Are Rationales Necessary and Sufficient? Tuning LLMs for Explainable Misinformation Detection
Summary
This paper proposes a pipeline for fine-tuning LLMs specifically for explainable misinformation detection and introduces LonsRex, a data synthesis method to generate necessary and sufficient rationales, addressing limitations of naive filtering based solely on label correctness.
arXiv:2605.19285v1 Announce Type: new
Abstract: The rapid spread of misinformation on social media platforms has become a formidable challenge. To mitigate its proliferation, Misinformation Detection (MD) has emerged as a critical research topic. Traditional MD approaches based on small models typically perform binary classification through a black-box process. Recently, the rise of Large Language Models (LLMs) has enabled explainable MD, where models generate rationales that explain their decisions, thereby enhancing transparency. Existing explainable MD methods primarily focus on crafting sophisticated prompts to elicit rationales from off-the-shelf LLMs. In this work, we propose a pipeline to fine-tune a dedicated LLM specifically for explainable MD. Our pipeline begins by collecting large-scale fact-checked articles, and then uses multiple strong LLMs to produce veracity predictions and rationales. To ensure high-quality training data, we leverage a filtering strategy that selects only the correct instances for fine-tuning. While this pipeline is intuitive and prevalent, our experiments reveal that naive filtering based solely on label correctness is insufficient in practice and suffers from two critical limitations: (1) Coarse-grained labels cause insufficient rationales: Rationales filtered solely based on binary labels are insufficient to adequately support their decisions; (2) Over-verification behavior causes unnecessary rationales: Stronger LLMs tend to exhibit over-verification behavior, producing excessively verbose and unnecessary rationales. To address these issues, we introduce LONSREX, a novel data synthesis pipeline to Locate Necessary and Sufficient Rationales for Explainable MD. Specifically, we propose a metric that quantifies the contribution of each verification step to the final prediction, thereby evaluating its necessity and sufficiency. Experimental results demonstrate the effectiveness of LONSREX.
View Cached Full Text
Cached at: 05/20/26, 08:25 AM
# Are Rationales Necessary and Sufficient? Tuning LLMs for Explainable Misinformation Detection Source: [https://arxiv.org/html/2605.19285](https://arxiv.org/html/2605.19285) Bing Wang[0000\-0002\-1304\-3718](https://orcid.org/0000-0002-1304-3718)College of Computer Science and Technology, Jilin UniversityChangchunChinaTongyi Lab, Alibaba GroupHangzhouChina[wangbing1416@gmail\.com](https://arxiv.org/html/2605.19285v1/mailto:[email protected])Rui Miao[0000\-0002\-2917\-2311](https://orcid.org/0000-0002-2917-2311)School of Artificial Intelligence, Jilin UniversityChangchunChina[miaorui24@mails\.jlu\.edu\.cn](https://arxiv.org/html/2605.19285v1/mailto:[email protected]),Ximing Li[0000\-0001\-8190\-5087](https://orcid.org/0000-0001-8190-5087)College of Computer Science and Technology, Jilin UniversityChangchunChinaRIKEN AIPTokyoJapan[liximing86@gmail\.com](https://arxiv.org/html/2605.19285v1/mailto:[email protected]),Chen Shen[0000\-0002\-7534\-0830](https://orcid.org/0000-0002-7534-0830),Shaotian Yan[0000\-0002\-9029\-0763](https://orcid.org/0000-0002-9029-0763)Tongyi Lab, Alibaba GroupHangzhouChina[zjushenchen@gmail\.com](https://arxiv.org/html/2605.19285v1/mailto:[email protected])[yanshaotian@gmail\.com](https://arxiv.org/html/2605.19285v1/mailto:[email protected]),Changchun Li[0000\-0002\-8001\-2655](https://orcid.org/0000-0002-8001-2655)College of Computer Science and Technology, Jilin UniversityChangchunChina[changchunli93@gmail\.com](https://arxiv.org/html/2605.19285v1/mailto:[email protected]),Kaiyuan Liu[0009\-0005\-4717\-0087](https://orcid.org/0009-0005-4717-0087)College of Computer Science and Technology, Zhejiang UniversityHangzhouChina[12421281@zju\.edu\.cn](https://arxiv.org/html/2605.19285v1/mailto:[email protected]),Xiaosong Yuan[0000\-0001\-5748\-5174](https://orcid.org/0000-0001-5748-5174)College of Computer Science and Technology, Jilin UniversityChangchunChina[yuanxs19@mails\.jlu\.edu\.cn](https://arxiv.org/html/2605.19285v1/mailto:[email protected])andJieping Ye[0000\-0001\-8662\-5818](https://orcid.org/0000-0001-8662-5818)Tongyi Lab, Alibaba GroupHangzhouChina[yejieping\.ye@alibaba\-inc\.com](https://arxiv.org/html/2605.19285v1/mailto:[email protected]) \(2026\) ###### Abstract\. The rapid spread of misinformation on social media platforms has become a formidable challenge\. To mitigate its proliferation, Misinformation Detection \(MD\) has emerged as a critical research topic\. Traditional MD approaches based on small models typically perform binary classification,*e\.g\.,*real and fake, through a black\-box process\. Recently, the rise of Large Language Models \(LLMs\) has enabledexplainable MD, where models generate rationales that explain their decisions, thereby enhancing transparency\. Existing explainable MD methods primarily focus on crafting sophisticated prompts to elicit rationales from off\-the\-shelf LLMs\. In this work, we propose a pipeline to fine\-tune a dedicated LLM specifically for explainable MD\. Our pipeline begins by collecting large\-scale fact\-checked articles, and then uses multiple strong LLMs to produce veracity predictions and rationales\. To ensure high\-quality training data, we leverage a filtering strategy that selects only the correct instances for fine\-tuning\. While this pipeline is intuitive and prevalent, our experiments reveal that naive filtering based solely on label correctness is insufficient in practice and suffers from two critical limitations: \(1\)Coarse\-grained labels cause insufficient rationales: Rationales filtered solely based on binary labels are insufficient to adequately support their decisions; \(2\)Over\-verification behavior causes unnecessary rationales: Stronger LLMs tend to exhibit over\-verification behavior, producing excessively verbose and unnecessary rationales\. To address these issues, we introduceLonsRex, a novel data synthesis pipeline toLocateNecessary andSufficientRationales forExplainable MD\. Specifically, we propose a metric that quantifies the contribution of each verification step to the final prediction, thereby evaluating its necessity and sufficiency\. Experimental results demonstrate thatLonsReximproves the accuracy of baseline LLMs by approximately 22\.97% and is comparable to larger LLMs\. We will publicly release our 316k raw data and the filtered version byLonsRex\. Social media, misinformation detection, large language models, explainability, supervised fine\-tuning, data filtering ††copyright:acmlicensed††journalyear:2026††doi:XXXXXXX\.XXXXXXX††conference:the 32st ACM SIGKDD Conference on Knowledge Discovery and Data Mining; 9–13 August, 2026; Jeju, Korea††isbn:978\-1\-4503\-XXXX\-X/18/06††ccs:Computing methodologies Artificial intelligence††ccs:Information systems Social networks## 1\.Introduction On today’s social media platforms,*e\.g\.,*X\.com and Reddit, vast numbers of users actively disseminate and consume information\. Unfortunately, this information ecosystem is inevitably polluted by a substantial amount of deliberately fabricated misinformation, which poses serious threats to users’ financial security and psychological well\-being\(Lazeret al\.,[2018](https://arxiv.org/html/2605.19285#bib.bib1); Vosoughiet al\.,[2018](https://arxiv.org/html/2605.19285#bib.bib2)\)\. To mitigate the spread of such misinformation content, Misinformation Detection \(MD\) has emerged as a critical and active research topic, aiming to automatically and rapidly identify misinformation on social media to facilitate timely intervention and control\(Huet al\.,[2024](https://arxiv.org/html/2605.19285#bib.bib16); Wanget al\.,[2025b](https://arxiv.org/html/2605.19285#bib.bib17); Modzelewskiet al\.,[2025](https://arxiv.org/html/2605.19285#bib.bib3)\)\. Figure 1\.A representative case demonstrate one rationale generated by LLMs for fake claims\.Typically, prior MD methods involve training discriminative models to classify claims on social media into binary labels,*i\.e\.,*real or fake, by incorporating various external features,*e\.g\.,*emotion signals\(Zhanget al\.,[2021](https://arxiv.org/html/2605.19285#bib.bib25); Liuet al\.,[2025b](https://arxiv.org/html/2605.19285#bib.bib26)\), commonsense knowledge\(Dunet al\.,[2021](https://arxiv.org/html/2605.19285#bib.bib28); Wanget al\.,[2025a](https://arxiv.org/html/2605.19285#bib.bib27)\), and publisher intent\(Wanget al\.,[2024a](https://arxiv.org/html/2605.19285#bib.bib23),[2025d](https://arxiv.org/html/2605.19285#bib.bib24)\)\. Despite their superior performance, the black\-box nature of these discriminative models prevents them from revealing the underlying reasoning behind their predictions\. Accordingly, to improve decision transparency, recent research has shifted towardexplainable MD, which seeks to generatenatural language rationalesin addition to the binary output\(Yanget al\.,[2022](https://arxiv.org/html/2605.19285#bib.bib22); Wanget al\.,[2024b](https://arxiv.org/html/2605.19285#bib.bib20),[2025c](https://arxiv.org/html/2605.19285#bib.bib21)\)\. They primarily leverage Large Language Models \(LLMs\) to design sophisticated prompts to directly elicit rationales\(Huet al\.,[2024](https://arxiv.org/html/2605.19285#bib.bib16); Modzelewskiet al\.,[2025](https://arxiv.org/html/2605.19285#bib.bib3)\)or summarize retrieved external evidence to serve as rationales\(Wanget al\.,[2024b](https://arxiv.org/html/2605.19285#bib.bib20); Yueet al\.,[2024](https://arxiv.org/html/2605.19285#bib.bib29)\)\. For example, PCoT\(Modzelewskiet al\.,[2025](https://arxiv.org/html/2605.19285#bib.bib3)\)employs a system prompt grounded in persuasion knowledge; L\-Defense\(Wanget al\.,[2024b](https://arxiv.org/html/2605.19285#bib.bib20)\)clusters retrieved evidence into supporting and opposing groups and uses an LLM to summarize them, respectively\. Although these approaches outperform straightforward zero\-shot prompting with LLMs, they heavily depend on the capabilities of the foundation LLMs and the quality of the retrieved evidence, which limits their potential\. In this work, we take a different perspective and design a data\-centric pipeline thatfine\-tunes a lightweight LLM specifically for explainable MD\. Specifically, our pipeline begins by collecting 316k publicly available fact\-checked claims, each labeled as either real or fake\. We then prompt multiple strong LLMs to generate corresponding step\-by\-step rationales for these claims, as illustrated in Fig\.[1](https://arxiv.org/html/2605.19285#S1.F1)\. Finally, some heuristic filtering criteria,*e\.g\.,*correctness\(Chenet al\.,[2025](https://arxiv.org/html/2605.19285#bib.bib30); Luoet al\.,[2025](https://arxiv.org/html/2605.19285#bib.bib31)\)and formatting standards\(Guoet al\.,[2025](https://arxiv.org/html/2605.19285#bib.bib4)\), are applied to select higher\-quality rationale\-label pairs for supervised fine\-tuning of a lightweight LLM,*e\.g\.,*Qwen3\-4B\-Instruct\(Yanget al\.,[2025](https://arxiv.org/html/2605.19285#bib.bib5)\)\. While this pipeline is intuitive and prevalent, our experiments further reveal that naive filtering based solely on binary label correctness is insufficient in practice for explainable MD and suffers from two critical limitations:\(1\) Coarse\-grained labels cause insufficient rationales: As illustrated in Fig\.[1](https://arxiv.org/html/2605.19285#S1.F1), the rationale*w\.r\.t*a claim typically involves more than ten verification steps\. Relying solely on the correctness of binary labels to filter training samples is insufficient to evaluate the quality of each verification step\. Therefore, the training data inevitably contain rationales that are meaningless or even noisy, which is empirically analyzed in Sec\.[2\.1](https://arxiv.org/html/2605.19285#S2.SS1)\.\(2\) Over\-verification behavior causes unnecessary rationales: Our experiments in Sec\.[2\.2](https://arxiv.org/html/2605.19285#S2.SS2)demonstrate that LLMs with stronger reasoning capabilities tend to generate more complex verification processes, leading to overly strict claim verification\. Such over\-verification not only increases the likelihood of labeling claims as fake, but also degrades the performance of LLMs trained on these rationales, as excessive and unnecessary verification steps introduce noise and reduce training efficiency\. To alleviate these issues, we update our training pipeline and propose a new method, namelyLocatingNecessary andSufficientRationales forExplainable MD \(LonsRex\)\. Specifically, the basic idea ofLonsRexis to design a perplexity\-based metric that quantifies the contribution of each verification step within rationales to the prediction, and to select necessary and sufficient rationales that contain fewer useless verification steps and more useful ones\. To achieve this, we first evaluate the contribution of each verification step by measuring the change in the LLM’s prediction perplexity when that step is masked\. Building on this counterfactual attribution framework, we introduce two complementary attribution metrics:self\-attributionandmutual\-attributionscores\. The self\-attribution score evaluates the average contribution of individual verification steps within a single rationale; The mutual\-attribution score summarizes diverse verification perspectives across multiple rationales, and assumes that frequently occurring high\-contribution perspectives are generally more important; accordingly, it assesses the average contribution of perspectives within each rationale\. During the implementation, we generate 948k rationales and predictions, and select 200k high\-quality ones to fine\-tune LLMs of varying sizes,*e\.g\.,*4B, 2B, and 1\.5B parameters\. The resulting LLMs significantly outperform their foundation counterparts and achieve performance comparable to that of much larger open\-source LLMs\. Our contributions can be summarized as the following three\-fold: - •We empirically identify that rationale\-based training data for explainable MD, generated by LLMs, often contains an amount of unnecessary and insufficient rationales\. - •To address this, we propose a novel pipelineLonsRexthat fine\-tunes LLMs to generate high\-quality MD rationales by explicitly locating those necessary and sufficient rationales\. - •We further construct an evaluation benchmark for explainable MD and demonstrate that our fine\-tuned LLMs consistently outperform their foundation LLMs and achieve competitive results against much larger open\-source LLMs\. ## 2\.Preliminary Empirical Analysis We empirically evaluate the training rationales generated by advanced LLMs and find that they contain numerous unnecessary and insufficient rationales, even when their predictions are correct\. Task formulation of explainable MD\.Formally, given a claim𝐱\\mathbf\{x\}from social media platforms, explainable MD aims to predict a veracity labely∈\{0,1\}y\\in\\\{0,1\\\}\(0/1 means fake/real\) and generate a corresponding rationale𝐄=\{𝐞1,…,𝐞L\}\\mathbf\{E\}=\\\{\\mathbf\{e\}\_\{1\},\\ldots,\\mathbf\{e\}\_\{L\}\\\}\. Here,𝐞l\\mathbf\{e\}\_\{l\}denotes thell\-th verification step that analyzes the claim from a specific perspective, as illustrated in Fig\.[1](https://arxiv.org/html/2605.19285#S1.F1)\. In this work, we collect a large\-scale fact\-checked dataset𝒟=\{\(𝐱i,yi\)\}i=1\|𝒟\|\\mathcal\{D\}=\\\{\(\\mathbf\{x\}\_\{i\},y\_\{i\}\)\\\}\_\{i=1\}^\{\|\\mathcal\{D\}\|\}from open\-sourced repositories\. We leverage advanced LLMs to generate a set of candidate rationales and predictions\{\(𝐄ik,y^ik\)\}k=1K\\\{\(\\mathbf\{E\}\_\{ik\},\\widehat\{y\}\_\{ik\}\)\\\}\_\{k=1\}^\{K\}for each claim𝐱i\\mathbf\{x\}\_\{i\}, whereKKdenotes the number of candidates per claim\. Then, we apply heuristic filtering strategies, such as filtering based on veracity labelsy^\\widehat\{y\}, to curate a high\-quality subset𝒟^=\{\(𝐱i,𝐄i,y^i\)\}i=1\|𝒟^\|\\mathcal\{\\widehat\{D\}\}=\\\{\(\\mathbf\{x\}\_\{i\},\\mathbf\{E\}\_\{i\},\\widehat\{y\}\_\{i\}\)\\\}\_\{i=1\}^\{\|\\mathcal\{\\widehat\{D\}\}\|\}, where\|𝒟^\|<K\|𝒟\|\|\\mathcal\{\\widehat\{D\}\}\|<K\|\\mathcal\{D\}\|\. Upon this refined subset, we fine\-tune a lightweight LLM, parameterized by𝜽\\boldsymbol\{\\theta\}, by minimizing the negative log\-likelihood loss as follows: \(1\)𝐦𝐢𝐧𝜽ℒ\(𝜽\)=−1\|𝒟^\|∑\(𝐱i,𝐄i,y^i\)∈𝒟^logP𝜽\(𝐄i,y^i∣𝐱i\)\.\\mathop\{\\boldsymbol\{\\min\}\}\_\{\\boldsymbol\{\\theta\}\}\\mathcal\{L\}\(\\boldsymbol\{\\theta\}\)=\-\\frac\{1\}\{\|\\mathcal\{\\widehat\{D\}\}\|\}\\sum\\nolimits\_\{\(\\mathbf\{x\}\_\{i\},\\mathbf\{E\}\_\{i\},\\widehat\{y\}\_\{i\}\)\\in\\mathcal\{\\widehat\{D\}\}\}\\log P\_\{\\boldsymbol\{\\theta\}\}\\left\(\\mathbf\{E\}\_\{i\},\\widehat\{y\}\_\{i\}\\mid\\mathbf\{x\}\_\{i\}\\right\)\. Evaluation settings\.Our work aims to evaluate the necessity and sufficiency of rationales by measuring the contribution of each verification step to the veracity prediction\. We hypothesize that a rationale isunnecessaryif it contains a high proportion of useless steps; conversely, it is deemedinsufficientif it contains a relatively low ratio of useful steps\. To quantify this, inspired by counterfactual attribution\(Chenet al\.,[2023](https://arxiv.org/html/2605.19285#bib.bib36)\), we measure the marginal impact of each step by observing the change in the model’s output probability after it is masked\. Formally, given a complete rationale𝐄=\{𝐞l\}l=1L\\mathbf\{E\}=\\\{\\mathbf\{e\}\_\{l\}\\\}\_\{l=1\}^\{L\}and its perturbed version𝐄∖l=𝐄∖\{𝐞l\}\\mathbf\{E\}\_\{\\setminus l\}=\\mathbf\{E\}\\setminus\\\{\\mathbf\{e\}\_\{l\}\\\}with thell\-th step removed, the contributionΔl\\Delta\_\{l\}of step𝐞l\\mathbf\{e\}\_\{l\}to the predictiony^\\widehat\{y\}is defined as follows: \(2\)Δl=logP𝜽\(y^∣𝐱,𝐄\)−logP𝜽\(y^∣𝐱,𝐄∖l\),\\Delta\_\{l\}=\\log P\_\{\\boldsymbol\{\\theta\}\}\\left\(\\widehat\{y\}\\mid\\mathbf\{x\},\\mathbf\{E\}\\right\)\-\\log P\_\{\\boldsymbol\{\\theta\}\}\\left\(\\widehat\{y\}\\mid\\mathbf\{x\},\\mathbf\{E\}\_\{\\setminus l\}\\right\),whereP𝜽\(y^∣⋅\)P\_\{\\boldsymbol\{\\theta\}\}\\left\(\\widehat\{y\}\\mid\\cdot\\right\)denotes the confidence score assigned by the off\-the\-shelf LLM𝜽\\boldsymbol\{\\theta\}to the predictiony^\\widehat\{y\}\. A higherΔl\\Delta\_\{l\}indicates that𝐞l\\mathbf\{e\}\_\{l\}is a critical piece of evidence, whereas aΔl≈0\\Delta\_\{l\}\\approx 0suggests that the step is redundant or even obstructive to the verification process\. ### 2\.1\.Unnecessary and Insufficient Rationales Even Though Correct Predictions Because during the training rationale filtering process, using label correctness is a straightforward and naive approach\. However, for MD, the labels consist solely of real and fake\. Our experiments in this section reveal that samples filtered using such coarse\-grained labels, even when the final predictions are correct, still contain a large number of unnecessary and insufficient rationales\. Figure 2\.Distribution ofΔ\\Deltafor correct / incorrect rationales\.Figure 3\.Distribution of the number of minimal sufficient verification stepsκ\\kappafor correct rationales\.Figure 4\.Precision and recall scores of instruct\-tuned and reasoning LLMs on real and fake classes, respectively\.Correct but unnecessary rationales\.In Fig\.[2](https://arxiv.org/html/2605.19285#S2.F2), we illustrate the distribution of step contributionΔ\\Deltafor correct and incorrect rationales, based on data from Qwen3\-32B and Qwen3\-235B\-A22B, during the fine\-tuning of the Qwen3\-4B\-Instruct\. The experimental results show that average 42\.85% of the steps in correct rationales have aΔ<0\\Delta<0, meaning their contribution to the prediction is actually negative, and the majority of steps have aΔ\\Deltaconcentrated around 0, indicating they provide no contribution to the prediction\. Therefore, we can draw a preliminary experimental conclusion: ⋆\\boldsymbol\{\\star\}Conclusion 1\.Even when theveracity predictions are correct, there still exist a number of rationales containing useless, even negative, verification steps,*i\.e\.,*unnecessary rationales\. Furthermore, in Fig\.[2](https://arxiv.org/html/2605.19285#S2.F2), we also observe that in the incorrectly predicted samples, an average of 41\.55% of the steps have a contributionΔ<0\\Delta<0\. This indicates that these steps have a negative contribution to the incorrect prediction, meaning they actually provide a positive contribution toward the ground\-truth label\. Therefore, another empirical conclusion is thateven when predictions are incorrect, their rationales still contain partially useful steps\. Correct but insufficient rationales\.To evaluate the sufficiency of the rationales, we rank the verification steps within each rationale by their contributions\{Δl\}l=1L\\\{\\Delta\_\{l\}\\\}\_\{l=1\}^\{L\}\. We then select the top\-κ\\kappasteps and compute the label’s prediction probability using only these steps\. A rationale is deemed sufficient if this restricted probability meets a predefined thresholdϵ=0\.01\\epsilon=0\.01relative to the original log\-probability\. Formally, sufficiency is achieved if \(3\)logP𝜽\(y^∣𝐱,𝐄top\-κ\)≥\(1−ϵ\)logP𝜽\(y^∣𝐱,𝐄\),\\log P\_\{\\boldsymbol\{\\theta\}\}\\left\(\\widehat\{y\}\\mid\\mathbf\{x\},\\mathbf\{E\}\_\{\\text\{top\-\}\\kappa\}\\right\)\\geq\(1\-\\epsilon\)\\log P\_\{\\boldsymbol\{\\theta\}\}\\left\(\\widehat\{y\}\\mid\\mathbf\{x\},\\mathbf\{E\}\\right\),where𝐄top\-κ\\mathbf\{E\}\_\{\\text\{top\-\}\\kappa\}denotes the set of theκ\\kappahighest\-contributing steps\. We then quantify the sufficiency of each rationale by calculating the minimumκ\\kapparequired to satisfy this condition\. The resulting distribution ofκ\\kappavalues of correct rationales is illustrated in Fig\.[3](https://arxiv.org/html/2605.19285#S2.F3)\. In practice, when humans judge misinformation, they often only require a very small number of key pieces of evidence to determine its falsity\. However, Fig\.[3](https://arxiv.org/html/2605.19285#S2.F3)shows that rationales achieving sufficiency with only two steps account for approximately 2\.8%, and those achieving sufficiency within four steps account for only about 12\.9%\. Therefore, we can draw an experimental conclusion: ⋆\\boldsymbol\{\\star\}Conclusion 2\.Even when theveracity predictions are correct, the proportion ofrationales that are sufficient\(containing useful steps\)is very small\. ### 2\.2\.Over\-Verification of LLMs In this section, we find that training data generated by advanced LLMs often suffers from the problem of over\-verification, particularly with LLMs possessing stronger reasoning capabilities,*e\.g\.,*gpt\-oss\-120b\(OpenAI,[2025](https://arxiv.org/html/2605.19285#bib.bib37)\)\. This over\-verification issue further increases the proportion of unnecessary rationales\. Over\-verification behavior especially in strong reasoning LLMs\.First, the results in Table[3](https://arxiv.org/html/2605.19285#S4.T3)demonstrate the MD performance of gpt\-oss\-120b, a strong open\-source reasoning LLM, compared to several instruct\-tuned LLMs of similar scale\. The results reveal that this reasoning LLM consistently exhibits inferior detection performance\. More specifically, in Fig\.[4](https://arxiv.org/html/2605.19285#S2.F4), we further compare the precision and recall of two instruct\-tuned LLMs and two reasoning LLMs on the real and fake classes, respectively\. It is clearly observed that reasoning LLMs achieve higher recall on the fake class, indicating a tendency for reasoning LLMs to classify information as fake more frequently\. To investigate the underlying cause of this phenomenon, we provide a representative case in Table[7](https://arxiv.org/html/2605.19285#S4.T7)\. We find that reasoning LLMs generally generate a significantly higher number of verification steps compared to instruct\-tuned LLMs, a finding further corroborated by Fig\.[5](https://arxiv.org/html/2605.19285#S2.F5)\. Therefore, if any single verification step concludes the information is fake, the LLM tends to predict the final label as fake\. This abundance of verification steps leads the LLM to be biased toward the fake prediction, which is referred to as theover\-verification problem\. These analyses induce a conclusion: ⋆\\boldsymbol\{\\star\}Conclusion 3\.Stronger LLMs, particularly reasoning ones, always suffer from an over\-verification problem in MD\. Specifically,their tendency to perform more verification steps causes their predictions to be biased toward the fake class\. Figure 5\.Distribution of the number of verification steps\.More unnecessary rationales due to over\-verification\.The over\-verification problem leads to an increased number of verification steps, which intuitively results in more unnecessary rationales\. To verify this, Fig\.[6](https://arxiv.org/html/2605.19285#S2.F6)shows the distribution of the ratio of unnecessary rationales for gpt\-oss\-120b and a similarly sized instruct\-tuned LLM, Qwen3\-235B\-A22B, formalized as follows: \(4\)ratio=1L∑l=1L𝟙\(logP𝜽\(y^∣𝐱,𝐄\)−logP𝜽\(y^∣𝐱,𝐄∖l\)<ζ\),\\text\{ratio\}=\\frac\{1\}\{L\}\\sum\\nolimits\_\{l=1\}^\{L\}\\mathbbm\{1\}\\left\(\\log P\_\{\\boldsymbol\{\\theta\}\}\(\\widehat\{y\}\\mid\\mathbf\{x\},\\mathbf\{E\}\)\-\\log P\_\{\\boldsymbol\{\\theta\}\}\(\\widehat\{y\}\\mid\\mathbf\{x\},\\mathbf\{E\}\_\{\\setminus l\}\)<\\zeta\\right\),where we fixζ=0\\zeta=0and deem a verification step unnecessary only if it contributes negatively\. As shown in Fig\.[6](https://arxiv.org/html/2605.19285#S2.F6), when fine\-tuning the same LLM, the ratio of unnecessary rationales from gpt\-oss\-120b is substantially higher, indeed, nearly half of all generated rationales consist entirely of useless steps \(ratio≈1\.0\\approx 1\.0\)\. This finding strongly suggests thatthe over\-verification behavior of LLMs leads to an excess of unnecessary rationales\. In summary, we observe that rationales produced by state\-of\-the\-art LLMs are consistently unnecessary and insufficient\. This insight motivates us to design a data filtering metric that selects only those rationales that are both necessary and sufficient for training LLMs\. Figure 6\.Distribution of the ratio of unnecessary rationales\. ## 3\.The ProposedLonsRexMethod In this section, we provide an overview ofLonsRexin Sec\.[3\.1](https://arxiv.org/html/2605.19285#S3.SS1), and the detailed descriptions of its two key scores in Secs\.[3\.2](https://arxiv.org/html/2605.19285#S3.SS2)and[3\.3](https://arxiv.org/html/2605.19285#S3.SS3)\. ### 3\.1\.Overview ofLonsRex The primary idea ofLonsRexis to design a metric to evaluate the necessity and sufficiency of each rationale and then select higher\-quality training data based on this metric\. Generally,LonsRexcomprises two main scoring components: theself\-attribution score, which directly assesses the contribution of each verification step within a rationale to measure its necessity and sufficiency, and themutual\-attribution score, which generates multiple rationales for each claim and evaluates their necessity by comparing their consistency\. The overview ofLonsRexis depicted in Fig\.[7](https://arxiv.org/html/2605.19285#S3.F7)\. Figure 7\.Overview ofLonsRex\. Given a claim, we generateKKrationales using advanced LLMs, which are then filtered by basic heuristics andself\-attributionandmutual\-attributionscores\. Finally, we use the filtered rationales to tune a lightweight LLM\.Basic heuristic filtering\.Formally, given the generated rationales\{\(𝐱i,𝐄ik,y^ik\)\}i,k=1\|𝒟\|,K\\left\\\{\(\\mathbf\{x\}\_\{i\},\\mathbf\{E\}\_\{ik\},\\widehat\{y\}\_\{ik\}\)\\right\\\}\_\{i,k=1\}^\{\|\\mathcal\{D\}\|,K\}formulated in Sec\.[2](https://arxiv.org/html/2605.19285#S2), we first apply several basic heuristic data filtering criteria\. Specifically, wediscardinstances where: \(1\) the predicted answer is incorrect,*i\.e\.,*y^ik≠yi\\widehat\{y\}\_\{ik\}\\neq y\_\{i\}; \(2\) the LLM outputy^ik\\widehat\{y\}\_\{ik\}does not explicitly generate either\\boxed\{real\}or\\boxed\{fake\}; \(3\) the rationale exceeds the token limit of 4,096 and is thus cut\-off; \(4\) the rationale contains unrecognizable characters or unusual special symbols; or \(5\) the rationale exhibits repetitive or degenerate patterns,*e\.g\.,*copycat behavior\. We denote the dataset after this initial filtering as𝒟0=\{\(𝐱i,𝐄i,y^i\)\}i=1\|𝒟0\|\\mathcal\{D\}\_\{0\}=\\left\\\{\(\\mathbf\{x\}\_\{i\},\\mathbf\{E\}\_\{i\},\\widehat\{y\}\_\{i\}\)\\right\\\}\_\{i=1\}^\{\|\\mathcal\{D\}\_\{0\}\|\}\. Self\-attribution score\.Given one instance\(𝐱i,𝐄i,y^i\)∈𝒟0\(\\mathbf\{x\}\_\{i\},\\mathbf\{E\}\_\{i\},\\widehat\{y\}\_\{i\}\)\\in\\mathcal\{D\}\_\{0\}, we aim to calculate a self\-attribution scoreϕs\\phi\_\{\\text\{s\}\}by examining the contribution of each verification step within its rationale𝐄i\\mathbf\{E\}\_\{i\}\. Specifically, following a similar approach to that in Sec\.[2](https://arxiv.org/html/2605.19285#S2), we compute the necessity scoresnecs\_\{\\text\{nec\}\}for each rationale by taking the averageΔ\\Deltaas defined in Eq\. \([2](https://arxiv.org/html/2605.19285#S2.E2)\) across all its verification steps, and we incorporate the ratio of unnecessary steps in Eq\. \([4](https://arxiv.org/html/2605.19285#S2.E4)\) as a penalty term\. The sufficiency scoressufs\_\{\\text\{suf\}\}for each rationale is measured by the probability of its top\-κ\\kappasteps in Eq\. \([3](https://arxiv.org/html/2605.19285#S2.E3)\)\. The final metricϕs=snec×\(1−ssuf\)\\phi\_\{\\text\{s\}\}=s\_\{\\text\{nec\}\}\\times\(1\-s\_\{\\text\{suf\}\}\)combines both the necessity and sufficiency scores\. Mutual\-attribution score\.Given a claim and its correspondingKKrationales\{\(𝐱i,𝐄ik,y^ik\)\}k=1K\\\{\(\\mathbf\{x\}\_\{i\},\\mathbf\{E\}\_\{ik\},\\widehat\{y\}\_\{ik\}\)\\\}\_\{k=1\}^\{K\}generated by multiple LLMs, we compute the mutual\-attribution scoreϕm\\phi\_\{\\text\{m\}\}by comparing these rationales\. Specifically, this metric follows a basic assumption:if a high\-contribution verification perspective frequently appears across different rationales for the same claim, it is more likely to be necessary\. Accordingly, we cluster the verification steps within the rationales\{𝐄ik\}k=1K\\\{\\mathbf\{E\}\_\{ik\}\\\}\_\{k=1\}^\{K\}intoMMdifferent verification perspectives\{𝐯m\}m=1M\\\{\\mathbf\{v\}\_\{m\}\\\}\_\{m=1\}^\{M\},*e\.g\.,*source verification and scientific validity assessment, and then compute the mutual information between each perspective and the predicted label to evaluate its contribution\. A rationale that incorporates a greater number of high\-contribution perspectives is deemed more necessary and is therefore assigned a higherϕm\\phi\_\{\\text\{m\}\}score\. Finally, we filter samples by\(ϕs\+ϕm\)/2\(\\phi\_\{\\text\{s\}\}\+\\phi\_\{\\text\{m\}\}\)/2to produce the final dataset𝒟^=\{\(𝐱i,𝐄i,y^i\)\}i=1\|𝒟^\|\\mathcal\{\\widehat\{D\}\}=\\\{\(\\mathbf\{x\}\_\{i\},\\mathbf\{E\}\_\{i\},\\widehat\{y\}\_\{i\}\)\\\}\_\{i=1\}^\{\|\\mathcal\{\\widehat\{D\}\}\|\}, and fine\-tune LLMs by Eq\. \([1](https://arxiv.org/html/2605.19285#S2.E1)\)\. In the following sections, we describe the details of self\-attribution and mutual\-attribution scores, respectively\. ### 3\.2\.Self\-Attribution Score The goal of the self\-attribution score is to quantify whether a rationale provides a verification process that is both necessary and sufficient for supporting its prediction\. Given an instance\(𝐱,𝐄,y^\)\(\\mathbf\{x\},\\mathbf\{E\},\\widehat\{y\}\), where𝐱\\mathbf\{x\}is the input claim,y^\\widehat\{y\}is the predicted veracity label, and𝐄=\{𝐞il\}l=1\|𝐄\|\\mathbf\{E\}=\\\{\\mathbf\{e\}\_\{il\}\\\}\_\{l=1\}^\{\|\\mathbf\{E\}\|\}denotes the rationale containing the sequence of verification steps, we aim to compute a self\-attribution scoreϕs\(𝐄\)\\phi\_\{\\text\{s\}\}\(\\mathbf\{E\}\)by evaluating the contribution of each verification step within𝐄\\mathbf\{E\}\. Following the counterfactual attribution formulation in Sec\.[2](https://arxiv.org/html/2605.19285#S2), we measure the contribution of each verification step by removing it from the rationale and observing the changeΔl\\Delta\_\{l\}in prediction confidence as in Eq\. \([2](https://arxiv.org/html/2605.19285#S2.E2)\)\. Intuitively, a largerΔl\\Delta\_\{l\}indicates that the verification step is more critical to the prediction\. Necessity score\.We first evaluate whether a rationale is necessary by assessing whether it contains relatively few low\-contribution verification steps\. The necessity score is initially defined as the average contribution across all verification steps as \(5\)snec\(𝐄\)=max\(0,1\|𝐄\|∑l=1\|𝐄\|Δi,l\)\.s\_\{\\text\{nec\}\}\(\\mathbf\{E\}\)=\\max\\Big\(0,\\frac\{1\}\{\|\\mathbf\{E\}\|\}\\sum\\nolimits\_\{l=1\}^\{\|\\mathbf\{E\}\|\}\\Delta\_\{i,l\}\\Big\)\. However, a rationale may still contain redundant or detrimental steps\. To penalize such cases, we introduce an unnecessary\-step ratio, which measures the proportion of steps whose contribution is below a thresholdζ=0\\zeta=0\. The final necessity score is adjusted as \(6\)snec\(𝐄\)←snec\(𝐄\)⋅\(1−1\|𝐄\|∑l=1\|𝐄\|𝟙\(Δl<ζ\)\)\.s\_\{\\text\{nec\}\}\(\\mathbf\{E\}\)\\leftarrow s\_\{\\text\{nec\}\}\(\\mathbf\{E\}\)\\cdot\\Big\(1\-\\frac\{1\}\{\|\\mathbf\{E\}\|\}\\sum\\nolimits\_\{l=1\}^\{\|\\mathbf\{E\}\|\}\\mathbbm\{1\}\\left\(\\Delta\_\{l\}<\\zeta\\right\)\\Big\)\. Sufficiency score\.Besides necessity, a high\-quality rationale should contain a compact subset of steps that alone can support the prediction\. We therefore identify the top\-κ\\kappasteps ranked byΔi,l\\Delta\_\{i,l\}, denoted as𝐄top\-κ\\mathbf\{E\}\_\{\\text\{top\-\}\\kappa\}, whereκ\\kappais empirically fixed to33\. The sufficiency score is defined as the prediction confidence using only these steps: \(7\)ssuf\(𝐄\)=logP𝜽\(y^∣𝐱,𝐄top\-κ\)−logP𝜽\(y^∣𝐱,𝐄\)\.s\_\{\\text\{suf\}\}\(\\mathbf\{E\}\)=\\log P\_\{\\boldsymbol\{\\theta\}\}\\left\(\\widehat\{y\}\\mid\\mathbf\{x\},\\mathbf\{E\}\_\{\\text\{top\-\}\\kappa\}\\right\)\-\\log P\_\{\\boldsymbol\{\\theta\}\}\(\\widehat\{y\}\\mid\\mathbf\{x\},\\mathbf\{E\}\)\.We aim to identify aminimal sufficient rationale,*i\.e\.,*a higherssufs\_\{\\text\{suf\}\}indicates that a small subset of steps is already sufficient for the prediction, suggesting that the remaining steps are potentially redundant\. Finally, we combine necessity and sufficiency to quantify the overall quality of the rationale: \(8\)ϕs\(𝐄\)=snec\(𝐄\)⋅\(1−ssuf\(𝐄\)\)\.\\phi\_\{\\text\{s\}\}\(\\mathbf\{E\}\)=s\_\{\\text\{nec\}\}\(\\mathbf\{E\}\)\\cdot\\big\(1\-s\_\{\\text\{suf\}\}\(\\mathbf\{E\}\)\\big\)\.This formulation encourages rationales whose verification steps are consistently important while discouraging those containing redundant or easily compressible verification processes\. ### 3\.3\.Mutual\-Attribution Score Given a claim𝐱\\mathbf\{x\}, we assume that LLMs generateKKrationales and corresponding veracity predictions\{𝐄k,y^k\}k=1K\\\{\\mathbf\{E\}\_\{k\},\\widehat\{y\}\_\{k\}\\\}\_\{k=1\}^\{K\}\. Each rationale is represented as a sequence of verification steps𝐄k=\{𝐞kl\}l=1\|𝐄k\|\\mathbf\{E\}\_\{k\}=\\\{\\mathbf\{e\}\_\{kl\}\\\}\_\{l=1\}^\{\|\\mathbf\{E\}\_\{k\}\|\}\. For each step𝐞kl\\mathbf\{e\}\_\{kl\}, wherek∈\{1,…,K\}k\\in\\\{1,\\ldots,K\\\}andl∈\{1,…,\|𝐄k\|\}l\\in\\\{1,\\ldots,\|\\mathbf\{E\}\_\{k\}\|\\\}, we obtain its semantic representation𝐡kl\\mathbf\{h\}\_\{kl\}using a lightweight encoder,*e\.g\.,*BERT\. We then useKK\-means clustering to cluster all step representations intoMMverification perspectives𝒱=\{𝐯m\}m=1M\\mathcal\{V\}=\\\{\\mathbf\{v\}\_\{m\}\\\}\_\{m=1\}^\{M\}, for example, evaluating the reliability of the information source or consistency with commonsense knowledge\. Each rationale can thus be denoted as a subset of perspectives,*i\.e\.,*𝐄k⊆𝒱\\mathbf\{E\}\_\{k\}\\subseteq\\mathcal\{V\}\. Based on these perspectives, the mutual\-attribution score aims to assess the contribution of each verification perspective and considers rationales that contain a greater number of high\-contribution perspectives to be more necessary\. To quantify the contribution of a verification perspective𝐯m\\mathbf\{v\}\_\{m\}to the predictiony^\\widehat\{y\}, we measure the prediction change when removing this perspective: \(9\)Δm,k=logP𝜽\(y^k∣𝐱,𝐄k\)−logP𝜽\(y^k∣𝐱,𝐄k∖\{𝐯m\}\)\.\\Delta\_\{m,k\}=\\log P\_\{\\boldsymbol\{\\theta\}\}\\left\(\\widehat\{y\}\_\{k\}\\mid\\mathbf\{x\},\\mathbf\{E\}\_\{k\}\\right\)\-\\log P\_\{\\boldsymbol\{\\theta\}\}\\left\(\\widehat\{y\}\_\{k\}\\mid\\mathbf\{x\},\\mathbf\{E\}\_\{k\}\\setminus\\\{\\mathbf\{v\}\_\{m\}\\\}\\right\)\. Let𝒦m=\{k∣𝐯m∈𝐄k\}\\mathcal\{K\}\_\{m\}=\\\{k\\mid\\mathbf\{v\}\_\{m\}\\in\\mathbf\{E\}\_\{k\}\\\}denote the set of rationales containing the perspective𝐯m\\mathbf\{v\}\_\{m\}\. We further define an indicator random variable \(10\)𝒜m=\{1,𝐯m∈𝐄k,0,otherwise\.\\mathcal\{A\}\_\{m\}=\\begin\{cases\}1,&\\mathbf\{v\}\_\{m\}\\in\\mathbf\{E\}\_\{k\},\\\\ 0,&\\text\{otherwise\}\.\\end\{cases\}Its empirical occurrence probability is estimated as \(11\)P\(𝒜m=1\)≈\|𝒦m\|K\.P\(\\mathcal\{A\}\_\{m\}=1\)\\approx\\frac\{\|\\mathcal\{K\}\_\{m\}\|\}\{K\}\. We define the importance of each perspective by thejoint mutual informationbetween the prediction outcome and the perspective occurrence together with its contribution: \(12\)ϕ\(𝐯m\)=ℐ\(Y;𝒜m,Δm\)\.\\phi\(\\mathbf\{v\}\_\{m\}\)=\\mathcal\{I\}\(Y;\\mathcal\{A\}\_\{m\},\\Delta\_\{m\}\)\.To implement it, we first use the chain rule of mutual information, \(13\)ϕ\(𝐯m\)=ℐ\(Y;𝒜m,Δm\)=ℐ\(Y;𝒜m\)\+ℐ\(Y;Δm∣𝒜m\),\\phi\(\\mathbf\{v\}\_\{m\}\)=\\mathcal\{I\}\(Y;\\mathcal\{A\}\_\{m\},\\Delta\_\{m\}\)=\\mathcal\{I\}\(Y;\\mathcal\{A\}\_\{m\}\)\+\\mathcal\{I\}\(Y;\\Delta\_\{m\}\\mid\\mathcal\{A\}\_\{m\}\),where the first term is \(14\)ℐ\(Y;𝒜m\)\\displaystyle\\mathcal\{I\}\(Y;\\mathcal\{A\}\_\{m\}\)=𝔼𝒜m,Y\[logP\(Y∣𝒜m\)P\(Y\)\]\\displaystyle=\\mathbb\{E\}\_\{\\mathcal\{A\}\_\{m\},Y\}\\left\[\\log\\frac\{P\(Y\\mid\\mathcal\{A\}\_\{m\}\)\}\{P\(Y\)\}\\right\]∝P\(𝒜m=1\)𝔼Y∣𝒜m=1\[logP\(Y∣𝒜m=1\)\],\\displaystyle\\propto P\(\\mathcal\{A\}\_\{m\}=1\)\\mathbb\{E\}\_\{Y\\mid\\mathcal\{A\}\_\{m\}=1\}\\left\[\\log P\(Y\\mid\\mathcal\{A\}\_\{m\}=1\)\\right\],and the second term is ℐ\(Y;Δm∣\\displaystyle\\mathcal\{I\}\(Y;\\Delta\_\{m\}\\mid𝒜m\)=P\(𝒜m=1\)𝔼Y,Δm∣𝒜m=1\[logP\(Y∣Δm,𝒜m=1\)P\(Y∣𝒜m=1\)\]\\displaystyle\\mathcal\{A\}\_\{m\}\)=P\(\\mathcal\{A\}\_\{m\}=1\)\\mathbb\{E\}\_\{Y,\\Delta\_\{m\}\\mid\\mathcal\{A\}\_\{m\}=1\}\\left\[\\log\\frac\{P\(Y\\mid\\Delta\_\{m\},\\mathcal\{A\}\_\{m\}=1\)\}\{P\(Y\\mid\\mathcal\{A\}\_\{m\}=1\)\}\\right\]=P\(𝒜m=1\)𝔼Y,Δm∣𝒜m=1\[logP\(Y∣Δm,𝒜m=1\)\]\\displaystyle=P\(\\mathcal\{A\}\_\{m\}=1\)\\mathbb\{E\}\_\{Y,\\Delta\_\{m\}\\mid\\mathcal\{A\}\_\{m\}=1\}\[\\log P\(Y\\mid\\Delta\_\{m\},\\mathcal\{A\}\_\{m\}=1\)\]\(15\)−P\(𝒜m=1\)𝔼Y∣𝒜m=1\[logP\(Y∣𝒜m=1\)\]\.\\displaystyle\\quad\-P\(\\mathcal\{A\}\_\{m\}=1\)\\mathbb\{E\}\_\{Y\\mid\\mathcal\{A\}\_\{m\}=1\}\[\\log P\(Y\\mid\\mathcal\{A\}\_\{m\}=1\)\]\.By summing the two terms, the baseline likelihood cancels out, yielding \(16\)ϕ\(𝐯m\)=P\(𝒜m=1\)𝔼Y,Δm∣𝒜m=1\[logP\(Y∣Δm,𝒜m=1\)\]\.\\phi\(\\mathbf\{v\}\_\{m\}\)=P\(\\mathcal\{A\}\_\{m\}=1\)\\mathbb\{E\}\_\{Y,\\Delta\_\{m\}\\mid\\mathcal\{A\}\_\{m\}=1\}\\left\[\\log P\(Y\\mid\\Delta\_\{m\},\\mathcal\{A\}\_\{m\}=1\)\\right\]\.Assuming that the prediction depends on the perspective contribution only throughΔm\\Delta\_\{m\}, we approximate \(17\)P\(Y∣Δm,𝒜m=1\)≈P\(Y∣Δm\)\.P\(Y\\mid\\Delta\_\{m\},\\mathcal\{A\}\_\{m\}=1\)\\approx P\(Y\\mid\\Delta\_\{m\}\)\.Since the true conditional distribution is unknown, we introduce a variational approximationQ𝚽\(Y∣Δm\)Q\_\{\\boldsymbol\{\\Phi\}\}\(Y\\mid\\Delta\_\{m\}\), which yields the evidence lower bound as \(18\)ϕ\(𝐯m\)≥P\(𝒜m=1\)𝔼k∈𝒦m\[logQ𝚽\(Y=y^k∣Δm,k\)\]\.\\phi\(\\mathbf\{v\}\_\{m\}\)\\geq P\(\\mathcal\{A\}\_\{m\}=1\)\\mathbb\{E\}\_\{k\\in\\mathcal\{K\}\_\{m\}\}\\left\[\\log Q\_\{\\boldsymbol\{\\Phi\}\}\\left\(Y=\\widehat\{y\}\_\{k\}\\mid\\Delta\_\{m,k\}\\right\)\\right\]\.Following a logistic stability assumption, we model \(19\)Q𝚽\(Y=y^k∣Δm,k\)=\(1\+e−Δm,k\)−1\.Q\_\{\\boldsymbol\{\\Phi\}\}\\left\(Y=\\widehat\{y\}\_\{k\}\\mid\\Delta\_\{m,k\}\\right\)=\\left\(1\+e^\{\-\\Delta\_\{m,k\}\}\\right\)^\{\-1\}\.Substituting this form and using empirical estimation, we obtain \(20\)ϕ\(𝐯m\)\\displaystyle\\phi\(\\mathbf\{v\}\_\{m\}\)=−\|𝒦m\|K⋅1\|𝒦m\|∑k∈𝒦mlog\(1\+e−Δm,k\)\\displaystyle=\-\\frac\{\|\\mathcal\{K\}\_\{m\}\|\}\{K\}\\cdot\\frac\{1\}\{\|\\mathcal\{K\}\_\{m\}\|\}\\sum\\nolimits\_\{k\\in\\mathcal\{K\}\_\{m\}\}\\log\\left\(1\+e^\{\-\\Delta\_\{m,k\}\}\\right\)=−1K∑k∈𝒦mlog\(1\+e−Δm,k\)\.\\displaystyle=\-\\frac\{1\}\{K\}\\sum\\nolimits\_\{k\\in\\mathcal\{K\}\_\{m\}\}\\log\\left\(1\+e^\{\-\\Delta\_\{m,k\}\}\\right\)\. Finally, the importance of a rationale is computed by aggregating the importance of its verification perspectives: \(21\)ϕm\(𝐄k\)=∑𝐯m∈𝐄kϕ\(𝐯m\)\.\\phi\_\{m\}\(\\mathbf\{E\}\_\{k\}\)=\\sum\\nolimits\_\{\\mathbf\{v\}\_\{m\}\\in\\mathbf\{E\}\_\{k\}\}\\phi\(\\mathbf\{v\}\_\{m\}\)\. ## 4\.Experimental Evaluation In this section, we evaluate the LLM fine\-tuned using our proposedLonsRexpipeline on their performance in MD\. ### 4\.1\.Experimental Settings Training datasets\.To train a specialized LLM for explainable MD, we collect a total of 316k claims and their corresponding veracity labels from publicly available repositories\. Specifically, the sources and quantities of these claims are detailed in Table[1](https://arxiv.org/html/2605.19285#S4.T1)\. After downloading the raw claims, we perform a simple deduplication step: a claim was considered a duplicate if its first 100 tokens matched those of an existing claim in the dataset\. Table 1\.Sources and quantities of the training dataset\.Evaluation dataset\.To prevent data contamination, where evaluation samples inadvertently appear in the training data, we evaluate LLMs using the most recent MD benchmarks\.MultiDisandEUDisinfoare collected byModzelewskiet al\.\([2025](https://arxiv.org/html/2605.19285#bib.bib3)\)and contain only claims published after January 2024\.GossipCop\+\+andPolitiFact\+\+are compiled bySuet al\.\([2023](https://arxiv.org/html/2605.19285#bib.bib55)\)and include both human\-written and machine\-generated claims, and we use only the human\-written subset\. Statistics for these datasets are provided in Table[2](https://arxiv.org/html/2605.19285#S4.T2)\. Table 2\.Experimental results on three LLMs fine\-tuned byLonsRexacross four MD datasets\.Baselines\.We compare against 6 prompt methods specifically designed for MD, evaluating their performance across LLMs of varying sizes\. These methods are as follows: Vanilla directly prompts LLMs to make a veracity judgment; CoT\(Weiet al\.,[2022](https://arxiv.org/html/2605.19285#bib.bib44)\)instructs LLMs to reason step\-by\-step before generating its prediction; ARG\(Huet al\.,[2024](https://arxiv.org/html/2605.19285#bib.bib16)\)uses LLMs to generate textual and commonsense descriptions related to the claim, which are then provided as auxiliary context for veracity prediction;GenFend\(Nanet al\.,[2024](https://arxiv.org/html/2605.19285#bib.bib39)\)constructs diverse user personas and simulates user\-generated comments to inform the prediction; DMR\(Jinet al\.,[2025b](https://arxiv.org/html/2605.19285#bib.bib54)\)prompts LLMs to extract the key knowledge required to verify the claim\. PCoT\(Modzelewskiet al\.,[2025](https://arxiv.org/html/2605.19285#bib.bib3)\)first asks LLMs to identify the persuasive strategy employed in the claim, then uses this analysis to guide the veracity assessment\. The exact prompts used for each method are provided in our code repository\. Implementation details\.During data synthesis, for the collected set of 316k claims, we employ three LLMs: Qwen3\-32B, Qwen3\-Next\-80B\-A3B, and Qwen3\-235B\-A22B, as rationale generators\. These LLMs are selected for their favorable balance between strong performance and efficient reasoning speed \(see Table[8](https://arxiv.org/html/2605.19285#A0.T8)\)\. Each LLM generates one distinct rationale per claim using a temperature of 0\.6\. Both data generation and model evaluation are conducted using[SGLang](https://github.com/sgl-project/sglang)for model deployment and inference, with the maximum output length capped at 32,768 tokens\. During model fine\-tuning, we performe full\-parameter fine\-tuning using the[LlamaFactory](https://github.com/hiyouga/LlamaFactory)framework, setting the packing length to 32,768 to accelerate training\. We use a batch size of 64 and optimize the LLM for 6 epochs with the Adam optimizer at a learning rate of5×10−55\\times 10^\{\-5\}\. We also experimented with training for 15 epochs, which yielded performance comparable to that of the 6\-epoch setting\. During evaluation, we prompt LLMs to generate the tokens\\boxed\{real\}or\\boxed\{fake\}, and directly match these outputs against the ground\-truth labels to compute accuracy and F1 score\. All LLM deployment and training were conducted on 8 NVIDIA A100 GPUs\. ### 4\.2\.Main Results To evaluate the effectiveness of the data generated by ourLonsRexpipeline in improving foundation LLM performance, we train three lightweight LLMs,*e\.g\.,*Qwen3\-4B\-Instruct, Qwen2\.5\-1\.5B\-Instruct, and Gemma2\-2B\-it, on our curated dataset\. We then evaluate their performance under carefully designed prompting strategies, including CoT\(Weiet al\.,[2022](https://arxiv.org/html/2605.19285#bib.bib44)\), ARG\(Huet al\.,[2024](https://arxiv.org/html/2605.19285#bib.bib16)\),GenFend\(Nanet al\.,[2024](https://arxiv.org/html/2605.19285#bib.bib39)\), DMR\(Jinet al\.,[2025b](https://arxiv.org/html/2605.19285#bib.bib54)\), and PCoT\(Modzelewskiet al\.,[2025](https://arxiv.org/html/2605.19285#bib.bib3)\)\. The results across four prevalent MD benchmark datasets are reported in Table[2](https://arxiv.org/html/2605.19285#S4.T2)\. Generally,LonsRexconsistently outperforms all baselines\. As shown in Table[2](https://arxiv.org/html/2605.19285#S4.T2), it achieves the highest accuracy and F1 scores across all three foundation LLMs\. Specifically,LonsRex\-4B achieves an average improvement of approximately 19\.78% over the CoT baseline\. Moreover, the data generated byLonsRexachieves consistently strong performance across LLMs of three different scales, demonstrating the robustness of our data in effectively instilling verification capabilities\. Table 3\.Performance compared with SOTA LLMs\.LLMGossipCopPolitiFactMultiDisEUDisinfo[Qwen3\-32B](https://huggingface.co/Qwen/Qwen3-32B)62\.5386\.9686\.5776\.11[Llama\-3\.1\-70B](https://huggingface.co/meta-llama/Llama-3.1-70B-Instruct)63\.7082\.4785\.2279\.32[gpt\-oss\-120b](https://huggingface.co/openai/gpt-oss-120b)58\.1057\.1365\.6864\.83[DeepSeek\-V3\.2](https://huggingface.co/deepseek-ai/DeepSeek-V3.2)65\.1589\.1589\.0778\.87LonsRex\-4B67\.9887\.1192\.2885\.03LonsRex\-2B67\.1281\.5389\.5878\.34LonsRex\-1\.5B68\.5084\.9790\.6383\.15 ### 4\.3\.Compare with SOTA LLMs In this section, we compare our fine\-tuned lightweight LLMs against significantly larger open\-source LLMs,*e\.g\.,*Qwen3\-32B, Llama\-3\.1\-70B, gpt\-oss\-120b, and DeepSeek\-V3\.2\. The results are presented in Table[3](https://arxiv.org/html/2605.19285#S4.T3)\. The results reveal a compelling finding: our fine\-tuned LLMs significantly outperform Llama\-3\.1\-70B and gpt\-oss\-120b, and are comparable with DeepSeek\-V3\.2, despite having less than 5% of the parameters\. Notably, the large reasoning model gpt\-oss\-120b performs poorly \(*e\.g\.,*57\.13% onPolitiFact\+\+\)\. As analyzed in Sec\.[2\.2](https://arxiv.org/html/2605.19285#S2.SS2), this is attributed to the over\-verification behavior, where strong reasoning models tend to generate excessive verification steps, leading to a bias towards classifying claims as fake\.LonsRexmitigates this by explicitly filtering for necessary and sufficient rationales, preventing the model from learning these redundant, noise\-inducing reasoning patterns\. ### 4\.4\.Training Across Different Datasets We also investigate the impact of different data synthesis sources and filtering strategies\. We compareLonsRexagainst three variants: \(1\)w/ ALL, using all generated rationales only filtered by correctness, \(2\) using rationales generated by a single LLM, filtered only by correctness, and \(3\)w/ only label, exclusively using the label format“This message is\[Label\]\.”, without rationales\. Generally, our method consistently outperforms all ablation variants\. Specifically, compared to theall datasetting, which uses nearly three times more training examples,LonsRexachieves significantly better performance, particularly on the 4B\-scale LLM\. This highlights that the higher quality of data produced byLonsRexis more effective than simply increasing data scale\. It also indicates that filtering solely based on correctness leaves substantial noise in the rationales\. In contrast to using multiple source LLMs directly, even though our data originates from these same LLMs, our pipeline, with its careful filtering strategy, yields fine\-tuned LLMs that consistently surpass these source LLMs\. This further underscores the importance of rigorous data curation\. Finally, LLMs trained only with the label perform poorly, demonstrating that LLMs cannot reliably infer the underlying relationship between inputs and labels\. In MD, explicitly providing explainable rationales proves far more effective than relying on implicit learning from labels alone\. Table 4\.Performance compared with different datasets\. ### 4\.5\.Token Consumption To evaluate the efficiency ofLonsRex, we explicitly compare the average token consumption per query during inference in Table[5](https://arxiv.org/html/2605.19285#S4.T5)\. While it consumes more tokens than vanilla prompting due to the generation of rationales, it is significantly more efficient than complex CoT or knowledge\-based methods like DMR, which consumes 2̃596 tokens\. By internalizing the verification steps,LonsRexavoids the high computational overhead of multi\-turn queries and reading extensive external information, making it suitable for real\-time MD applications\. ### 4\.6\.Evaluation of Generated Rationales Finally, we assess the quality of the rationales generated byLonsRex\. We followWanget al\.\([2024b](https://arxiv.org/html/2605.19285#bib.bib20)\)in employing automated metrics using Qwen3\-235B\-A22B to evaluate the LLM\-generated rationales, includingMisleadingness\(M\),Informativeness\(I\), andReadability\(R\)\. Misleadingness measures the degree to which the generated rationale aligns with the ground\-truth veracity label, scored on a 1 \(not misleading\) – 5 \(very misleading\) scale\. Informativeness assesses whether the rationale provides new knowledge beyond the claim itself, also on a 1 \(not informative\) – 5 \(very informative\) scale\. Readability evaluates the fluency and readability of the rationale, again using a 1 \(poor\) – 5 \(excellent\) scale\. As shown in Table[6](https://arxiv.org/html/2605.19285#S4.T6),LonsRex\-4B achieves rationale quality scores that are competitive with, and occasionally superior to, the teacher LLMs,*e\.g\.,*Qwen3\-32B\. This indicates that our fine\-tuning process does not merely memorize labels but learns to construct coherent, logically sound verification chains\. Additionally, some representative cases are shown in Table[7](https://arxiv.org/html/2605.19285#S4.T7)\. Table 5\.Token consumption of different methods\.Table 6\.Evaluation of rationales generated by different LLMs\.Table 7\.A representative case from theMultiDisdataset generated by different LLMs\.Article: This wouldn’t be happening without climate change’, saidthe head of the United States’ Coral Reef Watch ProgramOcean temperatures that have gone haywire could make the current global coral bleaching the worst in history\. …More than62\.9%of the world’s coral reefs are badly hurting from a bleaching event that began last year…nearing the record of 65\.7% in 2017… …99\.7% of the coral reefshave been hit with ”very very severe” losses in staghorn and elkhorn species… …Thailand shutting off a tourist\-laden islandto try to save the coral there\. Meteorologists say aLa Nina… forecast to develop soon …, but Manzello said it may betoo little and too late\. “I still am very worried…” Manzello said\. “This wouldn’t be happening without climate change…” Manzello said\. On top of that are changes inEl Nino…;reduced sulfur pollution from shipsandan undersea volcano eruption\. Former top NASA climate scientistJames Hansensaid “acceleration of global warming is now hard to deny”… …scientists are hoping fora few hurricanesas they cool the oceans… …the world’s oceans last month broke a recordfor the hottest April on record… Coral reefs are key to seafood production…loss of coral is one of the big tipping points…countries agreed to try to hold toin the 2015 Paris climate agreement\.Llama3\.1\-70B\(534 tokens\): The message appears to be an article from a reputable source, citingNOAAandClimate Central\. The information is specific, withquotes from expertsadding credibility\. Mentions recent records and trends likethe hottest April on record, suggesting the info is up\-to\-date\. … No obvious contradictions found, therefore:realgpt\-oss\-120b\(1048 tokens\): We need to determine if the message is likely real or fake\. The content is a news\-like article quoting NOAA coral reef watch program head Derek Manzello… Check for inconsistencies:”head of the United States’ Coral Reef Watch Program”— NOAA’s Coral Reef Watch is a program, but the head is likely a scientist; Derek Manzello is a NOAA coral reef watch coordinator\. The quotes seem plausible\. However, some statements:”99\.7% of the coral reefs have been hit…”seems extreme; also”Thailand shutting off a tourist\-laden island”… plausible but needs checking\. …Also”reduced sulfur pollution from ships”and”an undersea volcano eruption”— sentence is odd mixing topics\. …The phrase”very very severe”seems informal\. Also”the reverse of La Nina”is simplistic\. Thus likelyfake\.Qwen3\-4B\-Think\-2507\(5,363 tokens\): Okay, step by step… Claims:62\.9% of world’s coral reefsbadly hurt \(nearing65\.7% in 2017\);99\.7% of Atlantic reefs”very very severe” losses;Thailand shutting off island;La Niña forecast too little, too late;hurricanes to cool reefs;hottest April on record\.NOAAhas Coral Reef Watch program; data shows major bleaching\. 62\.9% figure plausible\.99\.7% figure catches my eye— extremely high\. NOAA data shows Florida severe damage butnot 99\.7%\. …62 countries seeing damaged coralmaybe high\. Scientists hope for hurricanes — plausible\. Thailand closure real\. Key issue:99\.7% likely fabricated to alarm people\. NOAA uses ”over 50%” or ”severe in many areas” not exact 99\.7%\. Given overall alignment but99\.7% unsupported, leaningfake\.LonsRex\-4B\(678 tokens\): Let’s analyze this message step by step for veracity, using known scientific facts, credible sources, and plausibility\. 1\. Source Attribution: Derek Manzello, NOAA Coral Reef Watch Program \- Fact Check: Derek Manzellois a real personand was indeed the Coordinator ofNOAA’s Coral Reef Watch Programduring the relevant time period \(2016–2018\)\. He frequently appeared in media briefings and scientific reports on coral bleaching\. \- Verdict: Real\. 2\. “This wouldn’t be happening without climate change” \- This is a direct quote that matchesManzello’s public statements during the 2016–2017 mass coral bleaching events\. \-NOAA,IPCC, and other scientific bodies have consistently stated that mass coral bleaching events are driven by ocean warming due toanthropogenic climate change\. \- Verdict: Accurate\. … Conclusion: Every claim in the message: \- Is attributed toreal people\(Manzello,Hansen,McNoldy,Klotzbach\), \- Matchesverified scientific reportsfromNOAA,IPCC, and peer\-reviewed literature, \- Usesaccurate statistics from 2016–2017, \- Reflectsactual public statementsmade during that time\. There are no exaggerations, no fabrications, and no misleading edits\. Final Answer:real ## 5\.Related Works In this section, we summarize the related literature on LLM\-based misinformation detection and advanced CoT reasoning techniques\. ### 5\.1\.LLM\-based Misinformation Detection Typically, previous MD methods focus on trainingsmall discriminative modelsthat leverage external features,*e\.g\.,*linguistic style\(Przybyla,[2020](https://arxiv.org/html/2605.19285#bib.bib32)\), propagation patterns\(Bianet al\.,[2020](https://arxiv.org/html/2605.19285#bib.bib35); Wu and Hooi,[2023](https://arxiv.org/html/2605.19285#bib.bib33)\), and more modalities\(Qiet al\.,[2023](https://arxiv.org/html/2605.19285#bib.bib34)\)\. However, these models heavily depend on labeled data and possess limited world knowledge, making them hard to handle rapidly evolving, out\-of\-distribution news events\. Recently, partial studies have begun leveraginglarge generative modelsto aid in misinformation detection and explanation\. One line of research leverages LLMs to retrieve or synthesize external information that augments small discriminative models in their decision\-making\(Huet al\.,[2024](https://arxiv.org/html/2605.19285#bib.bib16); Nanet al\.,[2024](https://arxiv.org/html/2605.19285#bib.bib39); Wanget al\.,[2025b](https://arxiv.org/html/2605.19285#bib.bib17)\)\. For example, ARG\(Huet al\.,[2024](https://arxiv.org/html/2605.19285#bib.bib16)\)generates textual descriptions and commonsense rationales, and DELL\(Wanet al\.,[2024](https://arxiv.org/html/2605.19285#bib.bib38)\)simulates user comments to serve as auxiliary knowledge that aids smaller detectors\. Another line employs carefully crafted prompts or multi\-agent collaboration to enable LLMs to directly produce veracity judgments and explanatory justifications\(Honget al\.,[2025](https://arxiv.org/html/2605.19285#bib.bib6); Modzelewskiet al\.,[2025](https://arxiv.org/html/2605.19285#bib.bib3); Tonget al\.,[2025](https://arxiv.org/html/2605.19285#bib.bib42)\)\. For example, TED\(Liuet al\.,[2025a](https://arxiv.org/html/2605.19285#bib.bib41)\)orchestrates multiple agents in a structured debate process, and DYNAMO\(Jinet al\.,[2025a](https://arxiv.org/html/2605.19285#bib.bib40)\)decomposes complex articles into easily verifiable sub\-questions through Monte Carlo tree search, thereby enabling deep logical reasoning\. ### 5\.2\.Long Chain\-of\-Thought Reasoning Recently, the introduction of LLMs has shifted research focus toward enhancing long chain\-of\-thought reasoning\(Yuanet al\.,[2024](https://arxiv.org/html/2605.19285#bib.bib60); Guoet al\.,[2025](https://arxiv.org/html/2605.19285#bib.bib4); Yanget al\.,[2026](https://arxiv.org/html/2605.19285#bib.bib58)\)\. This paradigm involves generating reasoning trajectories prior to producing a final answer\(Weiet al\.,[2022](https://arxiv.org/html/2605.19285#bib.bib44); Kojimaet al\.,[2022](https://arxiv.org/html/2605.19285#bib.bib45)\), which closely mirrors explainable MD\. Existing pipelines typically leverage supervised fine\-tuning and reinforcement learning techniques\. In reinforcement learning\-based frameworks, meticulously designed rewards,*e\.g\.,*those utilizing label correctness as a signal\(Shaoet al\.,[2024](https://arxiv.org/html/2605.19285#bib.bib43)\), are employed to incentivize LLMs to generate reasoning trajectories\. However, in explainable MD, where ground\-truth labels are binary, such rewards are inherently sparse\. This often results in training instability and may even lead to degraded performance\(Duet al\.,[2023](https://arxiv.org/html/2605.19285#bib.bib48)\)\. To address these limitations, our work explores an alternative direction that aggregates reasoning traces generated by diverse models and develops a filtering mechanism to extract high\-quality trajectories for the supervised fine\-tuning of LLMs\(Hoet al\.,[2023](https://arxiv.org/html/2605.19285#bib.bib52); Wanget al\.,[2023](https://arxiv.org/html/2605.19285#bib.bib53); Kouet al\.,[2025](https://arxiv.org/html/2605.19285#bib.bib57); Yuanet al\.,[2026](https://arxiv.org/html/2605.19285#bib.bib59)\)\. For example, existing approaches often employ heuristic rules,*e\.g\.,*response format\(Muennighoffet al\.,[2025](https://arxiv.org/html/2605.19285#bib.bib49); Huanget al\.,[2025](https://arxiv.org/html/2605.19285#bib.bib61)\), question difficulty\(Yeet al\.,[2025](https://arxiv.org/html/2605.19285#bib.bib50)\), or answer accuracy\(Leiet al\.,[2025](https://arxiv.org/html/2605.19285#bib.bib51); Chenet al\.,[2025](https://arxiv.org/html/2605.19285#bib.bib30)\), or assess LLMs’ adaptability to the responses\(Zhanget al\.,[2025](https://arxiv.org/html/2605.19285#bib.bib47); Liuet al\.,[2026](https://arxiv.org/html/2605.19285#bib.bib46); Yanet al\.,[2026](https://arxiv.org/html/2605.19285#bib.bib56)\)\. In contrast, our work prioritizes data selection through counterfactual attribution by selectively masking intermediate reasoning steps to observe their impact on the final output\(Chenet al\.,[2025](https://arxiv.org/html/2605.19285#bib.bib30)\)\. ## 6\.Conclusion In this work, we address the growing need for transparency in MD by shifting the focus from black\-box classification to explainable MD\. While the prevailing trend leverages LLMs through prompting or naive fine\-tuning, our experiments reveal that existing data synthesis pipelines suffer from two critical bottlenecks:insufficientrationales caused by coarse\-grained labels andunnecessaryrationales caused by the over\-verification behavior\. These issues introduce noise and verbosity that hinder the training of LLMs\. To overcome these issues, we introduceLonsRex, a novel data\-centric pipeline\. By implementing a perplexity\-based metric grounded in self\-attribution and mutual\-attribution,LonsRexeffectively identifies and selects rationales that are both necessary and sufficient\. Our experimental results demonstrate that fine\-tuning lightweight LLMs \(*e\.g\.,*1\.5B to 4B parameters\) using our filtered dataset leads to a significant performance boost, approximately 22\.97% over foundation models, achieving results comparable to much larger state\-of\-the\-art open\-source LLMs\. Ultimately, by releasing our raw dataset and the refinedLonsRexversion, we provide a robust resource for the community to develop more transparent MD systems\. ## Acknowledgement This work was supported by the National Natural Science Foundation of China \(No\.62276113\) and Alibaba Research Intern Program\. ## References - T\. Bian, X\. Xiao, T\. Xu, P\. Zhao, W\. Huang, Y\. Rong, and J\. Huang \(2020\)Rumor detection on social media with bi\-directional graph convolutional networks\.InAAAI,pp\. 549–556\.Cited by:[§5\.1](https://arxiv.org/html/2605.19285#S5.SS1.p1.1)\. - X\. Chen, S\. Zhou, K\. Liang, X\. Sun, and X\. Liu \(2025\)Skip\-thinking: chunk\-wise chain\-of\-thought distillation enable smaller language models to reason better and faster\.InEMNLP,pp\. 12153–12168\.Cited by:[§1](https://arxiv.org/html/2605.19285#S1.p3.1),[§5\.2](https://arxiv.org/html/2605.19285#S5.SS2.p2.1)\. - Z\. Chen, Q\. Gao, A\. Bosselut, A\. Sabharwal, and K\. Richardson \(2023\)DISCO: distilling counterfactuals with large language models\.InACL,pp\. 5514–5528\.Cited by:[§2](https://arxiv.org/html/2605.19285#S2.p3.6)\. - Y\. Du, O\. Watkins, Z\. Wang, C\. Colas, T\. Darrell, P\. Abbeel, A\. Gupta, and J\. Andreas \(2023\)Guiding pretraining in reinforcement learning with large language models\.InICML,Vol\.202,pp\. 8657–8677\.Cited by:[§5\.2](https://arxiv.org/html/2605.19285#S5.SS2.p1.1)\. - Y\. Dun, K\. Tu, C\. Chen, C\. Hou, and X\. Yuan \(2021\)KAN: knowledge\-aware attention network for fake news detection\.InAAAI,pp\. 81–89\.Cited by:[§1](https://arxiv.org/html/2605.19285#S1.p2.1)\. - D\. Guo, D\. Yang, H\. Zhang, J\. Song, P\. Wang, Q\. Zhu, R\. Xu, R\. Zhang, S\. Ma, X\. Bi,et al\.\(2025\)DeepSeek\-r1 incentivizes reasoning in llms through reinforcement learning\.Nature645\(8081\),pp\. 633–638\.Cited by:[§1](https://arxiv.org/html/2605.19285#S1.p3.1),[§5\.2](https://arxiv.org/html/2605.19285#S5.SS2.p1.1)\. - N\. Ho, L\. Schmid, and S\. Yun \(2023\)Large language models are reasoning teachers\.InACL,pp\. 14852–14882\.Cited by:[§5\.2](https://arxiv.org/html/2605.19285#S5.SS2.p2.1)\. - R\. Hong, J\. Lang, J\. Xu, Z\. Cheng, T\. Zhong, and F\. Zhou \(2025\)Following clues, approaching the truth: explainable micro\-video rumor detection via chain\-of\-thought reasoning\.InWWW,pp\. 4684–4698\.Cited by:[§5\.1](https://arxiv.org/html/2605.19285#S5.SS1.p1.1)\. - B\. Hu, Q\. Sheng, J\. Cao, Y\. Shi, Y\. Li, D\. Wang, and P\. Qi \(2024\)Bad actor, good advisor: exploring the role of large language models in fake news detection\.InAAAI,pp\. 22105–22113\.Cited by:[§1](https://arxiv.org/html/2605.19285#S1.p1.1),[§1](https://arxiv.org/html/2605.19285#S1.p2.1),[§4\.1](https://arxiv.org/html/2605.19285#S4.SS1.p3.1),[§4\.2](https://arxiv.org/html/2605.19285#S4.SS2.p1.1),[§5\.1](https://arxiv.org/html/2605.19285#S5.SS1.p1.1)\. - X\. Huang, Rishabh, G\. Franke, Z\. Yang, J\. Bai, W\. Bai, J\. Bi, Z\. Ding, Y\. Duan, C\. Fan,et al\.\(2025\)Loong: synthesize long chain\-of\-thoughts at scale through verifiers\.CoRRabs/2509\.03059\.Cited by:[§5\.2](https://arxiv.org/html/2605.19285#S5.SS2.p2.1)\. - D\. Jin, J\. Yang, X\. Wang, J\. Zhang, S\. Li, and D\. He \(2025a\)A dynamic knowledge update\-driven model with large language models for fake news detection\.InIJCAI,pp\. 3000–3008\.Cited by:[§5\.1](https://arxiv.org/html/2605.19285#S5.SS1.p1.1)\. - M\. Jin, W\. Luo, S\. Cheng, X\. Wang, W\. Hua, R\. Tang, W\. Y\. Wang, and Y\. Zhang \(2025b\)Disentangling memory and reasoning ability in large language models\.InACL,pp\. 1681–1701\.Cited by:[§4\.1](https://arxiv.org/html/2605.19285#S4.SS1.p3.1),[§4\.2](https://arxiv.org/html/2605.19285#S4.SS2.p1.1)\. - T\. Kojima, S\. S\. Gu, M\. Reid, Y\. Matsuo, and Y\. Iwasawa \(2022\)Large language models are zero\-shot reasoners\.InNeurIPS,Cited by:[§5\.2](https://arxiv.org/html/2605.19285#S5.SS2.p1.1)\. - Z\. Kou, J\. Chen, X\. Cai, M\. Xie, B\. Liu, C\. Wang, L\. Feng, Y\. Jia, G\. Niu, M\. Sugiyama, and X\. Geng \(2025\)Rethinking toxicity evaluation in large language models: A multi\-label perspective\.CoRRabs/2510\.15007\.Cited by:[§5\.2](https://arxiv.org/html/2605.19285#S5.SS2.p2.1)\. - D\. M\. Lazer, M\. A\. Baum, Y\. Benkler, A\. J\. Berinsky, K\. M\. Greenhill, F\. Menczer, M\. J\. Metzger, B\. Nyhan, G\. Pennycook, D\. Rothschild,et al\.\(2018\)The science of fake news\.Science359\(6380\),pp\. 1094–1096\.Cited by:[§1](https://arxiv.org/html/2605.19285#S1.p1.1)\. - Z\. Lei, Z\. Tan, S\. Wang, Y\. Zhu, Z\. Chen, Y\. Dong, and J\. Li \(2025\)Learning from diverse reasoning paths with routing and collaboration\.InEMNLP,pp\. 2832–2845\.Cited by:[§5\.2](https://arxiv.org/html/2605.19285#S5.SS2.p2.1)\. - K\. Liu, S\. Yan, R\. Miao, B\. Wang, C\. Shen, J\. Zhang, and J\. Ye \(2026\)Where did this sentence come from? tracing provenance in LLM reasoning distillation\.InICLR,Cited by:[§5\.2](https://arxiv.org/html/2605.19285#S5.SS2.p2.1)\. - Y\. Liu, Y\. Liu, X\. Zhang, X\. Chen, and R\. Yan \(2025a\)The truth becomes clearer through debate\! multi\-agent systems with large language models unmask fake news\.InSIGIR,pp\. 504–514\.Cited by:[§5\.1](https://arxiv.org/html/2605.19285#S5.SS1.p1.1)\. - Z\. Liu, K\. Yang, Q\. Xie, C\. De Kock, S\. Ananiadou, and E\. Hovy \(2025b\)Raemollm: retrieval augmented llms for cross\-domain misinformation detection using in\-context learning based on emotional information\.InACL,pp\. 16508–16523\.Cited by:[§1](https://arxiv.org/html/2605.19285#S1.p2.1)\. - Y\. Luo, Y\. Song, X\. Zhang, J\. Liu, W\. Wang, G\. Chen, W\. Su, and B\. Zheng \(2025\)Deconstructing long chain\-of\-thought: A structured reasoning optimization framework for long cot distillation\.CoRRabs/2503\.16385\.Cited by:[§1](https://arxiv.org/html/2605.19285#S1.p3.1)\. - A\. Modzelewski, W\. Sosnowski, T\. Labruna, A\. Wierzbicki, and G\. D\. S\. Martino \(2025\)PCoT: persuasion\-augmented chain of thought for detecting fake news and social media disinformation\.InACL,pp\. 24959–24983\.Cited by:[§1](https://arxiv.org/html/2605.19285#S1.p1.1),[§1](https://arxiv.org/html/2605.19285#S1.p2.1),[§4\.1](https://arxiv.org/html/2605.19285#S4.SS1.p2.1),[§4\.1](https://arxiv.org/html/2605.19285#S4.SS1.p3.1),[§4\.2](https://arxiv.org/html/2605.19285#S4.SS2.p1.1),[§5\.1](https://arxiv.org/html/2605.19285#S5.SS1.p1.1)\. - N\. Muennighoff, Z\. Yang, W\. Shi, X\. L\. Li, L\. Fei\-Fei, H\. Hajishirzi, L\. Zettlemoyer, P\. Liang, E\. J\. Candès, and T\. Hashimoto \(2025\)S1: simple test\-time scaling\.CoRRabs/2501\.19393\.Cited by:[§5\.2](https://arxiv.org/html/2605.19285#S5.SS2.p2.1)\. - Q\. Nan, Q\. Sheng, J\. Cao, B\. Hu, D\. Wang, and J\. Li \(2024\)Let silence speak: enhancing fake news detection with generated comments from large language models\.InCIKM,pp\. 1732–1742\.Cited by:[§4\.1](https://arxiv.org/html/2605.19285#S4.SS1.p3.1),[§4\.2](https://arxiv.org/html/2605.19285#S4.SS2.p1.1),[§5\.1](https://arxiv.org/html/2605.19285#S5.SS1.p1.1)\. - OpenAI \(2025\)Gpt\-oss\-120b & gpt\-oss\-20b model card\.CoRRabs/2508\.10925\.Cited by:[§2\.2](https://arxiv.org/html/2605.19285#S2.SS2.p1.1)\. - P\. Przybyla \(2020\)Capturing the style of fake news\.InAAAI,pp\. 490–497\.Cited by:[§5\.1](https://arxiv.org/html/2605.19285#S5.SS1.p1.1)\. - P\. Qi, Y\. Bu, J\. Cao, W\. Ji, R\. Shui, J\. Xiao, D\. Wang, and T\. Chua \(2023\)FakeSV: A multimodal benchmark with rich social context for fake news detection on short video platforms\.InAAAI,pp\. 14444–14452\.Cited by:[§5\.1](https://arxiv.org/html/2605.19285#S5.SS1.p1.1)\. - Z\. Shao, P\. Wang, Q\. Zhu, R\. Xu, J\. Song, M\. Zhang, Y\. K\. Li, Y\. Wu, and D\. Guo \(2024\)DeepSeekMath: pushing the limits of mathematical reasoning in open language models\.CoRRabs/2402\.03300\.Cited by:[§5\.2](https://arxiv.org/html/2605.19285#S5.SS2.p1.1)\. - J\. Su, T\. Y\. Zhuo, J\. Mansurov, D\. Wang, and P\. Nakov \(2023\)Fake news detectors are biased against texts generated by large language models\.CoRRabs/2309\.08674\.Cited by:[§4\.1](https://arxiv.org/html/2605.19285#S4.SS1.p2.1)\. - Z\. Tong, Y\. Gu, H\. Liu, Q\. Liu, S\. Wu, H\. Shi, and X\. Zhang \(2025\)Generate first, then sample: enhancing fake news detection with llm\-augmented reinforced sampling\.InACL,pp\. 24276–24290\.Cited by:[§5\.1](https://arxiv.org/html/2605.19285#S5.SS1.p1.1)\. - S\. Vosoughi, D\. Roy, and S\. Aral \(2018\)The spread of true and false news online\.Science359\(6380\),pp\. 1146–1151\.Cited by:[§1](https://arxiv.org/html/2605.19285#S1.p1.1)\. - H\. Wan, S\. Feng, Z\. Tan, H\. Wang, Y\. Tsvetkov, and M\. Luo \(2024\)DELL: generating reactions and explanations for llm\-based misinformation detection\.InFindings of ACL,pp\. 2637–2667\.Cited by:[§5\.1](https://arxiv.org/html/2605.19285#S5.SS1.p1.1)\. - B\. Wang, X\. Li, C\. Li, B\. Fu, S\. Pei, and S\. Wang \(2024a\)Why misinformation is created? detecting them by integrating intent features\.InCIKM,pp\. 2304–2314\.Cited by:[§1](https://arxiv.org/html/2605.19285#S1.p2.1)\. - B\. Wang, X\. Li, C\. Li, B\. Zhao, B\. Fu, R\. Guan, and S\. Wang \(2025a\)Robust misinformation detection by visiting potential commonsense conflict\.InIJCAI,pp\. 7760–7768\.Cited by:[§1](https://arxiv.org/html/2605.19285#S1.p2.1)\. - B\. Wang, B\. Zhao, X\. Li, C\. Li, W\. Gao, and S\. Wang \(2025b\)Collaboration and controversy among experts: rumor early detection by tuning a comment generator\.InSIGIR,pp\. 468–478\.Cited by:[§1](https://arxiv.org/html/2605.19285#S1.p1.1),[§5\.1](https://arxiv.org/html/2605.19285#S5.SS1.p1.1)\. - B\. Wang, J\. Ma, H\. Lin, Z\. Yang, R\. Yang, Y\. Tian, and Y\. Chang \(2024b\)Explainable fake news detection with large language model via defense among competing wisdom\.InWWW,pp\. 2452–2463\.Cited by:[§1](https://arxiv.org/html/2605.19285#S1.p2.1),[§4\.6](https://arxiv.org/html/2605.19285#S4.SS6.p1.1)\. - P\. Wang, Z\. Wang, Z\. Li, Y\. Gao, B\. Yin, and X\. Ren \(2023\)SCOTT: self\-consistent chain\-of\-thought distillation\.InACL,pp\. 5546–5558\.Cited by:[§5\.2](https://arxiv.org/html/2605.19285#S5.SS2.p2.1)\. - Y\. Wang, Z\. Gu, S\. Zhang, S\. Zheng, T\. Wang, T\. Li, H\. Feng, and Y\. Xiao \(2025c\)LLM\-GAN: constructing generative adversarial network through large language models for explainable fake news detection\.InIEEE International Conference on Acoustics, Speech and Signal Processing,pp\. 1–5\.Cited by:[§1](https://arxiv.org/html/2605.19285#S1.p2.1)\. - Z\. Wang, D\. Wang, Q\. Sheng, J\. Cao, S\. Ma, and H\. Cheng \(2025d\)Exploring news intent and its application: A theory\-driven approach\.Information Processing & Management62\(6\),pp\. 104229\.Cited by:[§1](https://arxiv.org/html/2605.19285#S1.p2.1)\. - J\. Wei, X\. Wang, D\. Schuurmans, M\. Bosma, B\. Ichter, F\. Xia, E\. H\. Chi, Q\. V\. Le, and D\. Zhou \(2022\)Chain\-of\-thought prompting elicits reasoning in large language models\.InNeurIPS,Cited by:[§4\.1](https://arxiv.org/html/2605.19285#S4.SS1.p3.1),[§4\.2](https://arxiv.org/html/2605.19285#S4.SS2.p1.1),[§5\.2](https://arxiv.org/html/2605.19285#S5.SS2.p1.1)\. - J\. Wu and B\. Hooi \(2023\)DECOR: degree\-corrected social graph refinement for fake news detection\.InKDD,pp\. 2582–2593\.Cited by:[§5\.1](https://arxiv.org/html/2605.19285#S5.SS1.p1.1)\. - S\. Yan, K\. Liu, C\. Shen, B\. Wang, S\. Fan, J\. Zhang, Y\. Wu, Z\. Wang, and J\. Ye \(2026\)Distribution\-aligned sequence distillation for superior long\-cot reasoning\.CoRRabs/2601\.09088\.Cited by:[§5\.2](https://arxiv.org/html/2605.19285#S5.SS2.p2.1)\. - A\. Yang, A\. Li, B\. Yang, B\. Zhang, B\. Hui, B\. Zheng, B\. Yu, C\. Gao, C\. Huang, C\. Lv,et al\.\(2025\)Qwen3 technical report\.CoRRabs/2505\.09388\.Cited by:[§1](https://arxiv.org/html/2605.19285#S1.p3.1)\. - X\. Yang, L\. Li, H\. Zhou, T\. Zhu, X\. Qu, Y\. Fan, Q\. Wei, R\. Ye, L\. Kang, Y\. Qin, Z\. Kou, D\. Liu, Q\. Li, N\. Ding, S\. Chen, and J\. Shao \(2026\)Toward efficient agents: memory, tool learning, and planning\.CoRRabs/2601\.14192\.Cited by:[§5\.2](https://arxiv.org/html/2605.19285#S5.SS2.p1.1)\. - Z\. Yang, J\. Ma, H\. Chen, H\. Lin, Z\. Luo, and Y\. Chang \(2022\)A coarse\-to\-fine cascaded evidence\-distillation neural network for explainable fake news detection\.InCOLING,pp\. 2608–2621\.Cited by:[§1](https://arxiv.org/html/2605.19285#S1.p2.1)\. - Y\. Ye, Z\. Huang, Y\. Xiao, E\. Chern, S\. Xia, and P\. Liu \(2025\)LIMO: less is more for reasoning\.InCOLM,Cited by:[§5\.2](https://arxiv.org/html/2605.19285#S5.SS2.p2.1)\. - X\. Yuan, C\. Shen, S\. Yan, K\. Liu, X\. Zhang, S\. Fan, L\. Xie, W\. Wang, R\. Guan, Y\. Wang, and J\. Ye \(2026\)Differential fine\-tuning large language models towards better diverse reasoning abilities\.InICLR,Cited by:[§5\.2](https://arxiv.org/html/2605.19285#S5.SS2.p2.1)\. - X\. Yuan, C\. Shen, S\. Yan, X\. Zhang, L\. Xie, W\. Wang, R\. Guan, Y\. Wang, and J\. Ye \(2024\)Instance\-adaptive zero\-shot chain\-of\-thought prompting\.InNeurIPS,pp\. 125469–125486\.Cited by:[§5\.2](https://arxiv.org/html/2605.19285#S5.SS2.p1.1)\. - Z\. Yue, H\. Zeng, Y\. Lu, L\. Shang, Y\. Zhang, and D\. Wang \(2024\)Evidence\-driven retrieval augmented response generation for online misinformation\.InNAACL,pp\. 5628–5643\.Cited by:[§1](https://arxiv.org/html/2605.19285#S1.p2.1)\. - D\. Zhang, Q\. Dai, and H\. Peng \(2025\)The best instruction\-tuning data are those that fit\.InNeurIPS,Cited by:[§5\.2](https://arxiv.org/html/2605.19285#S5.SS2.p2.1)\. - X\. Zhang, J\. Cao, X\. Li, Q\. Sheng, L\. Zhong, and K\. Shu \(2021\)Mining dual emotion for fake news detection\.InWWW,pp\. 3465–3476\.Cited by:[§1](https://arxiv.org/html/2605.19285#S1.p2.1)\. Table 8\.Experimental results on more advanced LLMs across four MD benchmarks\.LLMs & Prompt\#paramGossipCop\+\+PolitiFact\+\+MultiDisEUDisinfoAvg\.Acc\.F1fake\{\}\_\{\\text\{fake\}\}F1real\{\}\_\{\\text\{real\}\}Acc\.F1fake\{\}\_\{\\text\{fake\}\}F1real\{\}\_\{\\text\{real\}\}Acc\.F1fake\{\}\_\{\\text\{fake\}\}F1real\{\}\_\{\\text\{real\}\}Acc\.F1fake\{\}\_\{\\text\{fake\}\}F1real\{\}\_\{\\text\{real\}\}\>\>100B parameters[DeepSeek\-V3\.2](https://huggingface.co/deepseek-ai/DeepSeek-V3.2)685B68\.9761\.1874\.1687\.8979\.7191\.3790\.0382\.8992\.9784\.1277\.3487\.7881\.53\+ CoT\(NeurIPS’22\)685B65\.1550\.0673\.2389\.1580\.9592\.4189\.0780\.2892\.4478\.8761\.6385\.4278\.22\+ ARG\(AAAI’24\)685B67\.1256\.1573\.7188\.5781\.0891\.8291\.1884\.6493\.8283\.2274\.4487\.5181\.11\+GenFend\(CIKM’24\)685B66\.6253\.2474\.0587\.6379\.6091\.1292\.6486\.5594\.9384\.6877\.9188\.2781\.44\+ DMR\(ACL’25\)685B66\.7555\.7073\.3988\.4980\.9191\.7689\.7382\.5292\.7382\.7374\.3386\.9980\.50\+ PCoT\(ACL’25\)685B65\.8051\.7373\.5284\.2874\.4188\.6591\.1383\.2293\.9785\.5876\.8789\.5379\.89[Qwen3\-235B\-A22B\-Ins\.](https://huggingface.co/Qwen/Qwen3-235B-A22B-Instruct-2507)235B64\.9866\.4463\.3784\.8877\.8388\.5388\.4381\.0291\.6875\.5668\.9779\.8477\.63\+ CoT\(NeurIPS’22\)235B63\.1557\.6267\.4084\.0271\.9988\.8286\.5278\.1890\.2577\.1667\.8482\.2976\.27\+ ARG\(AAAI’24\)235B66\.2060\.7470\.3284\.1976\.0488\.2189\.6381\.2792\.8380\.9971\.7185\.6878\.98\+GenFend\(CIKM’24\)235B64\.1062\.9765\.1682\.3873\.7586\.7489\.9882\.8292\.9377\.7269\.3582\.4977\.53\+ DMR\(ACL’25\)235B66\.7067\.3466\.0483\.6876\.1987\.5884\.8276\.9988\.6783\.0177\.6686\.2978\.75\+ PCoT\(ACL’25\)235B66\.7759\.0472\.0585\.4877\.7989\.2290\.0882\.9093\.0184\.9678\.5788\.4180\.69[gpt\-oss\-120b\(high effort\)](https://huggingface.co/openai/gpt-oss-120b)120B58\.0566\.8842\.8157\.4757\.2257\.7368\.2960\.7173\.4164\.3556\.4669\.8161\.10\+ CoT\(NeurIPS’22\)120B58\.1067\.1842\.0957\.1357\.0957\.1765\.6858\.5170\.7464\.8356\.5070\.4960\.46\+ ARG\(AAAI’24\)120B60\.7567\.2251\.0961\.1759\.2162\.9571\.0962\.0676\.6568\.3158\.3074\.4564\.44\+GenFend\(CIKM’24\)120B57\.9557\.4658\.4358\.4252\.9262\.7754\.0634\.2764\.6955\.7843\.3563\.7455\.32\+ DMR\(ACL’25\)120B57\.7264\.0448\.7161\.3457\.2264\.7367\.6456\.0574\.3964\.9052\.7272\.0961\.80\+ PCoT\(ACL’25\)120B64\.7065\.8363\.5068\.9064\.3072\.4584\.7776\.6988\.6980\.7873\.6184\.8874\.09\>\>10B,<<100B parameters[Qwen3\-Next\-80B\-A3B](https://huggingface.co/Qwen/Qwen3-Next-80B-A3B-Instruct)80B62\.3062\.1562\.4584\.9778\.2688\.5188\.9381\.9692\.0182\.8076\.7686\.3578\.95\+ CoT\(NeurIPS’22\)80B65\.9263\.0168\.4186\.5179\.6489\.9289\.1382\.2992\.1679\.2573\.2083\.0779\.38\+ ARG\(AAAI’24\)80B63\.1860\.2065\.7484\.7977\.8588\.4286\.7778\.6490\.4279\.3271\.1483\.8877\.53\+GenFend\(CIKM’24\)80B64\.9563\.0966\.6385\.8278\.9389\.3287\.8880\.4291\.2278\.3259\.7085\.1777\.62\+ DMR\(ACL’25\)80B61\.0864\.7856\.5078\.9573\.3482\.6169\.5462\.7974\.2272\.2168\.4675\.1769\.97\+ PCoT\(ACL’25\)80B66\.1057\.2771\.9086\.5978\.0890\.3490\.9383\.7693\.7078\.3472\.4582\.1679\.30[Llama\-3\.1\-70B\-Instruct](https://huggingface.co/meta-llama/Llama-3.1-70B-Instruct)70B58\.8854\.9462\.1869\.8558\.3676\.3663\.9353\.0670\.7171\.9455\.7679\.4564\.62\+ CoT\(NeurIPS’22\)70B63\.7044\.7572\.9782\.4770\.0987\.6185\.2266\.9790\.4879\.3262\.3685\.7474\.31\+ ARG\(AAAI’24\)70B62\.3244\.3371\.5384\.6274\.3989\.0184\.6769\.0389\.8178\.1357\.3485\.3074\.21\+GenFend\(CIKM’24\)70B61\.5040\.3671\.5882\.7371\.3387\.6583\.4261\.5689\.4371\.9453\.6279\.8871\.25\+ DMR\(ACL’25\)70B64\.1553\.2670\.9278\.9568\.4784\.2081\.6667\.3287\.2675\.7763\.0681\.9773\.08\+ PCoT\(ACL’25\)70B62\.0550\.3969\.2777\.4167\.2582\.7585\.8777\.1889\.7780\.8570\.5985\.8074\.93[R1\-Distill\-Qwen\-32B](https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-32B)671/32B65\.7554\.0272\.7088\.6580\.2992\.0384\.9173\.1089\.5272\.2853\.3980\.2775\.58\+ CoT\(NeurIPS’22\)671/32B66\.5257\.1272\.5489\.0881\.2492\.3083\.0170\.2388\.1173\.8858\.3780\.9776\.11\+ ARG\(AAAI’24\)671/32B65\.1759\.9169\.2170\.1862\.5675\.2373\.7460\.3680\.3767\.4755\.2274\.4667\.82\+GenFend\(CIKM’24\)671/32B62\.7243\.3772\.2186\.3475\.3490\.5587\.3774\.0791\.6578\.7661\.1485\.3875\.74\+ DMR\(ACL’25\)671/32B65\.3758\.0970\.5086\.7677\.8090\.5781\.1169\.4286\.3373\.1159\.3679\.9174\.86\+ PCoT\(ACL’25\)671/32B64\.7248\.0273\.3087\.5477\.5191\.3890\.3880\.8793\.5782\.7967\.8888\.2578\.85[Qwen3\-32B](https://huggingface.co/Qwen/Qwen3-32B)32B67\.7060\.2472\.8087\.2879\.0390\.8789\.7281\.7492\.8581\.5472\.6586\.0780\.21\+ CoT\(NeurIPS’22\)32B62\.5355\.1967\.8186\.9678\.9090\.5786\.5776\.6190\.5876\.1163\.7882\.1876\.48\+ ARG\(AAAI’24\)32B59\.4341\.4668\.9677\.8455\.2085\.2886\.5776\.6190\.5876\.1163\.7882\.1872\.00\+GenFend\(CIKM’24\)32B61\.7854\.3467\.1487\.0579\.2690\.5985\.1274\.8589\.4374\.5861\.6180\.9975\.56\+ DMR\(ACL’25\)32B62\.7557\.6466\.7587\.5479\.9490\.9675\.2064\.7180\.8874\.0264\.4479\.5373\.70\+ PCoT\(ACL’25\)32B67\.1254\.9874\.1086\.7677\.7490\.5889\.7780\.6093\.0683\.2876\.0087\.1780\.10[Qwen3\-30B\-A3B\-Instr\.](https://huggingface.co/Qwen/Qwen3-30B-A3B-Instruct-2507)30B67\.0368\.2465\.7178\.8772\.0583\.0188\.0380\.5891\.3476\.6771\.3480\.3376\.93\+ CoT\(NeurIPS’22\)30B64\.7566\.2863\.0783\.7678\.0087\.1381\.3172\.7585\.7876\.7472\.2180\.0075\.98\+ ARG\(AAAI’24\)30B63\.9255\.3769\.7381\.8772\.7086\.4388\.0378\.8391\.6576\.0463\.3382\.2175\.84\+GenFend\(CIKM’24\)30B67\.3261\.3971\.6883\.3375\.3287\.4289\.3380\.7292\.6276\.6772\.4379\.7881\.17\+ DMR\(ACL’25\)30B63\.6065\.7461\.1776\.2971\.1979\.8569\.5462\.8474\.1971\.5268\.3774\.1069\.87\+ PCoT\(ACL’25\)30B68\.7369\.0768\.3781\.3673\.6385\.5890\.3384\.2493\.0285\.0380\.7987\.7480\.66[gpt\-oss\-20b\(high effort\)](https://huggingface.co/openai/gpt-oss-20b)20B56\.2067\.1434\.3358\.7658\.6958\.8366\.3359\.9070\.9868\.3862\.9172\.4561\.24\+ CoT\(NeurIPS’22\)20B54\.8766\.7529\.7955\.0757\.0352\.9359\.6755\.8462\.8964\.1460\.8466\.9257\.23\+ ARG\(AAAI’24\)20B52\.2266\.8614\.4241\.7552\.1225\.6640\.6846\.5233\.4144\.2952\.2733\.1141\.94\+GenFend\(CIKM’24\)20B48\.3957\.8933\.3352\.7551\.7553\.7056\.4146\.8263\.0758\.9851\.6864\.3753\.26\+ DMR\(ACL’25\)20B55\.6563\.8042\.7761\.1757\.8464\.0163\.7854\.9569\.7165\.4657\.0271\.1360\.61\+ PCoT\(ACL’25\)20B63\.4866\.1660\.3375\.0069\.5978\.7786\.1778\.0389\.9177\.7970\.1082\.3374\.81<<10B parametersLonsRex\-4B\(Ours\)4B67\.9865\.2670\.2987\.1178\.5790\.7992\.2886\.4794\.6085\.0378\.1388\.6282\.09LonsRex\-2B\(Ours\)2B67\.1265\.0968\.9381\.5366\.4687\.2689\.5882\.1692\.6478\.3470\.1283\.0177\.69LonsRex\-1\.5B\(Ours\)1\.5B68\.5069\.0167\.9784\.9775\.5989\.1490\.6383\.1793\.5183\.1572\.3187\.8980\.49Figure 8\.Distribution ofΔ\\Deltafor correct / incorrect rationales\.
Similar Articles
Safety as a Constraint: Fine-Tuning a LLM Recommender to Explain Itself
This paper fine-tunes a large language model for a video streaming service to generate personalized, faithful, and harmless recommendation explanations, improving pass rates without compromising original performance.
What LLMs explain is not what they believe: Evaluating explanation sufficiency under models' own input beliefs
This paper formalizes the concept of explanation sufficiency for LLMs, proposes a new metric called SCSuff to evaluate free-text explanations using the model's own input beliefs, and demonstrates that current LLM explanations are generally insufficient.
Error as a Lens: Probing LLM Reasoning through Synthetic Misconception Generation
This paper presents a framework using LLMs to generate targeted synthetic misconceptions aligned to a five-class taxonomy adapted from Bloom's taxonomy, addressing the scarcity of labeled student error data in education research.
When LLMs Learn to Be Consistently Wrong: A Multi-Model Study of Linear Representations of Synthetic Deception
This paper studies synthetic dishonesty in LLMs by fine-tuning honest and deceptive variants of five transformer models and finding that robust, domain-invariant dishonesty representations can be rapidly entrenched via modest supervised fine-tuning, with implications for activation-based monitoring.
Training-Time Explainability for Multilingual Hate Speech Detection: Aligning Model Reasoning with Human Rationales
This paper proposes a training-time explainability framework for multilingual hate speech detection, aligning model reasoning with human rationales to improve classification performance and interpretability, evaluated on English and Hinglish datasets.