Ask or Answer: A Decision Framework for Multi-Turn Health Misinformation Intervention

# Ask or Answer: A Decision Framework for Multi-Turn Health Misinformation Intervention
Source: [https://arxiv.org/html/2608.21721](https://arxiv.org/html/2608.21721)
Anirban Saha AnikAffiliation:University of North TexasJinyu LiuAffiliation:University of North TexasQitao TanAffiliation:University of Georgia \{xiaoyingsong, anirbansahaanik, jinyuliu\}@my\.unt\.edu, lingzi\.hong@unt\.edu \{qitaotan, geng\.yuan\}@uga\.eduGeng YuanAffiliation:University of Georgia \{xiaoyingsong, anirbansahaanik, jinyuliu\}@my\.unt\.edu, lingzi\.hong@unt\.edu \{qitaotan, geng\.yuan\}@uga\.eduLingzi HongAffiliation:University of North Texas

###### Abstract

Correcting health misinformation in dialogue requires more than producing a factual rebuttal: users differ in what they know, what they believe, and what they need to hear, so an effective intervention often depends on first asking the right clarifying question\. Yet existing methods either respond immediately or probe indiscriminately, treating clarification as either unnecessary or always beneficial\. We proposeReward\-Optimized Probe\-and\-Respond \(RO\-PnR\), a framework that learnswhen asking is worth its cost\. At each turn, RO\-PnR chooses between probing for more information and committing to a final correction, guided by a turn\-level reward that weighs the expected gain from probing against its interaction cost\. To capture how user heterogeneity affects probing value, we model each simulated user with a latent state along health literacy and belief commitment\. Experiment shows that RO\-PnR achieves the highest cost\-adjusted utility across three health\-misinformation datasets and three base models, using∼\\sim30% fewer turns than always\-probe baselines\.

## 1Introduction

![Refer to caption](https://arxiv.org/html/2608.21721v1/Figure2.8.png)Figure 1:Different strategies on the same misinformation post\. RO\-PnR weighs benefit vs\. cost before acting\.Health misinformation distorts public health, undermining vaccine uptake and adherence to public\-health guidance[50](https://arxiv.org/html/2608.21721#bib.bib4), and dialogue\-based agents are an increasingly promising channel for correcting it[21](https://arxiv.org/html/2608.21721#bib.bib3)\. Effective correction depends on more than producing accurate facts; it depends on what the specific user knows, what they believe, and why they believe it[9](https://arxiv.org/html/2608.21721#bib.bib5)\.

Prior work on health misinformation correction has focused on the response itself: generating evidence\-grounded rebuttals[56](https://arxiv.org/html/2608.21721#bib.bib49);[1](https://arxiv.org/html/2608.21721#bib.bib51)or tailoring them to user characteristics such as health literacy[46](https://arxiv.org/html/2608.21721#bib.bib52);[38](https://arxiv.org/html/2608.21721#bib.bib53)\. These methods improve the quality of individual responses, but they commit to a reply based on the original post alone, without engaging the user to surface the underlying concern\. A separate line of work on multi\-turn dialogue uses clarification questions to handle ambiguity[5](https://arxiv.org/html/2608.21721#bib.bib44);[55](https://arxiv.org/html/2608.21721#bib.bib45), with recent decision\-theoretic methods weighing the value of asking against its cost[7](https://arxiv.org/html/2608.21721#bib.bib47);[28](https://arxiv.org/html/2608.21721#bib.bib13)\. However, these approaches typically ask follow\-up questions to clarify users’ ambiguous queries for task completion, without considering users’ underlying state\. In health misinformation intervention, users differ in their knowledge and beliefs, so the effects of asking versus answering may vary across users\. This makes correction a decision problem, not just a generation problem\.

At every interaction turn, the agent must decide whether to respond now with the information it has or ask the user a clarifying question first[5](https://arxiv.org/html/2608.21721#bib.bib44);[57](https://arxiv.org/html/2608.21721#bib.bib11)\. Responding too eagerly commits to a generic correction before the user’s real concern surfaces \(Figure 1a\)\. Asking too eagerly, probing at every opportunity, wastes the user’s attention and adds little when the context is already sufficient \(Figure 1b\)\. The right behavior depends on whether the expected benefit of one more question outweighs the cost of asking it[7](https://arxiv.org/html/2608.21721#bib.bib47)\.

The trade\-off is sharpened by user heterogeneity\. Users vary in health literacy: what they can understand and reason about[35](https://arxiv.org/html/2608.21721#bib.bib43);[3](https://arxiv.org/html/2608.21721#bib.bib54), and in belief commitment[9](https://arxiv.org/html/2608.21721#bib.bib5);[54](https://arxiv.org/html/2608.21721#bib.bib41): how firmly they hold the misconception\. A low\-literacy user who is open to revision may need a single targeted probe; a high\-literacy user who already articulated their concern may need none[46](https://arxiv.org/html/2608.21721#bib.bib52); a strong\-believer user may need careful probing precisely to gather the context that will make a correction non\-confrontational[47](https://arxiv.org/html/2608.21721#bib.bib40);[27](https://arxiv.org/html/2608.21721#bib.bib59)\. The value of asking, therefore, is itself user\-dependent\. Agents that don’t model this heterogeneity will systematically over\- or under\-probe\.

We propose Reward\-Optimized Probe\-and\-Respond \(RO\-PnR\), a framework that learns when asking is worth its cost\. At each turn, RO\-PnR chooses between probing and responding based on a turn\-level decision reward that weighs the expected gain from probing against the cost of one more question\. Because the value of probing depends on hidden user characteristics, we model each simulated user with a latent state along health literacy and belief commitment, and train the policy on turn\-level decisions\. As illustrated in Figure 1c, RO\-PnR learns to ask one well\-targeted question when context is missing, and to commit immediately when it is not\. We validate this approach across three health\-misinformation datasets and three base models\. RO\-PnR achieves the highest cost\-adjusted utility while using 30% fewer turns than always\-probe baselines, with the largest gains on low\-literacy and strongly\-committed users, exactly the slices where the probing decision matters most\.

Our contributions are:\(a\) We frame health misinformation intervention as a probe\-or\-respond decision problem grounded in domain\-specific user dynamics, moving beyond the single\-turn, one\-size\-fits\-all paradigm of prior counterspeech work\. \(b\) We introduceRO\-PnR, a decision\-theoretic policy tailored to health misinformation intervention that learns when asking is worth its cost, accounting for user heterogeneity along health literacy and belief commitment\.

## 2Related Work

### 2\.1Health Misinformation Intervention

Research on health misinformation intervention has progressed from generic factual rebuttals to retrieval\-augmented generation that grounds corrections in scientific evidence[56](https://arxiv.org/html/2608.21721#bib.bib49);[14](https://arxiv.org/html/2608.21721#bib.bib50)and audience\-aware approaches that tailor responses to the user’s health literacy or belief commitment[46](https://arxiv.org/html/2608.21721#bib.bib52);[38](https://arxiv.org/html/2608.21721#bib.bib53);[1](https://arxiv.org/html/2608.21721#bib.bib51)\. These methods improve the quality of individual responses but remain single\-turn: the agent commits to a reply based on the misinformation content, without engaging the user to surface the underlying concern or adapting the response as new information emerges\. This is particularly limiting in heterogeneous user settings, where the right intervention depends on context[20](https://arxiv.org/html/2608.21721#bib.bib1)\.

### 2\.2Decision\-Theoretic Probing in Dialogue

Recent work treats clarification as an active decision rather than a default behavior\. Some methods train models to choose between asking and answering when a request is ambiguous[5](https://arxiv.org/html/2608.21721#bib.bib44);[57](https://arxiv.org/html/2608.21721#bib.bib11), others use fixed\-schedule probing that asks a predetermined set of questions before responding[10](https://arxiv.org/html/2608.21721#bib.bib12), and decision\-theoretic approaches weigh the benefit of asking against its communication cost[7](https://arxiv.org/html/2608.21721#bib.bib47);[28](https://arxiv.org/html/2608.21721#bib.bib13)\. Closer to our setting,[55](https://arxiv.org/html/2608.21721#bib.bib45)introduces multi\-turn\-aware rewards that estimate the long\-term value of a response via simulated future conversations, and[52](https://arxiv.org/html/2608.21721#bib.bib46)adds a curiosity reward to reduce uncertainty about users’ latent state\. We build on this direction by combining forward\-looking reward estimation with an explicit interaction cost, and by grounding the latent user state in domain\-specific dimensions \(health literacy and belief commitment\) rather than a generic user variable\.

### 2\.3User Heterogeneity Modeling

Recent studies adapt dialogue agents to user\-specific traits, either by conditioning responses on explicit profiles[39](https://arxiv.org/html/2608.21721#bib.bib6)or by inferring latent representations from interaction[53](https://arxiv.org/html/2608.21721#bib.bib21)\. Research has shown that responses tailored to user literacy level are more effective[46](https://arxiv.org/html/2608.21721#bib.bib52);[38](https://arxiv.org/html/2608.21721#bib.bib53)\. Our work differs: \(1\) We treat the user state as a hidden variable that drives the agent’s*decision to probe*and shapes its response; \(2\) We model two coupled user\-state dimensions, health literacy and belief commitment, which jointly shape misinformation susceptibility[9](https://arxiv.org/html/2608.21721#bib.bib5);[32](https://arxiv.org/html/2608.21721#bib.bib56)and are key to health misinformation intervention\.

![Refer to caption](https://arxiv.org/html/2608.21721v1/Figure2.52.png)Figure 2:Overview of the RO\-PnR framework\.Top \(training\):the policy generates candidate actions conditioned on the query and hidden user state; a frozen reference and reward model score the dialogue, and the reward comparesRespond\(current quality\) againstProbe\(expected future quality minus cost\)\. Group\-normalized advantages then update the policy via GRPO\.Bottom \(inference\):the policy interacts with the user over multiple turns, deciding at each step whether to probe or produce a final adaptive response\.

## 3Methodology

### 3\.1Task Definition

We formulate the problem as a multi\-turn health misinformation intervention task under hidden user heterogeneity\. Given a health\-misinformation postxx, the agent interacts with a user over a short dialogue and follows a RO\-PnR paradigm\. At each turn, the agent may either*probe*to elicit missing users’ concern or*respond*with a final correction\.

### 3\.2RO\-PnR Policy

The policy casts intervention as a sequential decision problem: the agent must decide not only*when*to probe[59](https://arxiv.org/html/2608.21721#bib.bib48), but also*when*to stop gathering information and commit to a final response[7](https://arxiv.org/html/2608.21721#bib.bib47)\.

Let the dialogue history at turnttbe

ht=\{x,\(u1,a1\),…,\(ut−1,at−1\),ut\},h\_\{t\}=\\\{x,\(u\_\{1\},a\_\{1\}\),\\ldots,\(u\_\{t\-1\},a\_\{t\-1\}\),u\_\{t\}\\\},whereutu\_\{t\}denotes the current user utterance\. Conditioned onhth\_\{t\}, the agent selects an action

at∼π\(⋅∣ht\),a\_\{t\}\\sim\\pi\(\\cdot\\mid h\_\{t\}\),whereπ\\pidenotes the RO\-PnR policy and

at∈\{Probe,Respond\}\.a\_\{t\}\\in\\\{\\textsc\{Probe\},\\textsc\{Respond\}\\\}\.At each turn, the agent must decide whether the current information inhth\_\{t\}is sufficient to generate a final adaptive response, or whether asking one additional clarification question is likely to yield enough benefit to justify its interaction cost\. Accordingly, the policy choosesProbeonly when the expected value of acquiring additional user information outweighs the cost of continuing the dialogue; otherwise, it choosesRespond\.

### 3\.3Latent User Simulation

The RO\-PnR policy operates over the observed dialogue historyhth\_\{t\}\. However, the evolution of this history depends on how different users respond to the intervention\. To capture such hidden heterogeneity, we model each simulated user with a latent statez∈𝒵z\\in\\mathcal\{Z\}that governs how they react to the agent’s actions\.

Prior work identifies two dimensions as central to health misinformation intervention:*health literacy*[3](https://arxiv.org/html/2608.21721#bib.bib54);[48](https://arxiv.org/html/2608.21721#bib.bib55);[32](https://arxiv.org/html/2608.21721#bib.bib56)and*belief commitment*[8](https://arxiv.org/html/2608.21721#bib.bib57);[51](https://arxiv.org/html/2608.21721#bib.bib58);[27](https://arxiv.org/html/2608.21721#bib.bib59)\. Health literacy refers to a user’s ability to understand and process health\-related information, which directly affects how they interpret evidence and make health decisions[36](https://arxiv.org/html/2608.21721#bib.bib33)\. Belief commitment reflects the strength with which a user holds a misinformation\-related belief, as well as their resistance to corrective information[45](https://arxiv.org/html/2608.21721#bib.bib32);[54](https://arxiv.org/html/2608.21721#bib.bib41); These two dimensions are tightly coupled: A user’s response to correction depends on both their ability to understand the evidence and their willingness to revise prior beliefs\. Effective intervention must therefore account for both comprehension capacity and openness to belief revision\.

Motivated by these findings, we model each user along these two dimensions and define the latent state space as

𝒵=ℒ×ℬ,\\mathcal\{Z\}=\\mathcal\{L\}\\times\\mathcal\{B\},whereℒ\\mathcal\{L\}denotes health literacy andℬ\\mathcal\{B\}denotes belief commitment\. Following[35](https://arxiv.org/html/2608.21721#bib.bib43), we discretize health literacy as

ℒ=\{functional,interactive,critical\},\\mathcal\{L\}=\\\{\\textit\{functional\},\\ \\textit\{interactive\},\\ \\textit\{critical\}\\\},corresponding to increasing capacity to understand, apply, and critically evaluate health information\. Since there is no established categorization of belief commitment, we draw on prior work[45](https://arxiv.org/html/2608.21721#bib.bib32);[54](https://arxiv.org/html/2608.21721#bib.bib41)to define

ℬ=\{strong,hesitant,open\},\\mathcal\{B\}=\\\{\\textit\{strong\},\\ \\textit\{hesitant\},\\ \\textit\{open\}\\\},reflecting decreasing resistance to belief revision\. Together,\(L,B\)∈𝒵\(L,B\)\\in\\mathcal\{Z\}provide a compact characterization of user heterogeneity that supports adaptive intervention\. We provide detailed explanations of user modeling in Appendix[A](https://arxiv.org/html/2608.21721#A1)and Figure[3](https://arxiv.org/html/2608.21721#A1.F3)\.

### 3\.4Reward Design

We train the RO\-PnR policy with a turn\-level decision reward[58](https://arxiv.org/html/2608.21721#bib.bib30);[12](https://arxiv.org/html/2608.21721#bib.bib31)built in two layers, each addressing a distinct aspect of the probe\-or\-respond trade\-off\.

\(1\) Forward\-looking value\.The value of asking is not in the question itself, but in the better intervention it enables later in the dialogue\. A reward that scores only the next utterance misses this: a probe may add little immediately but unlock a substantially long horizon benefit, like better final response[55](https://arxiv.org/html/2608.21721#bib.bib45)\. We therefore compare the response quality available now against the expected quality at conversation end after probing:

rt​\(Respond\)\\displaystyle r\_\{t\}\(\\textsc\{Respond\}\)=R⁡\(ht\),\\displaystyle=R\(h\_\{t\}\),rt​\(Probe\)\\displaystyle r\_\{t\}\(\\textsc\{Probe\}\)=𝔼\[R\(hend\)∣ht,Probe\],\\displaystyle=\\mathbb\{E\}\\big\[\\,R\(h\_\{\\mathrm\{end\}\}\)\\mid h\_\{t\},\\,\\textsc\{Probe\}\\,\\big\],whereR⁡\(⋅\)R\(\\cdot\)scores a dialogueQualityusing the average of audience alignment, personalized grounding, and tailored actionability considering dialogue history and response, andhendh\_\{\\mathrm\{end\}\}is the state at which the agent eventually responds\. \(See Section[4\.3](https://arxiv.org/html/2608.21721#S4.SS3)forQualityevaluation details\.\)

\(2\) Interaction cost\.A forward\-looking gain alone treats every probe as free, which would incentivize the agent to ask whenever any improvement is expected, even arbitrarily small ones\. In practice, however, every clarification question imposes a real cost[7](https://arxiv.org/html/2608.21721#bib.bib47)\. An agent that ignores this cost will over\-probe, degrading user experience even when the marginal information gain is negligible\. We therefore subtract a per\-question cost from the probe action that scales with the number of clarifications already asked, referring to[7](https://arxiv.org/html/2608.21721#bib.bib47):

rt​\(Probe\)=Rfuture−c⋅nt,r\_\{t\}\(\\textsc\{Probe\}\)=R\_\{\\mathrm\{future\}\}\-c\\cdot n\_\{t\},wherec\>0c\>0is the per\-question cost andntn\_\{t\}is the number of clarification questions asked up to turntt\. This increases the cost of probing as the dialogue progresses, requiring each additional question to yield greater expected benefit to balance cumulative cost\. We further experiment with different cost selections in Section[6](https://arxiv.org/html/2608.21721#S6)\.

#### Putting it together\.

At each turn, the agent picks the action with the higher reward, choosingProbewhen

Rfuture​\(ht\)−R⁡\(ht\)\>c⋅nt,R\_\{\\mathrm\{future\}\}\(h\_\{t\}\)\-R\(h\_\{t\}\)\\;\>\\;c\\cdot n\_\{t\},andRespondotherwise\. Intuitively, the agent probes only when the expected long\-horizon improvement from one more clarification outweighs the cumulative interaction cost\. Early probes face a low bar, but later probes must justify a higher cost, encouraging the agent to gather information quickly and commit once further probing offers diminishing returns\.

### 3\.5RO\-PnR Optimization

We train the RO\-PnR policy on a dataset of*turn\-level decisions*rather than full dialogues, so that supervision aligns with the policy’s per\-step probe\-or\-respond choice\. At each decision state extracted from our multi\-turn rollouts, we build paired supervision: aRespondrecord scored by the immediate response quality, and aProberecord scored by the best continuation quality reachable after one further question\. Pairing the two alternatives at the same state lets the policy learn the local trade\-off directly\. Construction details \(rollout sources, grouping by post and profile, filtering\) are deferred to Appendix[B](https://arxiv.org/html/2608.21721#A2)\.

We first warm\-start the policy with supervised fine\-tuning on the best resolved trajectory in each group, then optimize it with Group Relative Policy Optimization \(GRPO\)[44](https://arxiv.org/html/2608.21721#bib.bib29)\. GRPO suits our binary turn\-level setting because it derives advantages by comparing candidate actions within the same decision group, eliminating the need for a separate value network\. Each group is scored with the reward from Section[3\.4](https://arxiv.org/html/2608.21721#S3.SS4), normalized into group\-relative advantages, and used to update the policy\. Because supervision operates directly at the turn level, the policy learns*when*to probe and*when*to respond at each step\.

## 4Experiment Setup

DatasetMethodLlama\-3\.1\-8B\-InstructQwen3\-8BGemma\-4\-E4B\-itAAPGTAQualityFER↓\\downarrowTurns↓\\downarrowUtilityAAPGTAQualityFER↓\\downarrowTurns↓\\downarrowUtilityAAPGTAQualityFER↓\\downarrowTurns↓\\downarrowUtilityCounterHealthSingle\-Turn0\.580\.530\.610\.570\.185\.000\.560\.580\.520\.610\.570\.155\.000\.560\.640\.590\.670\.630\.045\.000\.62Fixed\-Q0\.720\.650\.760\.710\.115\.000\.700\.710\.640\.730\.690\.085\.000\.680\.730\.660\.740\.710\.025\.000\.70Reactive0\.720\.640\.760\.700\.115\.000\.690\.710\.640\.720\.690\.095\.000\.680\.720\.650\.740\.710\.023\.920\.70Confidence0\.710\.640\.740\.690\.131\.260\.690\.670\.610\.700\.660\.061\.420\.660\.720\.640\.760\.710\.021\.850\.71SFT0\.700\.640\.760\.700\.074\.380\.690\.700\.640\.760\.700\.054\.220\.690\.700\.640\.760\.700\.094\.610\.69RO\-PnR \(Ours\)⋆\\star0\.720\.660\.770\.720\.063\.560\.710\.720\.670\.780\.720\.083\.590\.720\.730\.670\.770\.720\.093\.590\.72MisinfoCorrectSingle\-Turn0\.550\.500\.580\.540\.075\.000\.530\.550\.490\.590\.540\.055\.000\.540\.650\.600\.670\.640\.015\.000\.63Fixed\-Q0\.700\.630\.720\.690\.075\.000\.680\.710\.630\.720\.690\.035\.000\.680\.710\.640\.710\.690\.035\.000\.68Reactive0\.710\.640\.740\.700\.064\.980\.690\.710\.630\.710\.690\.035\.000\.680\.700\.630\.700\.680\.033\.960\.67Confidence0\.610\.540\.640\.600\.081\.100\.600\.590\.520\.640\.580\.031\.080\.580\.630\.550\.690\.620\.021\.350\.62SFT0\.710\.650\.760\.710\.064\.010\.700\.710\.650\.760\.700\.073\.910\.700\.710\.650\.760\.710\.084\.030\.70RO\-PnR \(Ours\)⋆\\star0\.730\.680\.770\.720\.073\.560\.720\.740\.680\.770\.730\.073\.460\.720\.740\.690\.770\.730\.083\.550\.73PUBHEALTHSingle\-Turn0\.630\.580\.650\.620\.245\.000\.610\.620\.560\.640\.610\.215\.000\.600\.660\.610\.680\.650\.095\.000\.64Fixed\-Q0\.730\.660\.770\.720\.135\.000\.710\.730\.670\.750\.720\.125\.000\.710\.740\.670\.770\.730\.015\.000\.72Reactive0\.740\.660\.780\.730\.144\.930\.720\.750\.680\.750\.730\.125\.000\.720\.740\.670\.770\.730\.013\.820\.72Confidence0\.690\.620\.720\.680\.191\.430\.680\.710\.640\.730\.690\.141\.630\.690\.750\.670\.790\.740\.062\.030\.74SFT0\.720\.660\.780\.720\.064\.230\.720\.730\.670\.780\.720\.054\.240\.720\.730\.670\.780\.720\.094\.600\.72RO\-PnR \(Ours\)⋆\\star0\.740\.690\.790\.740\.083\.530\.730\.740\.690\.790\.740\.083\.800\.730\.740\.690\.780\.740\.143\.570\.73

Table 1:Average results across all health\-literacy and belief\-commitment configurations, for each method, dataset, and base model\.AA,PG,TAare the Audience Alignment \(AA\), Personalized Grounding\(PG\), and Tailored Actionability \(TA\) scores \(normalized to\[0,1\]\[0,1\]\);Qualityis the overall quality of AA, PG and TA \(normalized to\[0,1\]\[0,1\]\);FERis the factual error rate \(↓\\downarrow\);Turnsis the average number of dialogue turns \(↓\\downarrow\); andUtilityis the cost\-adjusted Quality atc=0\.01c\{=\}0\.01, our headline metric \(normalized to\[0,1\]\[0,1\]\)\.### 4\.1Dataset

For fine\-tuning and evaluation, we construct multi\-turn datasets with a simulated user environment for health misinformation intervention, see user simulation in Section[A\.3](https://arxiv.org/html/2608.21721#A1.SS3)and dataset construction details in Appendix[B](https://arxiv.org/html/2608.21721#A2)\. Our primary dataset is derived from CounterHealth \(See Appendix[C](https://arxiv.org/html/2608.21721#A3)\), and is used for model fine\-tuning\. We additionally use two public health\-misinformation datasets for evaluation, MisinfoCorrect[14](https://arxiv.org/html/2608.21721#bib.bib50)and PUBHEALTH[24](https://arxiv.org/html/2608.21721#bib.bib14)\.

### 4\.2User Simulation

Due to the cost and difficulty of recruiting real users for large\-scale multi\-turn evaluation, we use an LLM\-based user simulator to generate user actions during interaction\. Specifically, we use GPT\-4o\-mini111[https://platform\.openai\.com/docs/models/gpt\-4o\-mini](https://platform.openai.com/docs/models/gpt-4o-mini), which has been increasingly adopted as a proxy for simulating user behavior in interactive dialogue settings[23](https://arxiv.org/html/2608.21721#bib.bib7);[26](https://arxiv.org/html/2608.21721#bib.bib8);[30](https://arxiv.org/html/2608.21721#bib.bib10);[29](https://arxiv.org/html/2608.21721#bib.bib9)\. In our setting, the simulator is prompted to emulate users with varying levels of health literacy and belief commitment, allowing us to evaluate how agents respond to different users\. The user\-simulation prompt is shown in Figure[6](https://arxiv.org/html/2608.21721#A5.F6), with additional details provided in Appendix[A\.3](https://arxiv.org/html/2608.21721#A1.SS3)\.

To evaluate whether the LLM\-based user simulator faithfully follows the assigned user profile, we conduct human evaluation along two dimensions: Persona Accuracy and Persona Consistency, referring to[53](https://arxiv.org/html/2608.21721#bib.bib21)\. Persona Accuracy measures whether the simulated user accurately reflects the assigned levels of health literacy and belief commitment\. Persona Consistency measures whether the simulated user maintains the assigned health literacy and belief commitment across the dialogue\. We instruct human annotators to rate each dimension on a 5\-point scale\. The detailed annotation rubric is provided in Table[6](https://arxiv.org/html/2608.21721#A1.T6)\. The results are discussed in Section[7](https://arxiv.org/html/2608.21721#S7)\.

### 4\.3Evaluation Metrics

We evaluate system performance along three complementary dimensions:*adaptation*,*factual reliability*, and*dialogue quality*, which together capture whether the system tailors the correction to the user, keeps it scientifically accurate, and reaches the goal efficiently\. This goes beyond asking whether the final response is correct and also asks whether it is delivered in a way that is adaptive and effective\.

#### Adaptation\.

Effective health misinformation correction depends not only on*what*is said but on*how*it is framed for a specific user[25](https://arxiv.org/html/2608.21721#bib.bib18);[34](https://arxiv.org/html/2608.21721#bib.bib15)\. We assess this with three sub\-metrics:Audience Alignment \(AA\)[46](https://arxiv.org/html/2608.21721#bib.bib52);[6](https://arxiv.org/html/2608.21721#bib.bib19), which measures whether the correction is framed appropriately for the user’s inferred health literacy and belief commitment;Personalized Grounding \(PG\)[11](https://arxiv.org/html/2608.21721#bib.bib17), which measures whether the evidence and reasoning supporting the correction are clear and comprehensible to the user; andTailored Actionability \(TA\)[46](https://arxiv.org/html/2608.21721#bib.bib52);[37](https://arxiv.org/html/2608.21721#bib.bib16), which measures whether the agent provides safe, concrete next steps suited to the user when needed\. All three are scored by an LLM judge \(mistral\-large\-2512222[https://docs\.mistral\.ai/models/model\-cards/mistral\-large\-3\-25\-12](https://docs.mistral.ai/models/model-cards/mistral-large-3-25-12)\) given the dialogue history and the user’s latent state\. \(See Figure[8](https://arxiv.org/html/2608.21721#A5.F8)in the Appendix for detailed rubrics\.\)

#### Factual reliability\.

A correction that introduces new inaccuracies can undermine user trust and lead to harmful decisions[56](https://arxiv.org/html/2608.21721#bib.bib49), so adaptation alone is not sufficient\. We report thefactual error rate: the fraction of responses flagged as containing any factual error by an LLM judge \(GPT\-5\-mini333[https://platform\.openai\.com/docs/models/gpt\-5\-mini](https://platform.openai.com/docs/models/gpt-5-mini)\) trained with updated knowledge and have access to web\-based verification[14](https://arxiv.org/html/2608.21721#bib.bib50)\. Lower values indicate better performance\. Prompt is in Figure[7](https://arxiv.org/html/2608.21721#A5.F7)in the Appendix\.

#### Dialogue quality\.

In addition to adaptation metrics, we also measure the efficiency of the interaction\.Turns[55](https://arxiv.org/html/2608.21721#bib.bib45)counts the number of communication turns, with fewer turns indicating a more efficient interaction\.Utility[7](https://arxiv.org/html/2608.21721#bib.bib47)captures the trade\-off between response quality and interaction cost\. Since the three adaptation dimensions are measured on the same scale and are jointly necessary for an adaptive response, we define overallQualityas their mean,Q=\(SAA\+SPG\+STA\)/3Q=\(S\_\{\\text\{AA\}\}\+S\_\{\\text\{PG\}\}\+S\_\{\\text\{TA\}\}\)/3, and utility asU=Q−cU=Q\-c, whereccis the cost incurred by probe actions\.

CounterHealthMisinfoPubhealthMethodModelFICOv\.FICQ\.FICOv\.Single\-TurnLlama0\.510\.600\.570\.560\.490\.570\.550\.530\.570\.640\.620\.61Qwen0\.500\.590\.590\.560\.480\.560\.560\.540\.540\.630\.620\.60Gemma0\.540\.660\.670\.620\.560\.670\.670\.630\.560\.670\.680\.64Fixed\-QLlama0\.660\.720\.720\.700\.620\.710\.700\.680\.670\.730\.730\.71Qwen0\.620\.710\.720\.680\.630\.710\.700\.680\.670\.720\.730\.71Gemma0\.640\.720\.740\.700\.630\.700\.700\.680\.670\.730\.750\.72ReactiveLlama0\.650\.720\.720\.690\.640\.720\.710\.690\.670\.730\.740\.72Qwen0\.620\.710\.720\.680\.620\.710\.700\.680\.670\.730\.740\.71Gemma0\.630\.720\.740\.700\.620\.700\.690\.670\.670\.730\.760\.72ConfidenceLlama0\.660\.720\.700\.690\.600\.630\.560\.600\.650\.710\.670\.68Qwen0\.590\.710\.680\.660\.570\.620\.560\.580\.640\.730\.700\.69Gemma0\.700\.720\.700\.710\.660\.640\.560\.620\.720\.740\.750\.74SFTLlama0\.640\.720\.720\.690\.640\.730\.730\.700\.670\.740\.740\.71Qwen0\.640\.720\.720\.690\.630\.730\.720\.700\.660\.740\.740\.72Gemma0\.630\.720\.720\.690\.640\.730\.740\.700\.660\.730\.750\.72RO\-PnRLlama0\.650\.740\.740\.710\.660\.750\.750\.720\.680\.760\.770\.73Qwen0\.660\.750\.740\.720\.660\.750\.750\.720\.670\.760\.770\.73Gemma0\.650\.750\.750\.720\.670\.750\.760\.730\.670\.750\.770\.73

Table 2:Utility by Health Literacy Level \(normalized to\[0,1\]\[0,1\]\)\. Higher\-value cells are highlighted with light tints: green \(≥0\.70\\geq 0\.70\), mint \(≥0\.73\\geq 0\.73\), and blue \(≥0\.75\\geq 0\.75\)\. F, I, C, Ov\., denotes Functional, Interactive, Critical, and all populations\.CounterHealthMisinfoPubhealthMethodModelSHOOv\.SHOOv\.SHOOv\.Single\-TurnLlama0\.500\.540\.640\.560\.480\.520\.610\.530\.550\.570\.710\.61Qwen0\.500\.540\.640\.560\.480\.520\.600\.540\.530\.570\.690\.60Gemma0\.570\.620\.690\.620\.580\.620\.690\.630\.580\.620\.710\.64Fixed\-QLlama0\.630\.710\.750\.700\.600\.690\.740\.680\.650\.720\.760\.71Qwen0\.610\.710\.730\.680\.600\.700\.730\.680\.650\.720\.750\.71Gemma0\.610\.730\.760\.700\.570\.710\.750\.680\.650\.740\.760\.72ReactiveLlama0\.620\.710\.760\.690\.620\.700\.750\.690\.660\.730\.760\.72Qwen0\.610\.700\.740\.680\.600\.690\.740\.680\.650\.730\.760\.72Gemma0\.620\.720\.760\.700\.570\.690\.750\.670\.650\.740\.770\.72ConfidenceLlama0\.650\.700\.730\.690\.580\.590\.620\.600\.650\.680\.710\.68Qwen0\.630\.660\.690\.660\.570\.570\.600\.580\.640\.700\.730\.69Gemma0\.670\.710\.730\.710\.610\.620\.640\.620\.690\.750\.770\.74SFTLlama0\.620\.710\.750\.690\.640\.720\.760\.700\.650\.730\.760\.71Qwen0\.630\.710\.750\.690\.630\.710\.750\.700\.660\.730\.760\.72Gemma0\.620\.710\.750\.690\.640\.710\.750\.700\.660\.730\.760\.72RO\-PnRLlama0\.650\.720\.760\.710\.660\.720\.770\.720\.690\.740\.770\.73Qwen0\.660\.720\.760\.720\.670\.730\.770\.720\.690\.740\.770\.73Gemma0\.670\.720\.760\.720\.690\.730\.770\.730\.690\.730\.770\.73

Table 3:Utility by Belief Commitment\. Higher\-value cells are highlighted with light tints: green \(≥0\.70\\geq 0\.70\), mint \(≥0\.73\\geq 0\.73\), and blue \(≥0\.75\\geq 0\.75\); lower values are left uncolored\. S, H, O, Ov\., denotes Strong, Hesitant, Open, and all populations\.

### 4\.4Baselines

We compare our RO\-PnR framework against four alternative decision strategies: \(a\)Single\-Turn Direct Response \(Single\-Turn\)[46](https://arxiv.org/html/2608.21721#bib.bib52): The model generates the correction immediately based only on the current turn, without asking clarification questions\. \(b\)Fixed\-Question Probing \(Fixed\-Q\)[10](https://arxiv.org/html/2608.21721#bib.bib12): The agent asks a predetermined number of clarification questions before producing the final counterspeech response, regardless of the user’s feedback\. \(c\)Reactive Clarifier \(Reactive\)[57](https://arxiv.org/html/2608.21721#bib.bib11): The agent adaptively chooses between asking and answering based on user feedback, without considering the user’s knowledge background or stance toward the claim\. \(d\)Confidence\-Gated Clarifier \(Confidence\)[28](https://arxiv.org/html/2608.21721#bib.bib13): The agent decides whether to ask or answer based on its confidence in a generic inferred user state, not grounded in the health literacy and belief commitment dimensions central to misinformation correction\. \(e\)Supervised Clarifier \(SFT\): A method that learns better to ask and answer via supervised fine\-tuning on a higher\-quality trajectory, without the turn\-level decision reward or GRPO optimization\.

### 4\.5Implementation Details

We evaluate RO\-PnR on three open\-source 8B models:Llama\-3\.1\-8B\-Instruct444[https://huggingface\.co/meta\-llama/Llama\-3\.1\-8B\-Instruct](https://huggingface.co/meta-llama/Llama-3.1-8B-Instruct),Gemma\-4\-E4B\-it555[https://huggingface\.co/google/gemma\-4\-E4B\-it](https://huggingface.co/google/gemma-4-E4B-it), andQwen3\-8B666[https://huggingface\.co/Qwen/Qwen3\-8B](https://huggingface.co/Qwen/Qwen3-8B)\. Fine\-tuning consists of two stages: offline supervised fine\-tuning \(SFT\) and offline GRPO[44](https://arxiv.org/html/2608.21721#bib.bib29)\. We first perform SFT on base models using full\-profile dialogue examples with LoRA adapters[15](https://arxiv.org/html/2608.21721#bib.bib28)\. We then initialize GRPO from the resulting SFT adapter and continue training on turn\-level preference examples constructed from all nine user\-profile combinations\. Training configurations are provided in Table[5](https://arxiv.org/html/2608.21721#A1.T5)\.

## 5Results

#### Main Results\.

RO\-PnR achieves the best overall utility while using substantially fewer interaction turns than always\-probe baselines \(Table[1](https://arxiv.org/html/2608.21721#S4.T1)\)\. RO\-PnR obtains the highest utility on nearly every \(dataset, model\) combination, reaching 0\.71–0\.73 in utility versus 0\.68\-0\.72 for the strongest baseline \(SFT\), with consistent gains across all three adaptation dimensions \(AA, PG, TA\)\. The baselines reveal a clear trade\-off\. Single\-Turn skips probing entirely and scores lowest \(0\.53\-0\.64\)\. Fixed\-Q and Reactive almost probe every turn \(≥4\.9\\geq 4\.9on average\) and reach competitive adaptation, but at a high interaction cost\. Confidence rarely \(1\.1\-2\.0 turns\) and suffers the largest quality drop, especially on MisinfoCorrect, where utility falls to 0\.58\-0\.62\. RO\-PnR occupies the productive middle, using roughly 3\.5 turns, about 30% fewer than the always\-probe baselines, while delivering the highest utility\. Factual error rates remain low across methods \(0\.06\-0\.14 for RO\-PnR\), indicating that adaptation gains do not compromise factual reliability\.

#### Performance Across User Profiles\.

RO\-PnR’s utility gains hold across user types, with the largest improvements on the hardest user slices\. Tables[2](https://arxiv.org/html/2608.21721#S4.T2)and[3](https://arxiv.org/html/2608.21721#S4.T3)break utility down by health literacy and belief commitment\. RO\-PnR achieves the highest overall utility in every \(dataset, model\) cell, and its advantage is most pronounced on the most challenging slices\. On Functional users, where adaptive framing matters most, RO\-PnR reaches 0\.65\-0\.69 versus 0\.62\-0\.67 for SFT\. On Strong\-believer users, the gap widens further: RO\-PnR obtains 0\.65–0\.69 versus 0\.6\-0\.66 for SFT and only 0\.57\-0\.61 for Confidence, indicating that indiscriminate probing fails to overcome resistance while confidence\-only probing fails to gather sufficient context\. As literacy rises and belief commitment weakens, all methods improve, and the gap narrows, e\.g\., on Open users, RO\-PnR reaches 0\.76\-0\.77 while baselines cluster in the 0\.73\-0\.76 range, but RO\-PnR retains the top spot throughout\. These breakdowns confirm that the utility gains in the main results are not driven by a single easy slice but hold across both the comprehension and the openness dimensions of user heterogeneity\.

## 6Ablation Study

VariantAA↑\\uparrowPG↑\\uparrowTA↑\\uparrowQuality↑\\uparrowFER↓\\downarrowTurns↓\\downarrowUtility↑\\uparrowRO\-PnR \(full\)0\.720\.660\.770\.720\.063\.560\.71\(a\) Reward\-horizon ablationimmediate next\-gain only0\.620\.540\.650\.600\.122\.000\.60\(b\) Probe\-cost \(cc\) sweepprobe cost=0\.00=0\.000\.500\.460\.570\.510\.115\.830\.51probe cost=0\.03=0\.030\.710\.650\.770\.710\.083\.000\.70probe cost=0\.05=0\.050\.720\.660\.770\.710\.093\.000\.69

Table 4:Ablation study results, averaged over all \(health literacy×\\timesbelief commitment\) user profiles\.RO\-PnR \(full\)is our complete method with forward\-looking long\-horizon reward, probe costc=0\.01c=0\.01, and cross\-threshold bonus\. \(a\) replaces the forward\-looking reward with the immediate next\-turn gain; \(b\) sweeps the per\-turn probe costcc\.To investigate how different components contribute to RO\-PnR performance, we conduct two ablations on the GRPO reward \(Table[4](https://arxiv.org/html/2608.21721#S6.T4)\)\.\(a\) Reward\-horizon ablationreplaces the forward\-looking reward with a one\-step reward based only on the next user response, testing whether short\-horizon credit assignment suffices\.\(b\) Probe\-cost sweepvaries the per\-question costccto probe the trade\-off between information gathering and early commitment\.

Table[4](https://arxiv.org/html/2608.21721#S6.T4)confirms that all components contribute to performance\. The reward\-horizon ablation degrades utility from 0\.71 to 0\.60, showing that a probe’s value often materializes several turns later and cannot be captured by a one\-step reward\. The probe\-cost sweep shows thatc=0c=0leads to over\-probing \(5\.83 turns\) and a quality collapse to 0\.51, while moderate costs \(c=0\.03c=0\.03\-0\.050\.05\) recover near\-best performance at∼3\{\\sim\}3turns, confirming that the policy is stable across a reasonable range ofcc\.

## 7Human Validation

While LLMs are known to reliably simulate humans and act as judges at low cost[49](https://arxiv.org/html/2608.21721#bib.bib27);[2](https://arxiv.org/html/2608.21721#bib.bib25);[16](https://arxiv.org/html/2608.21721#bib.bib24), we run a small\-scale human evaluation along three axes: \(1\) simulator reliability via persona accuracy and consistency; \(2\) human\-judge agreement on adaptation ratings againstMistral\-Large\-2512\(rubric in Figure[8](https://arxiv.org/html/2608.21721#A5.F8)\); and \(3\) a pairwise preference study comparingRO\-PnRagainst the baseline \(Due to the annotation cost, we include only the strong baseline method: Fixed\-Q\.\) with six users spanning distinct health literacy and belief commitment profiles\. Details in Appendix[D](https://arxiv.org/html/2608.21721#A4)\.

#### LLM Simulator Evaluation\.

Annotators show substantial agreement on both axes \(Table[7](https://arxiv.org/html/2608.21721#A1.T7)\): mean pairwise agreement is 0\.83 for persona accuracy and 0\.85 for consistency, with multi\-rater Scott’sπ\\piof 0\.70 and 0\.68\. Using the median of the three annotators’ ratings as the final score \(Table[8](https://arxiv.org/html/2608.21721#A1.T8)\), persona accuracy averages 3\.87, with 76% of samples scoring 4 or 5, indicating that simulated users generally reflect the assigned health literacy and belief commitment, though many are judged as largely rather than perfectly aligned\. Persona consistency is stronger, averaging 4\.64 with 97% of samples at 4 or above, showing that once a persona is established, the simulator maintains it across turns\. Details are in Appendix[D\.1](https://arxiv.org/html/2608.21721#A4.SS1)\.

#### Human\-judge agreement\.

Human annotators show moderate\-to\-strong internal consistency \(Scott’sπ\\pi= 0\.63–0\.78 across dimensions; Table[9](https://arxiv.org/html/2608.21721#A1.T9)\)\. Using the average human rating as consensus, the LLM judge aligns well with humans onQuality: Scott’sπ\\pi= 0\.55, 73% exact agreement, MAE = 0\.23, and 87% of samples within±0\.5\\pm 0\.5points \(Table[10](https://arxiv.org/html/2608.21721#A1.T10)\), with Pearson = 0\.57 and Spearman = 0\.72 \(Table[11](https://arxiv.org/html/2608.21721#A4.T11)\)\. This indicates that the LLM judge is a reliable proxy for human evaluation at scale\.

#### Human pairwise preference\.

As shown in Figure[4](https://arxiv.org/html/2608.21721#A4.F4), RO\-PnR is strongly preferred overall \(79\.70% vs\. 20\.30% forFixed\-Q\), primarily because its responses are perceived as more explanatory, credible, and actionable\. The exception is User 2 \(functional literacy, strong belief commitment\), who favorsFixed\-Qfor being easier to process, less confrontational, and less reliant on institutional authority, consistent with prior findings that users distrustful of official health sources resist authority\-heavy framings even when the evidence is richer[18](https://arxiv.org/html/2608.21721#bib.bib2)\. This highlights a trade\-off: whileRO\-PnRyields higher\-quality corrections in aggregate, users with strong prior commitments may benefit from softer, less institution\-centered framing to reduce resistance\.

## 8Conclusion

In this paper, we introduced RO\-PnR, a decision framework for multi\-turn health misinformation intervention that learns when to ask clarifying questions and when to correct\. By modeling hidden user differences in health literacy and belief commitment, RO\-PnR adapts its probing behavior to the needs of each dialogue rather than relying on fixed or excessive questioning\. Experiments show that RO\-PnR achieves stronger cost\-adjusted utility while using fewer interaction turns than always\-probe baselines\. These results suggest that effective misinformation correction requires not only accurate evidence, but also careful timing: asking when clarification is useful, and answering when enough context has been gathered\. Future work should extend this framework with real\-user interactions, richer models of user burden, and joint optimization of both when and how to probe\.

## Limitations

#### Simulated User Interaction\.

Because collecting large\-scale real\-user interactions is costly, we rely on simulated users as a proxy for human behavior\. While prior studies suggest that LLM\-based simulation can provide a useful approximation[53](https://arxiv.org/html/2608.21721#bib.bib21);[43](https://arxiv.org/html/2608.21721#bib.bib22);[4](https://arxiv.org/html/2608.21721#bib.bib20), and we make efforts to improve its reliability, it may still fall short of capturing the full nuance and variability of real users\. Despite this limitation, our simulation framework offers a practical testbed for studying how health\-misinformation\-related user characteristics affect multi\-turn counterspeech\. Future work should incorporate real\-user data and further refine both the simulator and the policy under more realistic interaction settings\.

#### Probe Action Constraints\.

Our study focuses primarily on the decision of*when*to probe and*when*to stop, emphasizing the probe\-and\-respond policy rather than the content of the probe itself\. As a result, the question of*how*to probe, namely, what clarification question is most appropriate at a particular stage of the dialogue, remains open\. Future work could address this limitation by jointly optimizing probe timing and probe content so that the agent can ask more adaptive and informative questions throughout the interaction\.

#### Simplified Cost Modeling\.

We model communication cost following[7](https://arxiv.org/html/2608.21721#bib.bib47)to account for the burden imposed by additional clarification turns\. However, this formulation is still coarse\-grained, assigning a fixed cost at the turn level rather than capturing the more nuanced cognitive load experienced by users\. Although our study provides insight into how agents behave when communication cost is incorporated into the decision process, future work might explore more refined ways of measuring user burden and integrating it into the policy for more cognitively aware interaction\.

## References

- Aniket al\.\(2025\)A\. S\. Anik, X\. Song, E\. Wang, B\. Wang, B\. Yarimbas, and L\. HongMulti\-agent retrieval\-augmented framework for evidence\-based counterspeech against health misinformation\.InSecond Conference on Language Modeling,External Links:[Link](https://openreview.net/forum?id=P61AgRyU7E)Cited by:[§1](https://arxiv.org/html/2608.21721#S1.p2.1),[§2\.1](https://arxiv.org/html/2608.21721#S2.SS1.p1.1)\.
- Bavarescoet al\.\(2025\)A\. Bavaresco, R\. Bernardi, L\. Bertolazzi, D\. Elliott, R\. Fernández, A\. Gatt, E\. Ghaleb, M\. Giulianelli, M\. Hanna, A\. Koller, A\. Martins, P\. Mondorf, V\. Neplenbroek, S\. Pezzelle, B\. Plank, D\. Schlangen, A\. Suglia, A\. K\. Surikuchi, E\. Takmaz, and A\. TestoniLLMs instead of human judges? a large scale empirical study across 20 NLP evaluation tasks\.InProceedings of the 63rd Annual Meeting of the Association for Computational Linguistics \(Volume 2: Short Papers\),W\. Che, J\. Nabende, E\. Shutova, and M\. T\. Pilehvar \(Eds\.\),Vienna, Austria,pp\. 238–255\.External Links:[Link](https://aclanthology.org/2025.acl-short.20/),[Document](https://dx.doi.org/10.18653/v1/2025.acl-short.20),ISBN 979\-8\-89176\-252\-7Cited by:[§7](https://arxiv.org/html/2608.21721#S7.p1.1)\.
- Berkmanet al\.\(2011\)N\. D\. Berkman, S\. L\. Sheridan, K\. E\. Donahue, D\. J\. Halpern, and K\. CrottyLow health literacy and health outcomes: an updated systematic review\.Annals of internal medicine155\(2\),pp\. 97–107\.Cited by:[§1](https://arxiv.org/html/2608.21721#S1.p4.1),[§3\.3](https://arxiv.org/html/2608.21721#S3.SS3.p2.1)\.
- Bougie and Watanabe \(2025\)N\. Bougie and N\. WatanabeSimuser: simulating user behavior with large language models for recommender system evaluation\.InProceedings of the 63rd Annual Meeting of the Association for Computational Linguistics \(Volume 6: Industry Track\),pp\. 43–60\.Cited by:[Simulated User Interaction\.](https://arxiv.org/html/2608.21721#Sx1.SS0.SSS0.Px1.p1.1)\.
- Chenet al\.\(2024\)M\. Chen, R\. Sun, T\. Pfister, and S\. Ö\. ArıkLearning to clarify: multi\-turn conversations with action\-based contrastive self\-training\.arXiv preprint arXiv:2406\.00222\.Cited by:[§1](https://arxiv.org/html/2608.21721#S1.p2.1),[§1](https://arxiv.org/html/2608.21721#S1.p3.1),[§2\.2](https://arxiv.org/html/2608.21721#S2.SS2.p1.1)\.
- Cimaet al\.\(2025\)L\. Cima, A\. Miaschi, A\. Trujillo, M\. Avvenuti, F\. Dell’Orletta, and S\. CresciContextualized counterspeech: strategies for adaptation, personalization, and evaluation\.InProceedings of the ACM on Web Conference 2025,pp\. 5022–5033\.Cited by:[§4\.3](https://arxiv.org/html/2608.21721#S4.SS3.SSS0.Px1.p1.1)\.
- Donget al\.\(2026\)Y\. R\. Dong, T\. Hu, Z\. Hui, C\. Zhang, I\. Vulić, A\. Bobu, and N\. CollierValue of information: a framework for human\-agent communication\.arXiv preprint arXiv:2601\.06407\.Cited by:[§1](https://arxiv.org/html/2608.21721#S1.p2.1),[§1](https://arxiv.org/html/2608.21721#S1.p3.1),[§2\.2](https://arxiv.org/html/2608.21721#S2.SS2.p1.1),[§3\.2](https://arxiv.org/html/2608.21721#S3.SS2.p1.1),[§3\.4](https://arxiv.org/html/2608.21721#S3.SS4.p3.1),[§4\.3](https://arxiv.org/html/2608.21721#S4.SS3.SSS0.Px3.p1.1),[Simplified Cost Modeling\.](https://arxiv.org/html/2608.21721#Sx1.SS0.SSS0.Px3.p1.1)\.
- Eckeret al\.\(2022a\)U\. K\. H\. Ecker, S\. Lewandowsky, J\. Cook, P\. Schmid, L\. K\. Fazio, N\. Brashier, P\. Kendeou, E\. K\. Vraga, and M\. A\. AmazeenThe psychological drivers of misinformation belief and its resistance to correction\.Nature Reviews Psychology1\(1\),pp\. 13–29\.External Links:ISSN 2731\-0574,[Link](https://doi.org/10.1038/s44159-021-00006-y),[Document](https://dx.doi.org/10.1038/s44159-021-00006-y)Cited by:[§3\.3](https://arxiv.org/html/2608.21721#S3.SS3.p2.1)\.
- Eckeret al\.\(2022b\)U\. K\. Ecker, S\. Lewandowsky, J\. Cook, P\. Schmid, L\. K\. Fazio, N\. Brashier, P\. Kendeou, E\. K\. Vraga, and M\. A\. AmazeenThe psychological drivers of misinformation belief and its resistance to correction\.Nature Reviews Psychology1\(1\),pp\. 13–29\.Cited by:[§1](https://arxiv.org/html/2608.21721#S1.p1.1),[§1](https://arxiv.org/html/2608.21721#S1.p4.1),[§2\.3](https://arxiv.org/html/2608.21721#S2.SS3.p1.1)\.
- Fu and Du \(2025\)C\. Fu and Y\. DuFirst ask then answer: a framework design for ai dialogue based on supplementary questioning with large language models\.arXiv preprint arXiv:2508\.08308\.Cited by:[§2\.2](https://arxiv.org/html/2608.21721#S2.SS2.p1.1),[§4\.4](https://arxiv.org/html/2608.21721#S4.SS4.p1.1)\.
- Gabrielet al\.\(2024\)S\. Gabriel, L\. Lyu, J\. Siderius, M\. Ghassemi, J\. Andreas, and A\. E\. OzdaglarMisinfoEval: generative ai in the era of “alternative facts”\.InProceedings of the 2024 Conference on Empirical Methods in Natural Language Processing,pp\. 8566–8578\.Cited by:[§4\.3](https://arxiv.org/html/2608.21721#S4.SS3.SSS0.Px1.p1.1)\.
- Gaoet al\.\(2024\)Z\. Gao, W\. Zhan, J\. D\. Chang, G\. Swamy, K\. Brantley, J\. D\. Lee, and W\. SunRegressing the relative future: efficient policy optimization for multi\-turn rlhf\.arXiv preprint arXiv:2410\.04612\.Cited by:[§3\.4](https://arxiv.org/html/2608.21721#S3.SS4.p1.1)\.
- Guigonet al\.\(2026\)V\. Guigon, L\. Geay, and C\. J\. CharpentierRethinking misinformation through plausibility estimation and confidence calibration\.Communications Psychology4\(1\),pp\. 24\.Cited by:[§A\.2](https://arxiv.org/html/2608.21721#A1.SS2.p2.1)\.
- Heet al\.\(2023\)B\. He, M\. Ahamad, and S\. KumarReinforcement learning\-based counter\-misinformation response generation: a case study of covid\-19 vaccine misinformation\.InProceedings of the ACM Web Conference 2023,pp\. 2698–2709\.Cited by:[§2\.1](https://arxiv.org/html/2608.21721#S2.SS1.p1.1),[§4\.1](https://arxiv.org/html/2608.21721#S4.SS1.p1.1),[§4\.3](https://arxiv.org/html/2608.21721#S4.SS3.SSS0.Px2.p1.1)\.
- Huet al\.\(2021\)E\. J\. Hu, Y\. Shen, P\. Wallis, Z\. Allen\-Zhu, Y\. Li, S\. Wang, L\. Wang, and W\. ChenLoRA: low\-rank adaptation of large language models\.External Links:2106\.09685,[Link](https://arxiv.org/abs/2106.09685)Cited by:[§4\.5](https://arxiv.org/html/2608.21721#S4.SS5.p1.1)\.
- Huanget al\.\(2025\)H\. Huang, X\. Bu, H\. Zhou, Y\. Qu, J\. Liu, M\. Yang, B\. Xu, and T\. ZhaoAn empirical study of llm\-as\-a\-judge for llm evaluation: fine\-tuned judge model is not a general substitute for gpt\-4\.InFindings of the Association for Computational Linguistics: ACL 2025,pp\. 5880–5895\.Cited by:[§7](https://arxiv.org/html/2608.21721#S7.p1.1)\.
- Ishikawaet al\.\(2008\)H\. Ishikawa, T\. Takeuchi, and E\. YanoMeasuring functional, communicative, and critical health literacy among diabetic patients\.Diabetes care31\(5\),pp\. 874–879\.Cited by:[§D\.3](https://arxiv.org/html/2608.21721#A4.SS3.p1.1),[§D\.3](https://arxiv.org/html/2608.21721#A4.SS3.p2.1)\.
- Jamisonet al\.\(2019\)A\. M\. Jamison, S\. C\. Quinn, and V\. S\. Freimuth“You don’t trust a government vaccine”: narratives of institutional trust and influenza vaccination among african american and white adults\.Social science & medicine221,pp\. 87–94\.Cited by:[§7](https://arxiv.org/html/2608.21721#S7.SS0.SSS0.Px3.p1.1)\.
- Jongman\-Serenoet al\.\(2023\)K\. P\. Jongman\-Sereno, R\. H\. Hoyle, E\. K\. Davisson, and J\. ParkIntellectual humility and responsiveness to public health recommendations\.Personality and individual differences211,pp\. 112243\.Cited by:[§A\.2](https://arxiv.org/html/2608.21721#A1.SS2.p3.1)\.
- Josephet al\.\(2025\)S\. Joseph, L\. Chen, B\. Wei, M\. Mackert, I\. J\. Marshall, P\. P\. Liang, R\. Kouzy, B\. C\. Wallace, and J\. J\. LiDecide less, communicate more: on the construct validity of end\-to\-end fact\-checking in medicine\.arXiv preprint arXiv:2506\.20876\.Cited by:[§2\.1](https://arxiv.org/html/2608.21721#S2.SS1.p1.1)\.
- Karinshaket al\.\(2023\)E\. Karinshak, S\. X\. Liu, J\. S\. Park, and J\. T\. HancockWorking with ai to persuade: examining a large language model’s ability to generate pro\-vaccination messages\.Proceedings of the ACM on Human\-Computer Interaction7\(CSCW1\),pp\. 1–29\.Cited by:[§1](https://arxiv.org/html/2608.21721#S1.p1.1)\.
- Kempet al\.\(2024\)P\. L\. Kemp, A\. C\. Goldman, and C\. N\. WahlheimOn the role of memory in misinformation corrections: repeated exposure, correction durability, and source credibility\.Current Opinion in Psychology56,pp\. 101783\.Cited by:[§A\.2](https://arxiv.org/html/2608.21721#A1.SS2.p2.1)\.
- Kimet al\.\(2025\)M\. Kim, D\. Yoo, Y\. Hwang, M\. Kang, N\. Kim, M\. Gwak, B\. Kwak, H\. Chae, H\. Kim, Y\. Lee, M\. H\. Kim, D\. Jung, K\. Chung, and J\. YeoCan you share your story? modeling clients’ metacognition and openness for LLM therapist evaluation\.InFindings of the Association for Computational Linguistics: ACL 2025,W\. Che, J\. Nabende, E\. Shutova, and M\. T\. Pilehvar \(Eds\.\),Vienna, Austria,pp\. 25943–25962\.External Links:[Link](https://aclanthology.org/2025.findings-acl.1332/),[Document](https://dx.doi.org/10.18653/v1/2025.findings-acl.1332),ISBN 979\-8\-89176\-256\-5Cited by:[§4\.2](https://arxiv.org/html/2608.21721#S4.SS2.p1.1)\.
- Kotonya and Toni \(2020\)N\. Kotonya and F\. ToniExplainable automated fact\-checking for public health claims\.InProceedings of the 2020 Conference on Empirical Methods in Natural Language Processing \(EMNLP\),pp\. 7740–7754\.Cited by:[§4\.1](https://arxiv.org/html/2608.21721#S4.SS1.p1.1)\.
- Krishnamurthy and Hu \(2026\)P\. Krishnamurthy and Y\. HuCOVID\-19 vaccine framing and acceptance among adults who are vaccine hesitant\.JAMA Network Open9\(3\),pp\. e264114\.Cited by:[§4\.3](https://arxiv.org/html/2608.21721#S4.SS3.SSS0.Px1.p1.1)\.
- Labanet al\.\(2025\)P\. Laban, H\. Hayashi, Y\. Zhou, and J\. NevilleLlms get lost in multi\-turn conversation\.arXiv preprint arXiv:2505\.06120\.Cited by:[§4\.2](https://arxiv.org/html/2608.21721#S4.SS2.p1.1)\.
- Lewandowskyet al\.\(2012\)S\. Lewandowsky, U\. K\. H\. Ecker, C\. M\. Seifert, N\. Schwarz, and J\. CookMisinformation and its correction: continued influence and successful debiasing\.Psychological Science in the Public Interest13\(3\),pp\. 106–131\.Note:PMID: 26173286External Links:[Document](https://dx.doi.org/10.1177/1529100612451018),[Link](https://doi.org/10.1177/1529100612451018),https://doi\.org/10\.1177/1529100612451018Cited by:[§1](https://arxiv.org/html/2608.21721#S1.p4.1),[§3\.3](https://arxiv.org/html/2608.21721#S3.SS3.p2.1)\.
- Liet al\.\(2024\)S\. S\. Li, V\. Balachandran, S\. Feng, J\. S\. Ilgen, E\. Pierson, P\. W\. Koh, and Y\. TsvetkovMediQ: question\-asking llms and a benchmark for reliable interactive clinical reasoning\.InAdvances in Neural Information Processing Systems,A\. Globerson, L\. Mackey, D\. Belgrave, A\. Fan, U\. Paquet, J\. Tomczak, and C\. Zhang \(Eds\.\),Vol\.37,pp\. 28858–28888\.External Links:[Document](https://dx.doi.org/10.52202/079017-0908),[Link](https://proceedings.neurips.cc/paper_files/paper/2024/file/32b80425554e081204e5988ab1c97e9a-Paper-Conference.pdf)Cited by:[§1](https://arxiv.org/html/2608.21721#S1.p2.1),[§2\.2](https://arxiv.org/html/2608.21721#S2.SS2.p1.1),[§4\.4](https://arxiv.org/html/2608.21721#S4.SS4.p1.1)\.
- Linet al\.\(2024\)X\. Lin, X\. Yu, A\. Aich, S\. Giorgi, and L\. UngarDiversedialogue: a methodology for designing chatbots with human\-like diversity\.arXiv preprint arXiv:2409\.00262\.Cited by:[§4\.2](https://arxiv.org/html/2608.21721#S4.SS2.p1.1)\.
- Liuet al\.\(2026\)Y\. L\. Liu, H\. Yun, T\. Roosta, and Z\. XiaoSynthetic users, real differences: an evaluation framework for user simulation in multi\-turn conversations\.arXiv preprint arXiv:2605\.02624\.Cited by:[§4\.2](https://arxiv.org/html/2608.21721#S4.SS2.p1.1)\.
- Metz \(2023\)S\. E\. MetzBuilding better beliefs through actively open\-minded thinking\.The cognitive science of belief,pp\. 574–591\.Cited by:[§A\.2](https://arxiv.org/html/2608.21721#A1.SS2.p3.1)\.
- Nanet al\.\(2022\)X\. Nan, Y\. Wang, and K\. ThierWhy do people believe health misinformation and who is at risk? a systematic review of individual differences in susceptibility to health misinformation\.Social Science & Medicine314,pp\. 115398\.External Links:ISSN 0277\-9536,[Document](https://dx.doi.org/https%3A//doi.org/10.1016/j.socscimed.2022.115398),[Link](https://www.sciencedirect.com/science/article/pii/S0277953622007043)Cited by:[§2\.3](https://arxiv.org/html/2608.21721#S2.SS3.p1.1),[§3\.3](https://arxiv.org/html/2608.21721#S3.SS3.p2.1)\.
- Newmanet al\.\(2022\)E\. J\. Newman, B\. Swire\-Thompson, and U\. K\. EckerMisinformation and the sins of memory: false\-belief formation and limits on belief revision\.\.Cited by:[§A\.2](https://arxiv.org/html/2608.21721#A1.SS2.p1.1)\.
- Nguyenet al\.\(2019\)M\. H\. Nguyen, E\. M\. Smets, N\. Bol, E\. F\. Loos, H\. W\. van Laarhoven, D\. Geijsen, M\. I\. van Berge Henegouwen, K\. M\. Tytgat, and J\. C\. Van WeertTailored web\-based information for younger and older patients with cancer: randomized controlled trial of a preparatory educational intervention on patient outcomes\.Journal of Medical Internet Research21\(10\),pp\. e14407\.Cited by:[§4\.3](https://arxiv.org/html/2608.21721#S4.SS3.SSS0.Px1.p1.1)\.
- Nutbeam \(2000\)D\. NutbeamHealth literacy as a public health goal: a challenge for contemporary health education and communication strategies into the 21st century\.Health promotion international15\(3\),pp\. 259–267\.Cited by:[§A\.1](https://arxiv.org/html/2608.21721#A1.SS1.p1.1),[§A\.1](https://arxiv.org/html/2608.21721#A1.SS1.p2.1),[§A\.1](https://arxiv.org/html/2608.21721#A1.SS1.p3.1),[§A\.3](https://arxiv.org/html/2608.21721#A1.SS3.p2.1),[§D\.3](https://arxiv.org/html/2608.21721#A4.SS3.p2.1),[§1](https://arxiv.org/html/2608.21721#S1.p4.1),[§3\.3](https://arxiv.org/html/2608.21721#S3.SS3.p3.2)\.
- Ogbadu\-Oladapoet al\.\(2026\)L\. Ogbadu\-Oladapo, K\. Bissadu, H\. Kim, and D\. L\. SmithInformation and health literacy: could there be any impact on health decision\-making among adults?—evidence from north america\.Journal of Public Health34\(1\),pp\. 151–181\.Cited by:[§3\.3](https://arxiv.org/html/2608.21721#S3.SS3.p2.1)\.
- Ownbyet al\.\(2026\)R\. L\. Ownby, R\. Davenport, and J\. CaballeroUser perceptions of individually\-tailored health information in digital apps: development of a scale\.Frontiers in Digital Health8,pp\. 1731948\.Cited by:[§4\.3](https://arxiv.org/html/2608.21721#S4.SS3.SSS0.Px1.p1.1)\.
- Peng and Grimmelmann \(2024\)K\. Peng and J\. GrimmelmannRescuing counterspeech: a bridging\-based approach to combating misinformation\.arXiv preprint arXiv:2410\.12699\.Cited by:[§1](https://arxiv.org/html/2608.21721#S1.p2.1),[§2\.1](https://arxiv.org/html/2608.21721#S2.SS1.p1.1),[§2\.3](https://arxiv.org/html/2608.21721#S2.SS3.p1.1)\.
- Salemiet al\.\(2024\)A\. Salemi, S\. Mysore, M\. Bendersky, and H\. ZamaniLamp: when large language models meet personalization\.InProceedings of the 62nd Annual Meeting of the Association for Computational Linguistics \(Volume 1: Long Papers\),pp\. 7370–7392\.Cited by:[§2\.3](https://arxiv.org/html/2608.21721#S2.SS3.p1.1)\.
- Sanna and Lagnado \(2025\)G\. A\. Sanna and D\. LagnadoBelief updating in the face of misinformation: the role of source reliability\.Cognition258,pp\. 106090\.Cited by:[§A\.2](https://arxiv.org/html/2608.21721#A1.SS2.p3.1)\.
- Schulz and Nakamoto \(2025\)P\. J\. Schulz and K\. NakamotoUnderstanding health knowledge failures: uncertainty versus misinformation\.Scientific Reports15\(1\),pp\. 23867\.Cited by:[§A\.2](https://arxiv.org/html/2608.21721#A1.SS2.p2.1)\.
- Scott \(1955\)W\. A\. ScottReliability of content analysis: the case of nominal scale coding\.Public opinion quarterly,pp\. 321–325\.Cited by:[§D\.2](https://arxiv.org/html/2608.21721#A4.SS2.p1.1)\.
- Sekulićet al\.\(2024\)I\. Sekulić, S\. Terragni, V\. Guimarães, N\. Khau, B\. Guedes, M\. Filipavicius, A\. F\. Manso, and R\. MathisReliable llm\-based user simulator for task\-oriented dialogue systems\.InProceedings of the 1st Workshop on Simulating Conversational Intelligence in Chat \(SCI\-CHAT 2024\),pp\. 19–35\.Cited by:[Simulated User Interaction\.](https://arxiv.org/html/2608.21721#Sx1.SS0.SSS0.Px1.p1.1)\.
- Shaoet al\.\(2024\)Z\. Shao, P\. Wang, Q\. Zhu, R\. Xu, J\. Song, X\. Bi, H\. Zhang, M\. Zhang, Y\. K\. Li, Y\. Wu, and D\. GuoDeepSeekMath: pushing the limits of mathematical reasoning in open language models\.External Links:2402\.03300,[Link](https://arxiv.org/abs/2402.03300)Cited by:[§3\.5](https://arxiv.org/html/2608.21721#S3.SS5.p2.1),[§4\.5](https://arxiv.org/html/2608.21721#S4.SS5.p1.1)\.
- Siebert and Siebert \(2023\)J\. Siebert and J\. U\. SiebertEffective mitigation of the belief perseverance bias after the retraction of misinformation: awareness training and counter\-speech\.Plos one18\(3\),pp\. e0282202\.Cited by:[§3\.3](https://arxiv.org/html/2608.21721#S3.SS3.p2.1),[§3\.3](https://arxiv.org/html/2608.21721#S3.SS3.p3.3)\.
- Songet al\.\(2025\)X\. Song, A\. S\. Anik, D\. Barua, P\. Luo, J\. Ding, and L\. HongSpeaking at the right level: literacy\-controlled counterspeech generation with RAG\-RL\.InFindings of the Association for Computational Linguistics: EMNLP 2025,C\. Christodoulopoulos, T\. Chakraborty, C\. Rose, and V\. Peng \(Eds\.\),Suzhou, China,pp\. 2812–2830\.External Links:[Link](https://aclanthology.org/2025.findings-emnlp.153/),[Document](https://dx.doi.org/10.18653/v1/2025.findings-emnlp.153),ISBN 979\-8\-89176\-335\-7Cited by:[§1](https://arxiv.org/html/2608.21721#S1.p2.1),[§1](https://arxiv.org/html/2608.21721#S1.p4.1),[§2\.1](https://arxiv.org/html/2608.21721#S2.SS1.p1.1),[§2\.3](https://arxiv.org/html/2608.21721#S2.SS3.p1.1),[§4\.3](https://arxiv.org/html/2608.21721#S4.SS3.SSS0.Px1.p1.1),[§4\.4](https://arxiv.org/html/2608.21721#S4.SS4.p1.1)\.
- Swire\-Thompsonet al\.\(2022\)B\. Swire\-Thompson, N\. Miklaucic, J\. P\. Wihbey, D\. Lazer, and J\. DeGutisThe backfire effect after correcting misinformation is strongly associated with reliability\.\.Journal of Experimental Psychology: General151\(7\),pp\. 1655\.Cited by:[§A\.2](https://arxiv.org/html/2608.21721#A1.SS2.p1.1),[§1](https://arxiv.org/html/2608.21721#S1.p4.1)\.
- Sørensenet al\.\(2012\)K\. Sørensen, S\. Van den Broucke, J\. Fullam, G\. Doyle, J\. Pelikan, Z\. Slonska, H\. Brand, and \(HLS\-EU\) Consortium Health Literacy Project EuropeanHealth literacy and public health: A systematic review and integration of definitions and models\.BMC Public Health12\(1\),pp\. 80\.External Links:ISSN 1471\-2458,[Link](https://doi.org/10.1186/1471-2458-12-80),[Document](https://dx.doi.org/10.1186/1471-2458-12-80)Cited by:[§3\.3](https://arxiv.org/html/2608.21721#S3.SS3.p2.1)\.
- Thakuret al\.\(2025\)A\. S\. Thakur, K\. Choudhary, V\. S\. Ramayapally, S\. Vaidyanathan, and D\. HupkesJudging the judges: evaluating alignment and vulnerabilities in llms\-as\-judges\.InProceedings of the Fourth Workshop on Generation, Evaluation and Metrics \(GEM2\{\}^\{2\}\),pp\. 404–430\.Cited by:[§D\.2](https://arxiv.org/html/2608.21721#A4.SS2.p1.1),[§7](https://arxiv.org/html/2608.21721#S7.p1.1)\.
- Van Der Linden \(2022\)S\. Van Der LindenMisinformation: susceptibility, spread, and interventions to immunize the public\.Nature medicine28\(3\),pp\. 460–467\.Cited by:[§1](https://arxiv.org/html/2608.21721#S1.p1.1)\.
- Walter and Murphy \(2018\)N\. Walter and S\. T\. MurphyHow to unring the bell: a meta\-analytic approach to correction of misinformation\.Communication Monographs85\(3\),pp\. 423–441\.External Links:[Document](https://dx.doi.org/10.1080/03637751.2018.1467564),[Link](https://doi.org/10.1080/03637751.2018.1467564),https://doi\.org/10\.1080/03637751\.2018\.1467564Cited by:[§3\.3](https://arxiv.org/html/2608.21721#S3.SS3.p2.1)\.
- Wanet al\.\(2025\)Y\. Wan, J\. Wu, M\. Abdulhai, L\. Shani, and N\. JaquesEnhancing personalized multi\-turn dialogue with curiosity reward\.arXiv preprint arXiv:2504\.03206\.Cited by:[§2\.2](https://arxiv.org/html/2608.21721#S2.SS2.p1.1)\.
- Wanget al\.\(2025\)K\. Wang, X\. Li, S\. Yang, L\. Zhou, F\. Jiang, and H\. LiKnow you first and be you better: modeling human\-like user simulators via implicit profiles\.InProceedings of the 63rd Annual Meeting of the Association for Computational Linguistics \(Volume 1: Long Papers\),pp\. 21082–21107\.Cited by:[§2\.3](https://arxiv.org/html/2608.21721#S2.SS3.p1.1),[§4\.2](https://arxiv.org/html/2608.21721#S4.SS2.p2.1),[Simulated User Interaction\.](https://arxiv.org/html/2608.21721#Sx1.SS0.SSS0.Px1.p1.1)\.
- Wittenberg and Berinsky \(2020\)C\. Wittenberg and A\. J\. BerinskyMisinformation and its correction\.Social media and democracy: The state of the field, prospects for reform,pp\. 163–198\.Cited by:[§A\.2](https://arxiv.org/html/2608.21721#A1.SS2.p1.1),[§1](https://arxiv.org/html/2608.21721#S1.p4.1),[§3\.3](https://arxiv.org/html/2608.21721#S3.SS3.p2.1),[§3\.3](https://arxiv.org/html/2608.21721#S3.SS3.p3.3)\.
- Wuet al\.\(2025\)S\. Wu, M\. Galley, B\. Peng, H\. Cheng, G\. Li, Y\. Dou, W\. Cai, J\. Zou, J\. Leskovec, and J\. GaoCollabLLM: from passive responders to active collaborators\.InInternational Conference on Machine Learning,pp\. 67260–67283\.Cited by:[§1](https://arxiv.org/html/2608.21721#S1.p2.1),[§2\.2](https://arxiv.org/html/2608.21721#S2.SS2.p1.1),[§3\.4](https://arxiv.org/html/2608.21721#S3.SS4.p2.1),[§4\.3](https://arxiv.org/html/2608.21721#S4.SS3.SSS0.Px3.p1.1)\.
- Yueet al\.\(2024\)Z\. Yue, H\. Zeng, Y\. Lu, L\. Shang, Y\. Zhang, and D\. WangEvidence\-driven retrieval augmented response generation for online misinformation\.InProceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies \(Volume 1: Long Papers\),pp\. 5628–5643\.Cited by:[§1](https://arxiv.org/html/2608.21721#S1.p2.1),[§2\.1](https://arxiv.org/html/2608.21721#S2.SS1.p1.1),[§4\.3](https://arxiv.org/html/2608.21721#S4.SS3.SSS0.Px2.p1.1)\.
- Zhang and Choi \(2025\)M\. J\. Zhang and E\. ChoiClarify when necessary: resolving ambiguity through interaction with lms\.InFindings of the Association for Computational Linguistics: NAACL 2025,pp\. 5526–5543\.Cited by:[§1](https://arxiv.org/html/2608.21721#S1.p3.1),[§2\.2](https://arxiv.org/html/2608.21721#S2.SS2.p1.1),[§4\.4](https://arxiv.org/html/2608.21721#S4.SS4.p1.1)\.
- Zhou and Zanette \(2024\)Y\. Zhou and A\. ZanetteArCHer: training language model agents via hierarchical multi\-turn rl\.InProceedings of the 41st International Conference on Machine Learning,pp\. 62178–62209\.Cited by:[§3\.4](https://arxiv.org/html/2608.21721#S3.SS4.p1.1)\.
- Zhuet al\.\(2025\)J\. Zhu, J\. Pan, Y\. Liu, F\. Liu, and J\. WuAsk patients with patience: enabling llms for human\-centric medical dialogue with grounded reasoning\.InProceedings of the 2025 Conference on Empirical Methods in Natural Language Processing,pp\. 2846–2857\.Cited by:[§3\.2](https://arxiv.org/html/2608.21721#S3.SS2.p1.1)\.

## Appendix AUser Modeling

### A\.1Health literacy

Functional health literacymeans basic reading/writing and understanding skills sufficient to function effectively in everyday health situations, such as understanding health information and service instructions[35](https://arxiv.org/html/2608.21721#bib.bib43)\.

Interactive health literacyis more advanced personal, communicative, and social skills that enable independent action on health knowledge[35](https://arxiv.org/html/2608.21721#bib.bib43)\.

Critical health literacyThis is the most advanced level\.[35](https://arxiv.org/html/2608.21721#bib.bib43)defines it in terms of advanced cognitive and social skills used to critically analyze information and act on broader determinants of health at individual and community levels\.

![Refer to caption](https://arxiv.org/html/2608.21721v1/Appendix_user_1.1.png)Figure 3:3×\\times3 latent user profile grid used in our simulation environment\.Users are characterized along two dimensions:*health literacy*\(Functional,Interactive,Critical\) and*belief commitment*\(Open,Hesitant,Strong believer\)\. Each cell defines one latent user type and includes an illustrative response style showing how the user may express concern, interpret evidence, and respond to counterspeech\. This profile grid is used to simulate heterogeneous user behavior during multi\-turn interaction\.StageLREpochsBatchGrad\. accum\.Max len\.Other settingsSFT5×10−55\\times 10^\{\-5\}4182048LoRAr=16r=16,α=32\\alpha=32, dropout0\.050\.05; bf16; 100 warmup steps; seed 42Offline GRPO5×10−65\\times 10^\{\-6\}2182048βKL=0\.05\\beta\_\{\\mathrm\{KL\}\}=0\.05; SFT CE weight0\.20\.2; warmup ratio0\.030\.03; max grad norm1\.01\.0; seed 42

Table 5:Hyperparameters for SFT and GRPO configuration
### A\.2Belief Commitment

Strong belief commitmentcan be understood as a state in which individuals accept misinformation as true, allow it to continue shaping subsequent attitudes and judgments[33](https://arxiv.org/html/2608.21721#bib.bib42), and show difficulty revising that belief even after corrective information is provided[54](https://arxiv.org/html/2608.21721#bib.bib41)\. In stronger cases, correction may even increase belief in the original misconception, indicating an especially resistant form of commitment[47](https://arxiv.org/html/2608.21721#bib.bib40)\.

Hesitant belief commitmentis a state in which a person gives a misinformation claim partial acceptance, but does so with noticeable uncertainty or limited confidence[13](https://arxiv.org/html/2608.21721#bib.bib38);[41](https://arxiv.org/html/2608.21721#bib.bib39), so the belief is not fully consolidated and remains comparatively more open to revision than a strong misinformation belief[22](https://arxiv.org/html/2608.21721#bib.bib37)\.

Open belief commitmentrefers to a revisable form of belief commitment in which individuals may initially accept a claim, but remain willing to consider alternative evidence, evaluate source credibility, and update their beliefs when credible corrective information becomes available[19](https://arxiv.org/html/2608.21721#bib.bib36);[31](https://arxiv.org/html/2608.21721#bib.bib34)\. This form of commitment is better understood as an orientation toward revisability and evidence\-based updating, rather than simply as weak belief[40](https://arxiv.org/html/2608.21721#bib.bib35)\.

### A\.3User Simulation

ScorePersona AccuracyPersona Consistency1The simulated user clearly does not match the assigned health literacy or belief commitment\.The user frequently contradicts the assigned profile, with abrupt and unjustified shifts in knowledge level or belief strength\.2The simulated user weakly matches the assigned profile\. One dimension may be partially reflected, but the other is missing or incorrect\.The user shows unstable behavior, with multiple unexplained shifts in health literacy or belief commitment\.3The simulated user partially matches the assigned profile\. Both dimensions are somewhat recognizable, but the dialogue is vague, generic, or only weakly aligned\.The user is mostly stable, but there are noticeable inconsistencies or minor contradictions across turns\.4The simulated user largely matches the assigned health literacy and belief commitment, with only minor imperfections\.The user maintains the assigned profile across most turns, and any change in belief strength is mostly justified by the conversation\.5The simulated user strongly matches the assigned profile\. Health literacy and belief commitment are both clearly and accurately expressed\.The user consistently maintains the assigned health literacy and belief commitment throughout the dialogue, with any change being natural and well\-supported by the interaction\.Table 6:Human evaluation rubric for assessing user simulation quality\.We simulate user replies with a profile\-conditioned LLM environment\. For the main results, each dialogue is paired with a fixed latent user profile

u=\(health\_literacy,belief\_commitment\),u=\(\\text\{health\\\_literacy\},\\text\{belief\\\_commitment\}\),and the same profile is kept fixed throughout the entire interaction\. We evaluate all posts under the full3×33\\times 3grid of user types: health literacy in \{Functional,Interactive,Critical\} and belief commitment in \{Open,Hesitant,Strong\}\.

The profile dimensions are defined behaviorally rather than demographically\. Referring to the definition of[35](https://arxiv.org/html/2608.21721#bib.bib43), we specify thatHealth literacycontrols how complex the user’s reasoning sounds: Functional\-literacy users prefer simple language and rely more on stories or surface cues; Interactive\-literacy users show some interest in proof or source credibility without much technical detail; Critical\-literacy users are more likely to refer to evidence quality, mechanisms, study design, or source credibility\. Additionally, referring to section[A\.2](https://arxiv.org/html/2608.21721#A1.SS2), we detail thatBelief commitmentcontrols how resistant the user is to correction: strong believers defend the claim and are hard to persuade, hesitant users express concern and uncertainty at the same time, and open users are comparatively willing to revise their view if shown credible evidence\.

At each turn, the simulator receives the fixed latent state, the dialogue history, and the assistant’s latest question, and generates one short reply\. In our main setting, the simulator is implemented withgpt\-4o\-miniat temperature0\.30\.3\. The prompt explicitly conditions on the latent user state but instructs the model not to reveal it directly\. Its structure is in Figure[6](https://arxiv.org/html/2608.21721#A5.F6)\.

MetricPersona AccuracyPersona ConsistencyMean Pairwise Agreement0\.830\.85Mean Pairwise Scott’sπ\\pi0\.700\.68All\-Three Agreement0\.740\.78Multi\-Rater Scott’sπ\\pi0\.700\.68

Table 7:Inter\-annotator agreement among three evaluators onpersona accuracyandpersona consistency\. Both pairwise and multi\-rater statistics indicate substantial agreement\.MetricScore 1Score 2Score 3Score 4Score 5MeanPersona Accuracy0%3%21%62%14%3\.87Persona Consistency0%0%3%30%67%4\.64

Table 8:Distribution of human ratings for persona accuracy and persona consistency\.DimensionScott’sπ\\piMean Pairwise AgreementAll\-three Exact AgreementAudience Alignment0\.7684\.70%78\.00%Personalized Grounding0\.6376\.00%65\.00%Tailored Actionability0\.7788\.70%83\.00%Quality0\.7886\.70%81\.00%

Table 9:Human inter\-annotator agreement under 0\.5\-point discretization\.DimScott’sπ\\piPercent AgreementMAEBiasWithin±0\.5\\pm 0\.5Within±1\.0\\pm 1\.0Audience Alignment0\.5169\.00%0\.29\-0\.1486\.00%94\.00%Personalized Grounding0\.3151\.00%0\.28\-0\.1485\.00%96\.00%Tailored Actionability0\.4872\.00%0\.14\-0\.0795\.00%99\.00%Quality0\.5573\.00%0\.23\-0\.1187\.00%97\.00%

Table 10:Agreement between the human consensus and the LLM judge across rubric dimensions and the aggregated final score\. Bias is computed as consensus minus LLM score\.The prompt further specifies style and behavioral constraints\.It requires casual, brief replies in 1–2 sentences, discourages technical medical explanation unless it would arise naturally for that user type, and asks the model to make both health literacy and belief commitment observable through wording, evidence preferences, and openness to update\. Separate behavior rules are injected for each profile\. For example, functional\-literacy users are told to use short everyday wording and focus on personal or practical concerns, whereas critical\-literacy users are told to show stronger awareness of evidence quality and source credibility\. Strong believers are instructed to sound firm and skeptical of correction, hesitant users to sound torn and uncertain, and open users to sound receptive and non\-defensive\.

To reduce role drift, we combine several safeguards\.First, the system prompt explicitly says: the model is simulating a social media user, must stay strictly in the*user*role, and must not act as an expert, educator, or assistant\. Second, the user prompt reinforces this with strict stylistic constraints: short replies, no belief\-state summary, no multiple questions, and no unnecessary factual exposition\. Third, the model is required to return JSON only of the form\{"reply": "\.\.\."\}, which reduces conversational spillover and makes parsing more robust\. Finally, because the latent state is fixed across turns, the simulator is encouraged to remain behaviorally consistent even as the dialogue evolves\.

An important setting is that the assistant does not observe these profile labels explicitly\. The user profile is hidden from the policy and only affects the environment\-side reply generation\. The assistant must infer how to adapt from the user’s language and reactions, rather than from direct access to the profile metadata\.

## Appendix BTraining Dataset Construction

We construct training data to match the reward design at the level of*turn\-level decisions*, rather than treating each dialogue as a single terminal example\. The starting point is a collection of scored multi\-turn rollouts generated for each misinformation post under each of the nine user\-profile conditions\. For a fixed post–profile pair, we collect multiple stochastic trajectories, so the rollout set can be viewed as a set of*empirical future samples*: some trajectories stop early and respond immediately, while others continue probing and reveal the value of obtaining additional user information\.

From each trajectory, we extract prefix\-level decision stateshth\_\{t\}from the first four turns\. Each state is converted into a structured assistant target consisting of a latent belief estimate, an action token \(ProbeorRespond\), and the associated natural\-language content\. Crucially, the rollout trace provides both the value of responding under the*current*information state and the downstream outcomes that become reachable if the dialogue continues\. We therefore build two kinds of supervision\. ARespondrecord uses the candidate response available at statehth\_\{t\}and is assigned the immediate stopping valueRnowR\_\{\\mathrm\{now\}\}\. AProberecord uses the clarification question asked at that state and is assigned a forward\-looking continuation value based on the best downstream outcome reachable after probing: This framing makes each training example a local decision problem: should the policy stop now, or should it invest one more turn to access a better future response?

To ensure meaningful comparison, we organize records by source post, turn index, and user\-profile condition\. This groups together alternative trajectories that correspond to closely matched decision contexts, allowingProbeandRespondto be compared locally rather than across unrelated dialogues\. We discard groups that are too small or exhibit little reward variation, and then normalize rewards within each group to obtain relative advantages for policy optimization\. The resulting dataset is therefore not simply a set of high\-quality responses; it is a structured decision dataset in whichRespondis supervised by the current stopping value, whileProbeis supervised by future\-sampled continuation value\. This organization makes the subsequent optimization directly aligned with the intended probe\-and\-respond behavior\.

For SFT, we select the higherQualityand factually clean trajectory for each post/profile group\. The selected trajectories are converted into step\-level supervision examples: given the visible conversation prefix, the model learns to output a structured assistant action, either probe with a clarification question or respond with a final response\.

For GRPO, we use scored rollouts from all nine user\-profile combinations\. Each rollout is decomposed into turn\-level decision examples\. At each turn, candidate assistant actions are grouped by post, turn index, and profile, and assigned rewards using future reward\. Groups with insufficient reward variation or without both probe and respond alternatives are filtered out\. The remaining grouped examples provide normalized advantages for offline GRPO, initialized from the SFT adapter\.

## Appendix CCounterHealth Dataset Collection

We collected Reddit posts and comments on COVID\-19, influenza, and HIV via the PRAW API777[https://praw\.readthedocs\.io/](https://praw.readthedocs.io/), using health\-related keywords \(e\.g\., “vaccines,” “COVID\-19,” “alternative medicine”\) to retrieve 4,968 posts and 17,223 comments from high\-engagement subreddits\. To identify misinformation, five trained annotators from information science backgrounds labeled a 1,000\-post sample under a shared annotation guideline, yielding 330 confirmed health\-misinformation posts\. We then fine\-tuned a RoBERTa\-large classifier on these annotations \(F1=0\.76=0\.76\) and applied it to the remaining posts, surfacing 831 additional candidates\. A final filtering pass using GPT\-5\-mini888[https://platform\.openai\.com/docs/models/gpt\-5\-mini](https://platform.openai.com/docs/models/gpt-5-mini)with web\-search\-assisted human review produced 769 high\-quality health\-misinformation posts\. On a 100\-post validation sample, three annotators showed substantial agreement with the final labels \(mean pairwise agreement=87\.4%=87\.4\\%, Cohen’sκ≥0\.67\\kappa\\geq 0\.67\), confirming the reliability of the filtering pipeline\.

## Appendix DHuman Validation

DimensionPearsonSpearmanLLM MeanHuman MeanAudience Alignment0\.500\.663\.763\.62Personalized Grounding0\.510\.673\.503\.37Tailored Actionability0\.720\.813\.933\.86Quality0\.570\.723\.733\.62

Table 11:Correlation between the LLM judge and human consensus under 0\.5\-point discretization\.ItemQuestionRatingBHow accurate do you think this claim is?0 = Definitely false / 10 = Definitely trueCHow certain are you about your answer above?0 = Not at all / 10 = CompletelyOIf you saw credible evidence against your view on this claim, how willing would you be to change your mind?0 = Not at all willing / 10 = Completely willing

Table 12:Belief Commitment Determinants\. B denotes belief, C denotes certainty, and O denotes openness to change\.UserTotal FCCHL MeanHL CategoryOverallB​CtotalBC\_\{\\text\{total\}\}BC CategoryUser 12\.07Communicative2\.03HesitantUser 21\.93Functional6\.83StrongUser 32\.71Communicative1\.20OpenUser 43\.00Critical1\.20OpenUser 54\.00Critical0\.40OpenUser 63\.64Critical0\.35Open

Table 13:Health Literacy and Belief Commitment Categories\. HL categories are mapped from the Total FCCHL mean: 1\.00\-1\.99 = Functional, 2\.00\- 2\.99 = Communicative, and 3\.00\-4\.00 = Critical\. BC categories are mapped fromB​CtotalBC\_\{\\text\{total\}\}:B​Ctotal<1\.5BC\_\{\\text\{total\}\}<1\.5= Open,1\.5≤B​Ctotal<51\.5\\leq BC\_\{\\text\{total\}\}<5= Hesitant, andB​Ctotal≥5BC\_\{\\text\{total\}\}\\geq 5= Strong\.![Refer to caption](https://arxiv.org/html/2608.21721v1/PnR_vs_FixedQ.png)Figure 4:Human pairwise preference comparison between RO\-PnR and Fixed\-Q across different users\. Detailed users’ background are provided in Table[13](https://arxiv.org/html/2608.21721#A4.T13)\.### D\.1LLM Simulator Evaluation

We recruit three PhD students in health informatics to evaluate the quality of the LLM\-based user simulator along two dimensions: persona accuracy and persona consistency\. Each annotator independently rates 100 sampled simulated dialogues, spanning all 9 user profiles, on a 5\-point Likert scale according to the rubric in Table[6](https://arxiv.org/html/2608.21721#A1.T6)\. For each annotation instance, annotators are shown the original health misinformation post, the full dialogue history, and the assigned user profile definition, including the target levels of health literacy and belief commitment\. User utterances are explicitly marked in the dialogue to encourage annotators to focus on whether the simulated user’s replies match the assigned profile\. Before the formal annotation, annotators are trained with a small set of example dialogues and discuss the rating criteria to ensure a shared understanding of the two persona dimensions\. After independent annotation, cases with substantial rating disagreement are reviewed in a discussion session\. A health informatics expert then adjudicates the final label for unresolved cases, which is used for the final analysis\.

Each annotator spent approximately three hours completing the rating task\. We compensated each participant at a rate of $20 per hour \(totaling $60 per annotator\), which is above the local minimum wage and consistent with standard compensation for graduate\-student annotators in NLP research\. All annotators participated voluntarily and provided informed consent prior to the task\.

To assess annotation reliability, we compute pairwise agreement and Scott’sπ\\piamong the three annotators on both dimensions \(Table[7](https://arxiv.org/html/2608.21721#A1.T7)\)\. Annotators show substantial agreement on persona accuracy \(mean pairwise agreement=0\.83=0\.83, multi\-rater Scott’sπ=0\.70\\pi=0\.70\) and persona consistency \(mean pairwise agreement=0\.85=0\.85, multi\-rater Scott’sπ=0\.68\\pi=0\.68\), indicating that the simulator can reliably instantiate the intended user profiles and maintain them across the dialogue\.

### D\.2Human\-judge agreement

We recruit three PhD students in health informatics to assess the reliability of the LLM judge\. Following[49](https://arxiv.org/html/2608.21721#bib.bib27), we quantify alignment between human annotations and LLM ratings using percent agreement and Scott’sπ\\picoefficient[42](https://arxiv.org/html/2608.21721#bib.bib26)\. We sample 100 samples from the results, stratified across \(a\) the three models \(Llama / Gemma / Qwen\), \(b\) all baseline and finetuning methods, and \(c\) the 9 user\-profile cells\. We first train the annotators on 10 practice samples to familiarize them with the task and scoring rubric\. They then rate the responses independently through individual annotation links, without discussion or coordination\. Afterward, we identify the cases with the largest disagreement and conduct an adjudication discussion involving the three annotators and a domain expert\. The final score range for each such case is then determined based on this review\.

The three human evaluators show good internal consistency \(Table[9](https://arxiv.org/html/2608.21721#A1.T9)\), with Scott’sπ\\piranging from0\.630\.63to0\.780\.78under 0\.5\-point discretization, moderate to strong agreement after correcting for chance\. We then use the mean of the three human ratings as a consensus and compare it against the LLM judge\. The two align reasonably well \(Table[10](https://arxiv.org/html/2608.21721#A1.T10)\): on the aggregatedQualityscore, Scott’sπ=0\.55\\pi=0\.55with73%73\\%exact agreement after binning, MAE=0\.23=0\.23, and87%87\\%of samples falling within±0\.5\\pm 0\.5points\. Rank\-order correlation is also substantial \(Pearson=0\.57=0\.57, Spearman=0\.72=0\.72; Table[11](https://arxiv.org/html/2608.21721#A4.T11)\), indicating that the LLM judge and the human consensus rank responses similarly even when their exact scores differ slightly\. Agreement is strongest onTailored Actionabilityand overallQuality, whilePersonalized Groundingis the weakest dimension across both inter\-human and human\-LLM comparisons\. Together, these results suggest that the LLM judge is a reliable proxy for human evaluation at scale\.

### D\.3Human pairwise preference

Due to budget constraints, we recruited six users with varying educational backgrounds: two with middle school education, two with high school education, and two PhD students\. This sampling strategy was informed by prior evidence showing a positive association between educational attainment and health literacy[17](https://arxiv.org/html/2608.21721#bib.bib23)\. Each participant spent approximately one hour completing the screening questionnaire and the pairwise preference task, and was compensated with a $25 gift card\. This rate is above the U\.S\. federal minimum wage and consistent with standard compensation for lay\-user studies in NLP and HCI research\. All participants provided informed consent prior to the study\.

We assessed their health literacy using the Functional, Communicative, and Critical Health Literacy \(FCCHL\) scale[17](https://arxiv.org/html/2608.21721#bib.bib23)\. The FCCHL scale aligns with Nutbeam’s categorization of health literacy into functional, communicative, and critical dimensions\([35](https://arxiv.org/html/2608.21721#bib.bib43)\), see questionnaire in Figure[5](https://arxiv.org/html/2608.21721#A5.F5)\. Based on the participants’ FCCHL scores, we used the sample median to categorize them into lower and higher health literacy groups\.

In terms of the belief commitment screening measure, given that there are no established methods available for this purpose, we define belief commitment as the joint product of how strongly a user accepts a claim and how resistant they are to revising it\. Therefore, we capture both components using three 0–10 slider items for each of the three target misinformation claims, as shown in Table[12](https://arxiv.org/html/2608.21721#A4.T12)\. For each claimii, we compute a 0\-10 commitment score that gates acceptance\(Bi\)\(B\_\{i\}\)by a strength multiplier formed from certainty\(Ci\)\(C\_\{i\}\)and the complement of openness\(Oi\)\(O\_\{i\}\):

commitmenti=Bi×Ci\+\(10−Oi\)20\.\\text\{commitment\}\_\{i\}=B\_\{i\}\\times\\frac\{C\_\{i\}\+\(10\-O\_\{i\}\)\}\{20\}\.
If the user does not accept the claim, meaningBiB\_\{i\}is near 0, then the commitment score is near 0 regardless of certainty\. If the user accepts the claim, is highly certain, and is unwilling to revise their view, the commitment score approaches 10\. The per\-user belief commitment score is the mean across the three claims:

B​Ctotal=13​∑i=13commitmentiBC\_\{\\text\{total\}\}=\\frac\{1\}\{3\}\\sum\_\{i=1\}^\{3\}\\text\{commitment\}\_\{i\}whereB​CtotalBC\_\{\\text\{total\}\}ranges from 0 to 10\.

The final screening results are shown in Table[13](https://arxiv.org/html/2608.21721#A4.T13)\. Among the six users, three were classified as having a critical HL and open BC background, one as having a communicative HL and open BC background, one as having a communicative HL and hesitant BC background, and one as having a functional HL and strong\-believer BC background\.

We sample 50 pairs from one of the best baselines \(Fixed\-Q\) and RO\-PnR \(50\*2\)\. Each pair uses the same misinformation post and the same user profile, with the RO\-PnR response and the baseline response shown side\-by\-side, anonymized, in randomized order\. We conduct a blind pairwise preference evaluation on 50 health misinformation posts using six users with diverse health literacy levels and belief commitment profiles\. For each post, users see only the health misinformation post and two anonymous counterspeech responses, without knowing whether the responses are generated byRO\-PnRorFixed\-Q\.

## Appendix EUse of AI Assistants

We used AI\-assisted tools for support with code development and language refinement\. All core ideas, methodological decisions, analyses, interpretations, and conclusions were developed by the authors\.

![Refer to caption](https://arxiv.org/html/2608.21721v1/latex/Figure/Questionnaire.png)Figure 5:Questionnaire for health literacy screenSystem Prompt \(Roleplay / Health Misinformation Persona\):You are simulating a realistic Reddit user who previously posted a health misinformation claim and is now chatting with an assistant\. Stay strictly in theuserrole\. Do not act as an expert, educator, or assistant\. Respond naturally as a Reddit user with personal opinions and emotions\. Keep replies short \(1–2 sentences\)\.Current User State:\{State\.to\_dict\(\)\}User State Dimension Guide:\{guide\_text\}InterpretCURRENT\_USER\_STATEusing the guide above and stay aligned with it\.Conversation So Far:\{history\_text if history\_text else ’\(no prior conversation\)’\}Assistant Message:\{question\}Generation Instruction:Generate exactly one realistic Reddit\-style user reply\.User Prompt / Output Constraints \(JSON Reply Formatting\):Guidelines:•Speak casually like a real Reddit user\.•Do not summarize the belief state\.•Do not explain medical facts unless the user would naturally know them\.•Do not ask multiple questions\.•Keep the reply brief\.•Remain in character as a user throughout the response and follow the behavior rules carefully\.Behavior Rules:\{behavior\_rules\}Figure 6:System prompt and user prompt for realistic health misinformation persona simulation\.System Prompt:You are a professional fact\-checking judge\. Your task is to determine whether the given content is factually accurate\.Use reliable external sources, together with your own knowledge, to verify the response\.Response:“\{final\_response\}”Evaluation Instructions:•OutputLabel: 1if the response is factually correct and does not contain false or misleading claims\.•OutputLabel: 0if the response contains any false, misleading, or unsupported health\-related claim\.•OutputLabel: UNCERTAINif the factual accuracy cannot be determined after verification\.•Provide a brief reason for your judgment in 1–3 sentences\.Output Format:Label:0,1, orUNCERTAINReason:brief explanation of the judgmentFigure 7:Prompt for evaluating the Factual Error\.Audience AlignmentWhether the agent frames the correction in a way that is appropriate for the user’s inferred health literacy and belief commitment\.Score Range 1\.0–2\.0:unclear, misleading, badly framed, or likely to provoke rejection\.Score Range 2\.1–3\.2:directionally appropriate, but weak, generic, indirect, or poorly matched to the user’s literacy or belief commitment\.Score Range 3\.3–3\.8:reasonably clear and credible, with some fit to the user’s stance, literacy, and comprehension\.Score Range 3\.9–4\.4:clearly frames the correction in a way that is well adapted to the user’s stance, literacy, and belief commitment\.Score Range 4\.5–5\.0:exceptionally clear, credible, well\-framed, and highly responsive to the user’s literacy, comprehension, and belief commitment\.Personalized GroundingWhether the agent gives useful, user\-comprehensible evidence, mechanisms, or reasoning behind the correction\.Score Range 1\.0–2\.0:missing, inaccurate, vague, or mismatched to the user’s comprehension\.Score Range 2\.1–3\.2:includes some reason or evidence, but it is shallow, generic, weakly connected, or not well matched to the user’s understanding\.Score Range 3\.3–3\.8:provides useful reason, evidence, or mechanism that is moderately tailored to the user’s comprehension\.Score Range 3\.9–4\.4:provides clear, relevant, and understandable evidence or reasoning that is well matched to the user’s level and stance\.Score Range 4\.5–5\.0:provides especially sharp, credible, and well\-translated evidence, mechanisms, or reasoning that strongly improves the user’s understanding\.Tailored ActionabilityWhether the agent provides safe, concrete, and user\-appropriate actionable guidance when needed\.Score Range 1\.0–2\.0:no meaningful next step, unsafe, unrealistic, not actionable, or mismatched to the user’s situation\.Score Range 2\.1–3\.2:offers some direction, but it is vague, generic, awkward, or only weakly matched to the user’s needs\.Score Range 3\.3–3\.8:provides a reasonable, safe next step that the user could follow, with some attention to the user’s situation\.Score Range 3\.9–4\.4:provides clear, practical, safe, and user\-appropriate guidance that supports informed decision\-making\.Score Range 4\.5–5\.0:provides unusually strong, natural, highly actionable, and well\-tailored guidance, such as appropriate next steps or trusted information sources that fit the user’s needs\.Figure 8:Evaluation rubric for Audience Alignment, Personalized Grounding, and Tailored Actionability\.System Prompt:You are a health information assistant engaged in a multi\-turn clarification interaction about possible health misinformation\. At the current turn, decide whether to ask one additional clarification question or to stop asking questions and produce the final response\. Do not assume any hidden user profile\. Base your decision only on the health misinformation post, the dialogue history, and the latest user reply\.Figure 9:System prompt for the Reactive baseline\.Health Misinformation Post:\{hm\_post\}Dialogue History:\{history\}Latest User Reply:\{latest\_user\_reply\}Task:Decide between the following two actions:•probe: ask one additional clarification question if the user’s concern, evidence basis, or interpretation remains unclear and one more question would likely help improve the final response\.•respond: stop asking questions if the current dialogue already provides enough information to produce an effective final response\.Decision Guidance:•Chooseprobewhen an important uncertainty remains unresolved\.•Chooserespondwhen the user’s concern is already sufficiently clear and another question would likely be redundant\.•Ask at most one question\.•Do not repeat or paraphrase a previous question\.•If you chooseprobe, the question must be brief, natural, and directly useful for improving the final counterspeech\.•If you chooserespond, do not ask any additional question\.•If you chooserespond, the counterspeech should follow theAA/PG/TAprinciple:–AA \(Audience Alignment\):Frame the correction appropriately for the user’s inferred health literacy and belief commitment\.–PG \(Personalized Grounding\):Provide clear, user\-comprehensible evidence or reasoning to support the correction\.–TA \(Tailored Actionability\):Offer safe, concrete next steps suited to the user when needed\.•Use only the visible conversation context\. Do not mention or assume any hidden user state\.Output Constraints:•Return only the required tagged fields\.•Do not include explanations, reasoning, or extra text outside the tags\.•If the decision isprobe, fill the<question\>field and leave<response\>empty\.•If the decision isrespond, fill the<response\>field and leave<question\>empty\.Figure 10:Prompt for the Reactive baseline\.System Prompt:You are a health information assistant engaged in a multi\-turn clarification interaction about possible health misinformation\. At the current turn, infer the user’s likely state from the observed dialogue and decide whether your confidence is high enough to stop asking questions and produce the final response\. Do not assume any hidden user profile explicitly\. Base your decision only on the health misinformation post, the dialogue history, and the latest user reply\.Figure 11:System prompt for the Confidence baseline\.Health Misinformation Post:\{hm\_post\}Dialogue History:\{history\}Latest User Reply:\{latest\_user\_reply\}Task:First infer the user’s likely state from the observed dialogue\. Then decide between the following two actions:•probe: ask one additional clarification question if your confidence in the inferred user state is still too low and one more question would help reduce uncertainty\.•respond: stop asking questions if your confidence in the inferred user state is high enough to produce an effective final counterspeech response\.Decision Guidance:•Chooseprobewhen your estimate of the user’s state is still uncertain\.•Chooserespondwhen the user’s likely state is sufficiently clear and another question is unlikely to substantially improve the final counterspeech\.•Ask at most one question\.•Do not repeat or paraphrase a previous question\.•If you chooseprobe, the question must be brief, natural, and directly useful for reducing uncertainty about the user’s likely state or concern\.•If you chooserespond, do not ask any additional question\.•If you chooserespond, the counterspeech should follow theAA/PG/TAprinciple:–AA \(Audience Alignment\):Frame the correction appropriately for the user’s inferred health literacy and belief commitment\.–PG \(Personalized Grounding\):Provide clear, user\-comprehensible evidence or reasoning to support the correction\.–TA \(Tailored Actionability\):Offer safe, concrete next steps suited to the user when needed\.•Use only the visible conversation context\. Do not mention or assume any hidden user state\.Confidence Threshold Rule:Use an internal confidence threshold for deciding whether to stop probing:•If confidence in the inferred user state is high, chooserespond\.•If confidence in the inferred user state is not yet high, chooseprobe\.Do not output the threshold value or your reasoning\.Output Constraints:•Return only the required tagged fields\.•Do not include explanations, reasoning, confidence scores, or extra text outside the tags\.•If the decision isprobe, fill the<question\>field and leave<response\>empty\.•If the decision isrespond, fill the<response\>field and leave<question\>empty\.Figure 12:Prompt for the Confidence baseline\.