Same Verdict, Different Readers: Personalizing Fact-Checking Explanations from Relative Preferences

1 Introduction

• The same explanation may not suit readers with different backgrounds, interests, and preferences, motivating adaptation of presentation while reusing reliable factual claims.Rewriting can personalize explanations without repeating the underlying verification process.Source: §1 · p. 1
• PERPL formulates faithful personalized fact-checking explanation rewriting, adapting an existing explanation to reader preferences while preserving its factual content.Personalization may change organization, emphasis, detail, and wording, but must preserve the verdict, factual claims, source attributions, and necessary qualifications.Source: §1 · p. 1
• Personalized rewriting faces two challenges: separating persona-specific preferences from general writing quality and learning beyond a limited set of predefined explanations.Candidate-order sensitivity can also make simulated preference judgments noisy.Source: §1 · p. 2
• PERPL constructs persona-specific supervision through simulated preference selection and cross-persona reciprocal pairs, then progressively trains a rewriter through warm-up learning and exploration-guided preference optimization.The construction uses position-balanced comparisons; exploration compares newly sampled rewrites and aggregates multiple highly ranked explanations through Borda ranking.Source: §1 · p. 2
• PERPL outperforms much larger general-purpose LLMs in simulated reader preference while maintaining high factual fidelity.The paper presents this as enabling a lightweight rewriter to achieve preference alignment without sacrificing fidelity.Source: §1 · p. 2

2 Problem Formulation

• The task is to adapt an existing fact-checking explanation to a reader while preserving fidelity to the claim’s evidence and verdict.The inputs are a claim, available evidence, a verdict, an original explanation, and a reader profile.Source: §2 · p. 2
• A reader profile captures demographic characteristics, news consumption habits, and cognitive dispositions.News habits include frequency and topics of interest.Source: §2 · p. 2
• The reader described by the profile is called a persona.The profile provides the reader information used to tailor the explanation.Source: §2 · p. 2

3 Methodology

• PERPL uses LLM-simulated reader preferences, elicited through candidate comparisons, to train a rewriter for personalized fact-checking explanations.The approach targets persona-specific preferences rather than general explanation quality.Source: §3 · p. 2, §3.1 · p. 2

3.1.1 Audited Style Candidate Construction.

• Four candidates share the same claim, evidence, verdict, and original explanation; fidelity and style audits require all four to pass for a claim to be retained.Auditors check factual support, verdict meaning, overstatement, intended style, and distinctness; candidates can undergo up to two revision rounds.Source: §3.1.1 · p. 3

3.1.2 Position-Balanced Preference Selection.

• A persona selects a candidate across four cyclic orderings only if it wins at least three, then must prefer it to the original in both orders for retention.Disagreement between presentation orders yields no preference.Source: §3.1.2 · pp. 3–4

3.1.3 Cross-Persona Reciprocal Pairing.

• For the same claim, reciprocal pairs retain two personas’ opposite preferences only after each persona reliably prefers its own selected explanation in both presentation orders.The paired inputs differ only in persona profile, directly supervising persona-specific preferences.Source: §3.1.3 · p. 4

3.2 Progressive Preference Learning

• Progressive Preference Learning starts with supervision over predefined explanations and expands it to simulated preferences over new rewrites generated by the rewriter.This progression is intended to let the model discover explanations beyond the predefined styles.Source: §3.2 · p. 4

3.2.1 Stage I: Warm-up Training.

• Stage I fine-tunes on persona-preferred explanations, then optimizes reciprocal pairs with a DPO-style loss and an auxiliary likelihood term.The auxiliary term reinforces generation of preferred explanations, and the resulting warm-up rewriter seeds later exploration.Source: §3.2.1 · p. 4

3.2.2 Stage II: Exploration-guided Preference Optimization.

• Stage II samples diverse rewrites, audits and filters them, ranks candidates with Borda scores, selects reliable positives, and pairs them with consistently less-preferred candidates.Each training pair is supported by both an order-consistent pairwise preference and the overall Borda ranking.Source: §3.2.2 · pp. 4–5

4.1 Experimental Setup

• The corpus contains 2,080 claims from MOCHEG, TRUE, and VeriTaS, split into 1,540 training, 269 validation, and 271 test claims.The evaluation uses ten personas per test claim, for 2,710 claim–persona pairs.Source: §4.1 · pp. 5–6, §4.2 · p. 6
• Ten personas per claim are formed by combining demographic characteristics, news-consumption habits, and cognitive dispositions.The implementation uses Qwen3.5-4B as the LoRA-fine-tuned rewriter and Qwen3.8-27B as the simulated reader for preference construction and Stage II exploration.Source: §4.1 · p. 6
• PERPL uses a lightweight Qwen3.5-4B rewriter fine-tuned with LoRA, while preference construction and Stage II exploration use Qwen3.8-27B as the simulated reader.Stage II samples four rewrites per input alongside a greedy-decoded reference and selects up to two positives.Source: §4.1 · p. 6

4.2 Personalized Rewriting Performance

• PERPL’s average DWR against Selection-Only is 81.75%, and its target-persona rewrites outperform mismatched-persona rewrites across evaluators.PERPL also achieves more wins than losses against all baselines under GPT and GLM, which were not used for preference construction or optimization.Source: §4.2 · pp. 6–7, Table 3 · p. 7
• PERPL achieves the best average FPW among eight zero-shot LLM baselines, including GPT-6-Astra, with the highest FPW under Qwen and GLM while remaining competitive under GPT.PERPL also achieves a high fidelity pass rate; GPT-6-Astra has the highest fidelity but lower preference win rates than PERPL under Qwen and GLM.Source: §4.2 · p. 6, Table 2 · p. 7

4.3.1 Effects of Key Training Components.

• Reciprocal preference supervision improves warm-up training, and full PERPL achieves higher FPW than Stage I Only through Stage II optimization.Using one positive, removing Borda ranking, or removing positive filtering reduces FPW; removing the auxiliary likelihood term reduces FPW, fidelity pass rate, and win rate.Source: Table 4 · p. 7, §4.3.1 · pp. 7–8

4.3.2 Position-Balanced Preference Selection.

• Four-Order (3/4) Agreement retains 76.7% of selections, achieves 71.0% recheck agreement, and has the highest persona differentiation at 46.4%.Four-Order (4/4) Agreement raises recheck agreement to 82.1% but retains only 31.1% of selections.Source: Table 5 · p. 8, §4.3.2 · p. 8

4.3.3 Effect of Stage II Exploration on Explanation Diversity.

• Stage II outputs are less similar to predefined style candidates than Warm-up outputs, with most claim-level points below the diagonal in Figure 3(a).Figure 3(b) reports higher median persona-diversity scores for PERPL than Warm-up: 0.08344 versus 0.06836, a 22.06% relative increase.Source: Figure 3 · p. 8, §4.3.3 · p. 8

• PERPL adapts existing fact-checking explanations to reader preferences while preserving factual fidelity.Its supervision comes from simulated preference selection and cross-persona reciprocal pairing, followed by warm-up training and exploration-guided preference optimization.Source: §5 · p. 8
• Across three fact-checking datasets, PERPL outperforms strong general-purpose LLMs in simulated reader preference while maintaining high factual fidelity.Source: §5 · p. 8
• Learned preferences are limited by the distributions captured by LLM-simulated readers; the authors propose collecting large-scale human preference data to validate and improve personalization for real readers.Source: §5 · p. 8

A Ethics and Privacy Statement

• The study uses public fact-checking datasets.Source: A Ethics and Privacy Statement · p. 10
• It uses synthetic personas without collecting personal data.Source: A Ethics and Privacy Statement · p. 10
• Generated explanations may contain errors or simulated preference biases, so factual verification is required before deployment.Source: A Ethics and Privacy Statement · p. 10

B Related Work

• The work is closely aligned with personalized explainable fact-checking.Source: B Related Work · p. 10
• The work is also closely aligned with LLM-based user simulation.Source: B Related Work · p. 10
• The related-work framing brings personalized explainable fact-checking and LLM-based user simulation together.Source: B Related Work · p. 10

B.1 Personalized Explainable Fact-Checking

• Existing personalized fact-checking methods use off-the-shelf LLM generation or inference-time refinement without explicitly learning persona-specific explanation preferences.Source: §B.1 · p. 10
• Gabriel et al. [8] generated explanations appealing to users’ existing beliefs and values, while Altoe et al. used affective modalities such as memes for personalized interventions.Source: §B.1 · p. 10
• Pang et al. proposed MURSE, which iteratively refines debunking passages using LLM-simulated user feedback at inference time.Source: §B.1 · p. 10
• This work formulates personalization as rewriting an existing explanation while preserving factual fidelity, using a lightweight open-weight rewriter trained through preference learning.Source: §B.1 · p. 10
• The trained rewriter achieves performance comparable to or better than much larger proprietary LLMs.Source: §B.1 · p. 10

B.2 LLM-based User Simulation

• Profile-conditioned LLMs have been used to simulate social interactions, behavioral experiments, and user behavior in recommendation and search.Source: §B.2 · p. 10
• Prior studies suggest LLM-based user simulation can obtain user-dependent feedback without extensive human participation.Source: §B.2 · p. 10
• The work simulates readers’ preferences over fact-checking explanations to provide scalable supervision without costly large-scale human preference collection.Source: §B.2 · p. 10

C Additional Implementation Details

• Both training stages run for one epoch, with LoRA rank 16 and scaling factor 32.Stage I uses effective batch size 16, weight decay 0.01, linear learning-rate decay, and 5% warm-up; Stage II uses a constant learning rate and zero weight decay.Source: C Additional Implementation Details · p. 10
• Stage I sets βI = 0.1 and α = 0.1, while Stage II sets βII = 1.0 and λ = 0.1.The Stage II exploration uses temperature 0.8 and nucleus sampling with p = 0.95.Source: C Additional Implementation Details · p. 10
• At inference, the rewriter uses greedy decoding with a maximum of 768 new tokens.Qwen3.8-27B at temperature zero supplies preference judgments.Source: C Additional Implementation Details · p. 10

D Case Study

• Each case randomly samples three of ten personas, and space limits exclude the original explanations and supporting evidence from the presentation.Source: D Case Study · p. 10
• For the carbon-dioxide claim, Personas A and B receive similar detailed reasoning, while low-deliberation, high-intuition Persona C receives a shorter verdict-and-justification explanation.A and B share high deliberation, low intuition, and interests in science and environmental topics.Source: D Case Study · p. 10
• For the zebra claim, Persona A receives detailed discussion of the photograph, diagnosis, and evidence limits; B focuses on evidence and biological interpretation, while C gets a concise verdict-focused explanation.Source: D Case Study · p. 10

E.1 Audited Style Candidate Construction

• Candidate explanations must use supplied evidence, preserve verdict scope and uncertainty, and avoid unsupported claims about images, provenance, or silent video.Source: §E.1 · p. 10
• The carbon-dioxide examples distinguish natural sinks’ current absorption rate from their total capacity, which the supplied material does not quantify.The three displayed outputs range from detailed reasoning to a concise verdict-and-evidence statement.Source: §E.1 · p. 11
• The appendix’s zebra example supports the claim through the photograph’s identification and reported pseudomelanism, while noting limits on age and genetic-mechanism evidence.The example identifies Tira as a zebra born at Masai Mara National Reserve in 2019 and attributes the diagnosis and explanation to named experts.Source: §E.1 · p. 11
• E.1 defines four candidate styles—Direct, Evidence-led, Reasoning-led, and Plain-language—with distinct organization or register while requiring supported facts and qualifications.Each style is generated in a separate request; prompts call for clear language, attribution, and no invented arguments or altered conclusions.Source: §E.1 · pp. 10, 12
• Separate audits check factual fidelity and communication style, including meaningful pairwise contrast without requiring different facts or arguments.Fidelity checks can pass, fail for repairable defects, or request source review; style checks assess organization, emphasis, and sustained register.Source: §E.1 · p. 12
• Revision uses auditor feedback for minimal factual or organizational corrections while treating feedback and other candidates as untrusted evidence.GPT-5.6-Sol generates and revises candidates, while GPT-5.6-Luna audits fidelity and style.Source: §E.1 · pp. 10, 12

E.2 Simulated Reader

• E.2 uses the same simulated-reader prompts across three evaluated LLMs, concealing candidate style labels, generation methods, and source identifiers.The profile, claim, verdict, and candidate explanations are shown; candidate order and agreement rules are controlled outside the prompt.Source: §E.2 · p. 12
• Simulated readers choose a uniquely preferred explanation only when the profile and texts provide a meaningful basis, and may return null.The profile template includes demographics, interests, news frequency, and reported deliberation and open-minded-thinking preferences.Source: §E.2 · p. 12