"doc_module":4,"doc_module_name":46,"category_name":94,"show_sort_weight":95,"slug":96},"Story & Novel",90,"story-novel",{"id":47,"doc_module":4,"doc_module_name":46,"category_name":98,"show_sort_weight":99,"slug":100},"Literature",80,"literature",{"id":52,"doc_module":4,"doc_module_name":46,"category_name":102,"show_sort_weight":103,"slug":104},"Exam",70,"exam",{"id":106,"doc_module":4,"doc_module_name":46,"category_name":107,"show_sort_weight":108,"slug":109},5,"Comic",60,"comic",{"id":111,"doc_module":4,"doc_module_name":46,"category_name":112,"show_sort_weight":113,"slug":114},6,"Technology",50,"technology",{"id":116,"doc_module":4,"doc_module_name":46,"category_name":117,"show_sort_weight":118,"slug":119},7,"Healthcare",40,"healthcare",{"id":11,"doc_module":4,"doc_module_name":46,"category_name":12,"show_sort_weight":29,"slug":121},"research-report",{"id":123,"doc_module":4,"doc_module_name":46,"category_name":124,"show_sort_weight":125,"slug":126},9,"Religion & Spirituality",20,"religion-spirituality",{"id":125,"doc_module":4,"doc_module_name":46,"category_name":128,"show_sort_weight":125,"slug":129},"World Cup","world-cup",{"id":131,"doc_module":4,"doc_module_name":46,"category_name":132,"show_sort_weight":131,"slug":133},10,"Lifestyle","lifestyle",{"id":135,"doc_module":4,"doc_module_name":46,"category_name":136,"show_sort_weight":106,"slug":137},19,"General","general"]