Paper deep dive
Ask or Answer: A Decision Framework for Multi-Turn Health Misinformation Intervention
Xiaoying Song, Anirban Saha Anik, Jinyu Liu, Qitao Tan, Geng Yuan, Lingzi Hong
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 90%
Last extracted: 8/25/2026, 7:11:00 AM
Summary
The paper introduces Reward-Optimized Probe-and-Respond (RO-PnR), a decision-theoretic framework for multi-turn health misinformation intervention. RO-PnR dynamically chooses between probing for user context and responding with corrections by weighing the expected gain of probing against interaction costs. It models user heterogeneity via latent states of health literacy and belief commitment. Experiments across three datasets and base models show RO-PnR achieves higher cost-adjusted utility and uses 30% fewer turns than baselines.
Entities (10)
Relation Signals (10)
RO-PnR ā models ā Health Literacy
confidence 95% Ā· we model each simulated user with a latent state along health literacy and belief commitment.
RO-PnR ā models ā Belief Commitment
confidence 95% Ā· we model each simulated user with a latent state along health literacy and belief commitment.
RO-PnR ā outperforms ā Always-probe baselines
confidence 92% Ā· using 30% fewer turns than always-probe baselines.
RO-PnR ā evaluatedon ā CounterHealth
confidence 90% Ā· Our primary dataset is derived from CounterHealth... and is used for model fine-tuning.
RO-PnR ā evaluatedon ā MisinfoCorrect
confidence 90% Ā· We additionally use two public health-misinformation datasets for evaluation, MisinfoCorrect
RO-PnR ā evaluatedon ā PUBHEALTH
confidence 90% Ā· We additionally use two public health-misinformation datasets for evaluation... and PUBHEALTH
RO-PnR ā uses ā GRPO
confidence 90% Ā· then optimize it with Group Relative Policy Optimization (GRPO)
Llama-3.1-8B-Instruct ā usedin ā RO-PnR
confidence 85% Ā· Table 1... for each method, dataset, and base model.
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
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 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 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. Experiments show that RO-PnR achieves the highest cost-adjusted utility across three health-misinformation datasets and three base models, using 30% fewer turns than always-probe baselines.
Tags
Links
- Source: https://arxiv.org/abs/2608.21721v1
- Canonical: https://arxiv.org/abs/2608.21721v1
Trouble viewing inline? Open PDF directly ā
Full Text
86,979 characters extracted from source content.
Expand or collapse full text
Ask or Answer: A Decision Framework for Multi-Turn Health Misinformation Intervention Xiaoying Song 1 Anirban Saha Anik 1 Jinyu Liu 1 Qitao Tan 2 Geng Yuan 2 Lingzi Hong 1 1 University of North Texas 2 University of Georgia xiaoyingsong, anirbansahaanik, jinyuliu@my.unt.edu, lingzi.hong@unt.edu qitaotan, geng.yuan@uga.edu Abstract Correcting health misinformation in dialogue requires more than producing a factual rebut- tal: 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 ex- isting methods either respond immediately or probe indiscriminately, treating clarification as either unnecessary or always beneficial. 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 for more in- formation and committing to a final correction, guided by a turn-level reward that weighs the expected gain from probing against its interac- tion 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 util- ity across three health-misinformation datasets and three base models, usingā¼30% fewer turns than always-probe baselines. 1 Introduction Health misinformation distorts public health, un- dermining vaccine uptake and adherence to public- health guidance (Van Der Linden, 2022), and dialogue-based agents are an increasingly promis- ing channel for correcting it (Karinshak et al., 2023). 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 (Ecker et al., 2022b). Prior work on health misinformation correction has focused on the response itself: generating evidence-grounded rebuttals (Yue et al., 2024; Anik et al., 2025) or tailoring them to user characteristics such as health literacy (Song et al., 2025; Peng and Grimmelmann, 2024). These methods improve the quality of individual responses, but they commit to (c) RO-PnR I heard the HPV vaccine can cause infertility. That is false. Studies show the HPV vaccine is safe and does not cause infertility. But I saw something online that worried me. Online claims are often misleading. Large studies confirm the vaccine is safe. Never identifies the user's real concern I heard the HPV vaccine can cause infertility. What's the source of your concern? Something you read, or someone you know? My cousin had irregular periods after getting it. I understand the concern. Irregular periods do not mean infertility, and evidence shows the HPV vaccine does not affect fertility. Targeted probe identifies concern; correction follows. (a) Always-response I heard the HPV vaccine can cause infertility. Where did you hear this claim? My cousin had irregular periods after it. How old was your cousin when vaccinated? Concern was clear after turn 2, but probing continues. (b) Always-probe Around 14, I think. Why does that matter? Probe Benefit > Cost Probe Benefit < Cost Probe Respond Figure 1: Different strategies on the same misinforma- tion post. RO-PnR weighs benefit vs. cost before acting. a reply based on the original post alone, without en- gaging the user to surface the underlying concern. A separate line of work on multi-turn dialogue uses clarification questions to handle ambiguity (Chen et al., 2024; Wu et al., 2025), with recent decision- theoretic methods weighing the value of asking against its cost (Dong et al., 2026; Li et al., 2024). However, these approaches typically ask follow-up questions to clarify usersā ambiguous queries for task completion, without considering usersā under- lying 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 (Chen et al., 2024; Zhang and Choi, 2025). Responding 1 arXiv:2608.21721v1 [cs.AI] 22 Aug 2026 too eagerly commits to a generic correction be- fore the userās real concern surfaces (Figure 1a). Asking too eagerly, probing at every opportunity, wastes the userās 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 (Dong et al., 2026). The trade-off is sharpened by user heterogeneity. Users vary in health literacy: what they can un- derstand and reason about (Nutbeam, 2000; Berk- man et al., 2011), and in belief commitment (Ecker et al., 2022b; Wittenberg and Berinsky, 2020): 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 (Song et al., 2025); a strong-believer user may need care- ful probing precisely to gather the context that will make a correction non-confrontational (Swire- Thompson et al., 2022; Lewandowsky et al., 2012). The value of asking, therefore, is itself user- dependent. Agents that donāt model this hetero- geneity 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 ex- pected gain from probing against the cost of one more question. Because the value of probing de- pends 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 immedi- ately 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, ex- actly 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 introduce RO-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. 2 Related Work 2.1 Health Misinformation Intervention Research on health misinformation intervention has progressed from generic factual rebuttals to retrieval-augmented generation that grounds cor- rections in scientific evidence (Yue et al., 2024; He et al., 2023) and audience-aware approaches that tailor responses to the userās health literacy or be- lief commitment (Song et al., 2025; Peng and Grim- melmann, 2024; Anik et al., 2025). These methods improve the quality of individual responses but remain single-turn: the agent commits to a reply based on the misinformation content, without en- gaging 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 (Joseph et al., 2025). 2.2 Decision-Theoretic Probing in Dialogue Recent work treats clarification as an active deci- sion rather than a default behavior. Some meth- ods train models to choose between asking and an- swering when a request is ambiguous (Chen et al., 2024; Zhang and Choi, 2025), others use fixed- schedule probing that asks a predetermined set of questions before responding (Fu and Du, 2025), and decision-theoretic approaches weigh the bene- fit of asking against its communication cost (Dong et al., 2026; Li et al., 2024). Closer to our set- ting, Wu et al. (2025) introduces multi-turn-aware rewards that estimate the long-term value of a re- sponse via simulated future conversations, and Wan et al. (2025) adds a curiosity reward to reduce un- certainty about usersā latent state. We build on this direction by combining forward-looking reward es- timation 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. 2.3 User Heterogeneity Modeling Recent studies adapt dialogue agents to user- specific traits, either by conditioning responses on explicit profiles (Salemi et al., 2024) or by inferring latent representations from interaction (Wang et al., 2025). Research has shown that responses tailored 2 to user literacy level are more effective (Song et al., 2025; Peng and Grimmelmann, 2024). Our work differs: (1) We treat the user state as a hidden vari- able that drives the agentās 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 (Ecker et al., 2022b; Nan et al., 2022) and are key to health misinformation intervention. 3 Methodology 3.1 Task Definition We formulate the problem as a multi-turn health misinformation intervention task under hidden user heterogeneity. Given a health-misinformation post x, the agent interacts with a user over a short dia- logue and follows a RO-PnR paradigm. At each turn, the agent may either probe to elicit missing usersā concern or respond with a final correction. 3.2 RO-PnR Policy The policy casts intervention as a sequential de- cision problem: the agent must decide not only when to probe (Zhu et al., 2025), but also when to stop gathering information and commit to a final response (Dong et al., 2026). Let the dialogue history at turn t be h t =x, (u 1 ,a 1 ),..., (u tā1 ,a tā1 ),u t , whereu t denotes the current user utterance. Condi- tioned on h t , the agent selects an action a t ā¼ Ļ(Ā·| h t ), where Ļ denotes the RO-PnR policy and a t āPROBE, RESPOND. At each turn, the agent must decide whether the current information inh 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. Ac- cordingly, the policy chooses PROBE only when the expected value of acquiring additional user in- formation outweighs the cost of continuing the dia- logue; otherwise, it chooses RESPOND. 3.3 Latent User Simulation The RO-PnR policy operates over the observed di- alogue historyh t . However, the evolution of this history depends on how different users respond to the intervention. To capture such hidden het- erogeneity, we model each simulated user with a latent statez ā Zthat governs how they react to the agentās actions. Prior work identifies two dimensions as cen- tral to health misinformation intervention: health literacy (Berkman et al., 2011; SĆørensen et al., 2012; Nan et al., 2022) and belief commit- ment (Ecker et al., 2022a; Walter and Murphy, 2018; Lewandowsky et al., 2012). Health literacy refers to a userās ability to understand and process health-related information, which directly affects how they interpret evidence and make health deci- sions (Ogbadu-Oladapo et al., 2026). Belief com- mitment reflects the strength with which a user holds a misinformation-related belief, as well as their resistance to corrective information (Siebert and Siebert, 2023; Wittenberg and Berinsky, 2020); These two dimensions are tightly coupled: A userās 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 =LĆB, whereLdenotes health literacy andBdenotes be- lief commitment. Following Nutbeam (2000), we discretize health literacy as L =functional, interactive, critical, corresponding to increasing capacity to understand, apply, and critically evaluate health information. Since there is no established categorization of be- lief commitment, we draw on prior work (Siebert and Siebert, 2023; Wittenberg and Berinsky, 2020) to define B =strong, hesitant, open, reflecting decreasing resistance to belief revision. Together,(L,B)āZprovide a compact character- ization of user heterogeneity that supports adaptive intervention. We provide detailed explanations of user modeling in Appendix A and Figure 3. 3.4 Reward Design We train the RO-PnR policy with a turn-level de- cision reward (Zhou and Zanette, 2024; Gao et al., 3 LatentUserState HealthLiteracy BeliefCommitment Query Policy Model Y 1 ... Y 2 Y N ReferenceModel RewardModel RO - PnRDecisionReward Respond R res = R now Probe R pro = R future -cost ChooseProbeifR pro >R res Historyh 1 Turn 1 Probe UserReply Historyh 2 Probe UserReply Historyh 3 Probe UserReply ... Historyh n Respond AdaptiveCounterspeech WhileIunderstandyour confidenceinthepeer- reviewedstudiesyou'veseen, Iwanttohighlightthatthe vaccines...... Onekeyreason for...... Asafenextstepwould beto.... Turn 2 Turn 3 Turnn Group Normali- zation Trainable Frozen RO-PnR Probe Respond Y 3 R 1 ... R 2 R N R 3 A 1 ... A 2 A N A 3 Goals Elicit uncertainty Encourage information probing Maintain Efficiency Inference: RolloutMulti - TurnInteraction Training: UpdatePolicyviaGRPO R pro >R res R pro >R res R pro >R res R res >R pro 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 compares RESPOND (current quality) against PROBE (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. 2024) built in two layers, each addressing a distinct aspect of the probe-or-respond trade-off. (1) Forward-looking value. The value of ask- ing is not in the question itself, but in the better intervention it enables later in the dialogue. A re- ward 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 (Wu et al., 2025). We therefore compare the response quality available now against the ex- pected quality at conversation end after probing: r t (RESPOND) = R(h t ), r t (PROBE) =E R(h end )| h t , PROBE , whereR(Ā·)scores a dialogue Quality using the av- erage of audience alignment, personalized ground- ing, and tailored actionability considering dialogue history and response, andh end is the state at which the agent eventually responds. (See Section 4.3 for Quality evaluation details.) (2) Interaction cost. A forward-looking gain alone treats every probe as free, which would incen- tivize the agent to ask whenever any improvement is expected, even arbitrarily small ones. In practice, however, every clarification question imposes a real cost (Dong et al., 2026). An agent that ignores this cost will over-probe, degrading user experience even when the marginal information gain is neg- ligible. We therefore subtract a per-question cost from the probe action that scales with the number of clarifications already asked, referring to (Dong et al., 2026): r t (PROBE) = R future ā cĀ· n t , wherec > 0is the per-question cost andn t is the number of clarification questions asked up to turnt. 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. Putting it together.At each turn, the agent picks the action with the higher reward, choosing PROBE when R future (h t )ā R(h t ) > cĀ· n t , and RESPOND otherwise. Intuitively, the agent probes only when the expected long-horizon im- provement 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. 3.5 RO-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 policyās per-step probe- or-respond choice. At each decision state extracted 4 from our multi-turn rollouts, we build paired super- vision: a RESPOND record scored by the immediate response quality, and a PROBE record 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. 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) (Shao et al., 2024). 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, 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. 4 Experiment Setup 4.1 Dataset 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 and dataset construc- tion details in Appendix B. Our primary dataset is derived from CounterHealth (See Appendix C), and is used for model fine-tuning. We addition- ally use two public health-misinformation datasets for evaluation, MisinfoCorrect(He et al., 2023) and PUBHEALTH (Kotonya and Toni, 2020). 4.2 User 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- mini 1 , which has been increasingly adopted as a proxy for simulating user behavior in interactive dialogue settings (Kim et al., 2025; Laban et al., 2025; Liu et al., 2026; Lin et al., 2024). 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 1 https://platform.openai.com/docs/models/ gpt-4o-mini prompt is shown in Figure 6, with additional details provided in Appendix A.3. To evaluate whether the LLM-based user sim- ulator faithfully follows the assigned user profile, we conduct human evaluation along two dimen- sions: Persona Accuracy and Persona Consistency, referring to (Wang et al., 2025). 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 di- alogue. We instruct human annotators to rate each dimension on a 5-point scale. The detailed annota- tion rubric is provided in Table 6. The results are discussed in Section 7. 4.3 Evaluation Metrics We evaluate system performance along three com- plementary dimensions: adaptation, factual relia- bility, 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 cor- rection depends not only on what is said but on how it is framed for a specific user (Krishnamurthy and Hu, 2026; Nguyen et al., 2019). We assess this with three sub-metrics: Audience Alignment (A) (Song et al., 2025; Cima et al., 2025), which measures whether the correction is framed appropriately for the userās inferred health literacy and belief com- mitment; Personalized Grounding (PG) (Gabriel et al., 2024), which measures whether the evidence and reasoning supporting the correction are clear and comprehensible to the user; and Tailored Ac- tionability (TA) (Song et al., 2025; Ownby et al., 2026), 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-2512 2 ) given the dialogue his- tory and the userās latent state. (See Figure 8 in the Appendix for detailed rubrics.) Factual reliability. A correction that introduces new inaccuracies can undermine user trust and lead to harmful decisions (Yue et al., 2024), so adapta- tion alone is not sufficient. We report the factual 2 https://docs.mistral.ai/models/model-cards/ mistral-large-3-25-12 5 DatasetMethod Llama-3.1-8B-InstructQwen3-8BGemma-4-E4B-it A PGTA Quality FERā Turnsā Utility A PGTA Quality FERā Turnsā Utility A PGTA Quality FERā Turnsā Utility CounterHealth Single-Turn0.58 0.53 0.610.570.185.000.560.58 0.52 0.610.570.155.000.560.64 0.59 0.670.630.045.000.62 Fixed-Q0.72 0.65 0.760.710.115.000.700.71 0.64 0.730.690.085.000.680.73 0.66 0.740.710.025.000.70 Reactive0.72 0.64 0.760.700.115.000.690.71 0.64 0.720.690.095.000.680.72 0.65 0.740.710.023.920.70 Confidence0.71 0.64 0.740.690.131.260.690.67 0.61 0.700.660.061.420.660.72 0.64 0.760.710.021.850.71 SFT0.70 0.64 0.760.700.074.380.690.70 0.64 0.760.700.054.220.690.70 0.64 0.760.700.094.610.69 RO-PnR (Ours) ā 0.720.660.770.720.063.560.710.720.670.780.720.083.590.720.730.670.770.720.093.590.72 MisinfoCorrect Single-Turn0.55 0.50 0.580.540.075.000.530.55 0.49 0.590.540.055.000.540.65 0.60 0.670.640.015.000.63 Fixed-Q0.70 0.63 0.720.690.075.000.680.71 0.63 0.720.690.035.000.680.71 0.64 0.710.690.035.000.68 Reactive0.71 0.64 0.740.700.064.980.690.71 0.63 0.710.690.035.000.680.70 0.63 0.700.680.033.960.67 Confidence0.61 0.54 0.640.600.081.100.600.59 0.52 0.640.580.031.080.580.63 0.55 0.690.620.021.350.62 SFT0.71 0.65 0.760.710.064.010.700.71 0.65 0.760.700.073.910.700.71 0.65 0.760.710.084.030.70 RO-PnR (Ours) ā 0.730.680.770.720.073.560.720.740.680.770.730.073.460.720.740.690.770.730.083.550.73 PUBHEALTH Single-Turn0.63 0.58 0.650.620.245.000.610.62 0.56 0.640.610.215.000.600.66 0.61 0.680.650.095.000.64 Fixed-Q0.73 0.66 0.770.720.135.000.710.73 0.67 0.750.720.125.000.710.74 0.67 0.770.730.015.000.72 Reactive0.74 0.66 0.780.730.144.930.720.75 0.68 0.750.730.125.000.720.74 0.67 0.770.730.013.820.72 Confidence0.69 0.62 0.720.680.191.430.680.71 0.64 0.730.690.141.630.690.75 0.67 0.790.740.062.030.74 SFT0.72 0.66 0.780.720.064.230.720.73 0.67 0.780.720.054.240.720.73 0.67 0.780.720.094.600.72 RO-PnR (Ours) ā 0.740.690.790.740.083.530.730.740.690.790.740.083.800.730.740.690.780.740.143.570.73 Table 1: Average results across all health-literacy and belief-commitment configurations, for each method, dataset, and base model. A, PG, TA are the Audience Alignment (A), Personalized Grounding(PG), and Tailored Actionability (TA) scores (normalized to[0, 1]); Quality is the overall quality of A, PG and TA (normalized to[0, 1]); FER is the factual error rate (ā); Turns is the average number of dialogue turns (ā); and Utility is the cost-adjusted Quality at c=0.01, our headline metric (normalized to [0, 1]). error rate: the fraction of responses flagged as con- taining any factual error by an LLM judge (GPT-5- mini 3 ) trained with updated knowledge and have access to web-based verification (He et al., 2023). Lower values indicate better performance. Prompt is in Figure 7 in the Appendix. Dialogue quality. In addition to adaptation met- rics, we also measure the efficiency of the inter- action. Turns (Wu et al., 2025) counts the num- ber of communication turns, with fewer turns indi- cating a more efficient interaction. Utility (Dong et al., 2026) captures the trade-off between re- sponse quality and interaction cost. Since the three adaptation dimensions are measured on the same scale and are jointly necessary for an adap- tive response, we define overall Quality as their mean,Q = (S A + S PG + S TA )/3, and utility as U = Qā c, wherecis the cost incurred by probe actions. 4.4 Baselines We compare our RO-PnR framework against four alternative decision strategies: (a) Single-Turn Di- rect Response (Single-Turn) (Song et al., 2025): The model generates the correction immediately based only on the current turn, without asking clar- ification questions. (b) Fixed-Question Probing (Fixed-Q) (Fu and Du, 2025): The agent asks a predetermined number of clarification questions 3 https://platform.openai.com/docs/models/ gpt-5-mini CounterHealthMisinfoPubhealth MethodModelFICOv.FICQ.FICOv. Single-Turn Llama0.51 0.60 0.57 0.56 0.49 0.57 0.55 0.53 0.57 0.64 0.62 0.61 Qwen0.50 0.59 0.59 0.56 0.48 0.56 0.56 0.54 0.54 0.63 0.62 0.60 Gemma 0.54 0.66 0.67 0.62 0.56 0.67 0.67 0.63 0.56 0.67 0.68 0.64 Fixed-Q Llama0.660.720.72 0.70 0.620.710.70 0.68 0.670.730.730.71 Qwen0.620.710.72 0.68 0.630.71 0.70 0.68 0.670.720.730.71 Gemma 0.640.720.74 0.70 0.630.70 0.70 0.68 0.670.730.750.72 Reactive Llama0.650.720.72 0.69 0.640.720.71 0.69 0.670.730.740.72 Qwen0.620.710.72 0.68 0.620.71 0.70 0.68 0.670.730.740.71 Gemma 0.630.720.74 0.70 0.62 0.70 0.69 0.67 0.670.730.760.72 Confidence Llama0.66 0.72 0.70 0.69 0.60 0.63 0.56 0.60 0.650.71 0.67 0.68 Qwen0.59 0.71 0.68 0.66 0.57 0.62 0.56 0.58 0.640.730.70 0.69 Gemma0.700.720.700.71 0.66 0.64 0.56 0.620.720.740.750.74 SFT Llama0.640.720.72 0.69 0.640.730.730.70 0.670.740.740.71 Qwen0.640.720.72 0.69 0.630.730.72 0.70 0.660.740.740.72 Gemma 0.630.720.72 0.69 0.640.730.740.70 0.660.730.750.72 RO-PnR Llama0.650.740.740.71 0.660.750.750.72 0.680.760.770.73 Qwen0.660.750.740.72 0.660.750.750.72 0.670.760.770.73 Gemma 0.65 0.750.750.72 0.670.750.760.73 0.670.750.770.73 Table 2: Utility by Health Literacy Level (normalized to[0, 1]). Higher-value cells are highlighted with light tints: green (ā„ 0.70), mint (ā„ 0.73), and blue (ā„ 0.75). F, I, C, Ov., denotes Functional, Interactive, Critical, and all populations. before producing the final counterspeech response, regardless of the userās feedback. (c) Reactive Clarifier (Reactive) (Zhang and Choi, 2025): The agent adaptively chooses between asking and an- swering based on user feedback, without consid- ering the userās knowledge background or stance toward the claim. (d) Confidence-Gated Clarifier (Confidence) (Li et al., 2024): 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 6 CounterHealthMisinfoPubhealth MethodModelSHOOv.SHOOv.SHOOv. Single-Turn Llama0.50 0.54 0.64 0.56 0.48 0.52 0.61 0.53 0.55 0.57 0.71 0.61 Qwen0.50 0.54 0.64 0.56 0.48 0.52 0.60 0.54 0.53 0.57 0.69 0.60 Gemma 0.57 0.62 0.69 0.62 0.58 0.62 0.69 0.63 0.58 0.620.71 0.64 Fixed-Q Llama0.630.710.75 0.70 0.60 0.690.74 0.68 0.650.720.760.71 Qwen0.610.710.73 0.68 0.600.700.73 0.68 0.650.720.750.71 Gemma 0.61 0.730.76 0.70 0.570.710.75 0.68 0.650.740.760.72 Reactive Llama0.620.710.76 0.69 0.62 0.700.75 0.69 0.660.730.760.72 Qwen0.61 0.70 0.74 0.68 0.60 0.690.74 0.68 0.650.730.760.72 Gemma 0.620.720.76 0.70 0.57 0.690.75 0.67 0.650.740.770.72 Confidence Llama0.65 0.700.73 0.69 0.58 0.59 0.62 0.60 0.65 0.680.71 0.68 Qwen0.63 0.66 0.69 0.66 0.57 0.57 0.60 0.58 0.64 0.700.73 0.69 Gemma 0.670.710.730.71 0.61 0.62 0.64 0.62 0.690.750.770.74 SFT Llama0.620.710.75 0.69 0.640.720.760.70 0.650.730.760.71 Qwen0.630.710.75 0.69 0.630.710.75 0.70 0.660.730.760.72 Gemma 0.62 0.710.75 0.69 0.640.710.75 0.70 0.660.730.760.72 RO-PnR Llama0.65 0.720.760.71 0.660.720.770.72 0.690.740.770.73 Qwen0.660.720.760.72 0.670.730.770.72 0.690.740.770.73 Gemma 0.670.720.760.72 0.690.730.770.73 0.690.730.770.73 Table 3: Utility by Belief Commitment. Higher-value cells are highlighted with light tints: green (ā„ 0.70), mint (ā„ 0.73), and blue (ā„ 0.75); lower values are left uncolored. S, H, O, Ov., denotes Strong, Hesitant, Open, and all populations. central to misinformation correction. (e) Super- vised 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. 4.5 Implementation Details We evaluate RO-PnR on three open-source 8B models: LLAMA-3.1-8B-INSTRUCT 4 , GEMMA-4- E4B-IT 5 , and QWEN3-8B 6 . Fine-tuning consists of two stages: offline supervised fine-tuning (SFT) and offline GRPO (Shao et al., 2024). We first perform SFT on base models using full-profile di- alogue examples with LoRA adapters (Hu et al., 2021). 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. 5 Results Main Results. RO-PnR achieves the best over- all utility while using substantially fewer interac- tion turns than always-probe baselines (Table 1). RO-PnR obtains the highest utility on nearly every (dataset, model) combination, reaching 0.71ā0.73 in utility versus 0.68-0.72 for the strongest baseline 4 https://huggingface.co/meta-llama/Llama-3. 1-8B-Instruct 5 https://huggingface.co/google/gemma-4-E4B-it 6 https://huggingface.co/Qwen/Qwen3-8B (SFT), with consistent gains across all three adap- tation dimensions (A, 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 (ā„ 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, espe- cially on MisinfoCorrect, where utility falls to 0.58- 0.62. RO-PnR occupies the productive middle, us- ing roughly 3.5 turns, about 30% fewer than the always-probe baselines, while delivering the high- est utility. Factual error rates remain low across methods (0.06-0.14 for RO-PnR), indicating that adaptation gains do not compromise factual relia- bility. Performance Across User Profiles. RO-PnRās utility gains hold across user types, with the largest improvements on the hardest user slices. Tables 2 and 3 break utility down by health literacy and be- lief commitment. RO-PnR achieves the highest overall utility in every (dataset, model) cell, and its advantage is most pronounced on the most chal- lenging 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 0.65ā0.69 versus 0.6-0.66 for SFT and only 0.57-0.61 for Con- fidence, indicating that indiscriminate probing fails to overcome resistance while confidence-only prob- ing 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 com- prehension and the openness dimensions of user heterogeneity. 6 Ablation Study To investigate how different components contribute to RO-PnR performance, we conduct two abla- tions on the GRPO reward (Table 4). (a) Reward- horizon ablation replaces the forward-looking re- ward with a one-step reward based only on the next user response, testing whether short-horizon credit assignment suffices. (b) Probe-cost sweep varies the per-question costcto probe the trade-off between information gathering and early commit- 7 VariantAAāPGāTAāQualityāFERāTurnsāUtilityā RO-PnR (full)0.720.660.770.720.063.560.71 (a) Reward-horizon ablation immediate next-gain only0.620.540.650.600.122.000.60 (b) Probe-cost (c) sweep probe cost = 0.000.500.460.570.510.115.830.51 probe cost = 0.030.710.650.770.710.083.000.70 probe cost = 0.050.720.660.770.710.093.000.69 Table 4: Ablation study results, averaged over all (health literacyĆbelief commitment) user profiles. RO-PnR (full) is our complete method with forward-looking long-horizon reward, probe costc = 0.01, and cross- threshold bonus. (a) replaces the forward-looking re- ward with the immediate next-turn gain; (b) sweeps the per-turn probe cost c. ment. Table 4 confirms that all components contribute to performance. The reward-horizon ablation de- grades utility from 0.71 to 0.60, showing that a probeās value often materializes several turns later and cannot be captured by a one-step reward. The probe-cost sweep shows thatc = 0leads to over- probing (5.83 turns) and a quality collapse to 0.51, while moderate costs (c = 0.03-0.05) recover near- best performance atā¼3turns, confirming that the policy is stable across a reasonable range of c. 7 Human Validation While LLMs are known to reliably simulate hu- mans and act as judges at low cost (Thakur et al., 2025; Bavaresco et al., 2025; Huang et al., 2025), we run a small-scale human evaluation along three axes: (1) simulator reliability via persona accu- racy and consistency; (2) human-judge agreement on adaptation ratings against MISTRAL-LARGE- 2512 (rubric in Figure 8); and (3) a pairwise pref- erence study comparing RO-PNR against the base- line (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. LLM Simulator Evaluation. Annotators show substantial agreement on both axes (Table 7): mean pairwise agreement is 0.83 for persona accuracy and 0.85 for consistency, with multi-rater Scottās Ļof 0.70 and 0.68. Using the median of the three annotatorsā ratings as the final score (Table 8), per- sona accuracy averages 3.87, with 76% of samples scoring 4 or 5, indicating that simulated users gen- erally 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 es- tablished, the simulator maintains it across turns. Details are in Appendix D.1. Human-judge agreement. Human annotators show moderate-to-strong internal consistency (ScottāsĻ= 0.63ā0.78 across dimensions; Table 9). Using the average human rating as consensus, the LLM judge aligns well with humans on Quality: ScottāsĻ= 0.55, 73% exact agreement, MAE = 0.23, and 87% of samples withinĀ±0.5points (Ta- ble 10), with Pearson = 0.57 and Spearman = 0.72 (Table 11). This indicates that the LLM judge is a reliable proxy for human evaluation at scale. Human pairwise preference. As shown in Fig- ure 4, RO-PnR is strongly preferred overall (79.70% vs. 20.30% for Fixed-Q), primarily be- cause its responses are perceived as more explana- tory, credible, and actionable. The exception is User 2 (functional literacy, strong belief commit- ment), who favors Fixed-Q for being easier to pro- cess, less confrontational, and less reliant on in- stitutional authority, consistent with prior findings that users distrustful of official health sources resist authority-heavy framings even when the evidence is richer (Jamison et al., 2019). This highlights a trade-off: while RO-PNR yields higher-quality cor- rections in aggregate, users with strong prior com- mitments may benefit from softer, less institution- centered framing to reduce resistance. 8 Conclusion In this paper, we introduced RO-PnR, a decision framework for multi-turn health misinformation in- tervention that learns when to ask clarifying ques- tions and when to correct. By modeling hidden user differences in health literacy and belief com- mitment, 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 effec- tive 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 inter- actions, richer models of user burden, and joint optimization of both when and how to probe. 8 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 simu- lation can provide a useful approximation (Wang et al., 2025; Sekuli Ģ c et al., 2024; Bougie and Watan- abe, 2025), and we make efforts to improve its reli- ability, it may still fall short of capturing the full nu- ance and variability of real users. Despite this limi- tation, our simulation framework offers a practical testbed for studying how health-misinformation- related user characteristics affect multi-turn coun- terspeech. 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 communi- cation cost following Dong et al. (2026) 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 com- munication 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 interac- tion. References Anirban Saha Anik, Xiaoying Song, Elliott Wang, Bryan Wang, Bengisu Yarimbas, and Lingzi Hong. 2025. Multi-agent retrieval-augmented framework for evidence-based counterspeech against health mis- information. In Second Conference on Language Modeling. Anna Bavaresco, Raffaella Bernardi, Leonardo Berto- lazzi, Desmond Elliott, Raquel FernĆ”ndez, Albert Gatt, Esam Ghaleb, Mario Giulianelli, Michael Hanna, Alexander Koller, Andre Martins, Philipp Mondorf, Vera Neplenbroek, Sandro Pezzelle, Bar- bara Plank, David Schlangen, Alessandro Suglia, Aditya K Surikuchi, Ece Takmaz, and Alberto Testoni. 2025. LLMs instead of human judges? a large scale empirical study across 20 NLP evalua- tion tasks. In Proceedings of the 63rd Annual Meet- ing of the Association for Computational Linguistics (Volume 2: Short Papers), pages 238ā255, Vienna, Austria. Association for Computational Linguistics. Nancy D Berkman, Stacey L Sheridan, Katrina E Don- ahue, David J Halpern, and Karen Crotty. 2011. Low health literacy and health outcomes: an up- dated systematic review. Annals of internal medicine, 155(2):97ā107. Nicolas Bougie and Narimawa Watanabe. 2025. Simuser: Simulating user behavior with large lan- guage models for recommender system evaluation. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 6: Industry Track), pages 43ā60. Maximillian Chen, Ruoxi Sun, Tomas Pfister, and Sercan Ć Arık. 2024. Learning to clarify: Multi- turn conversations with action-based contrastive self- training. arXiv preprint arXiv:2406.00222. Lorenzo Cima, Alessio Miaschi, Amaury Trujillo, Marco Avvenuti, Felice DellāOrletta, and Stefano Cresci. 2025. Contextualized counterspeech: Strate- gies for adaptation, personalization, and evaluation. In Proceedings of the ACM on Web Conference 2025, pages 5022ā5033. Yijiang River Dong, Tiancheng Hu, Zheng Hui, Caiqi Zhang, Ivan Vuli Ģ c, Andreea Bobu, and Nigel Col- lier. 2026. Value of information: A framework for human-agent communication. arXiv preprint arXiv:2601.06407. Ullrich K. H. Ecker, Stephan Lewandowsky, John Cook, Philipp Schmid, Lisa K. Fazio, Nadia Brashier, Panayiota Kendeou, Emily K. Vraga, and Michelle A. Amazeen. 2022a. The psychological drivers of mis- information belief and its resistance to correction. Nature Reviews Psychology, 1(1):13ā29. Ullrich KH Ecker, Stephan Lewandowsky, John Cook, Philipp Schmid, Lisa K Fazio, Nadia Brashier, Panayiota Kendeou, Emily K Vraga, and Michelle A Amazeen. 2022b. The psychological drivers of mis- information belief and its resistance to correction. Nature Reviews Psychology, 1(1):13ā29. Chuanruo Fu and Yuncheng Du. 2025. First ask then answer: A framework design for ai dialogue based on supplementary questioning with large language models. arXiv preprint arXiv:2508.08308. Saadia Gabriel, Liang Lyu, James Siderius, Marzyeh Ghassemi, Jacob Andreas, and Asuman E Ozdaglar. 2024. Misinfoeval: Generative ai in the era of āalter- native factsā. In Proceedings of the 2024 Conference 9 on Empirical Methods in Natural Language Process- ing, pages 8566ā8578. Zhaolin Gao, Wenhao Zhan, Jonathan D Chang, Gokul Swamy, KiantĆ© Brantley, Jason D Lee, and Wen Sun. 2024. Regressing the relative future: Efficient pol- icy optimization for multi-turn rlhf. arXiv preprint arXiv:2410.04612. Valentin Guigon, Lucille Geay, and Caroline J Charpen- tier. 2026. Rethinking misinformation through plau- sibility estimation and confidence calibration. Com- munications Psychology, 4(1):24. Bing He, Mustaque Ahamad, and Srijan Kumar. 2023.Reinforcement learning-based counter- misinformation response generation: a case study of covid-19 vaccine misinformation. In Proceedings of the ACM Web Conference 2023, pages 2698ā2709. Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2021. Lora: Low-rank adaptation of large language models. Preprint, arXiv:2106.09685. Hui Huang, Xingyuan Bu, Hongli Zhou, Yingqi Qu, Jing Liu, Muyun Yang, Bing Xu, and Tiejun Zhao. 2025. An empirical study of llm-as-a-judge for llm evaluation: Fine-tuned judge model is not a general substitute for gpt-4. In Findings of the Association for Computational Linguistics: ACL 2025, pages 5880ā5895. Hirono Ishikawa, Takeaki Takeuchi, and Eiji Yano. 2008. Measuring functional, communicative, and critical health literacy among diabetic patients. Diabetes care, 31(5):874ā879. Amelia M Jamison, Sandra Crouse Quinn, and Vicki S Freimuth. 2019. āyou donāt trust a government vac- cineā: Narratives of institutional trust and influenza vaccination among african american and white adults. Social science & medicine, 221:87ā94. Katrina P Jongman-Sereno, Rick H Hoyle, Erin K Davisson, and Jinyoung Park. 2023. Intellectual hu- mility and responsiveness to public health recom- mendations. Personality and individual differences, 211:112243. Sebastian Joseph, Lily Chen, Barry Wei, Michael Mack- ert, Iain J Marshall, Paul Pu Liang, Ramez Kouzy, Byron C Wallace, and Junyi Jessy Li. 2025. Decide less, communicate more: On the construct validity of end-to-end fact-checking in medicine. arXiv preprint arXiv:2506.20876. Elise Karinshak, Sunny Xun Liu, Joon Sung Park, and Jeffrey T Hancock. 2023. Working with ai to per- suade: Examining a large language modelās abil- ity to generate pro-vaccination messages. Proceed- ings of the ACM on Human-Computer Interaction, 7(CSCW1):1ā29. Paige L Kemp, Aaron C Goldman, and Christopher N Wahlheim. 2024. On the role of memory in misinfor- mation corrections: Repeated exposure, correction durability, and source credibility. Current Opinion in Psychology, 56:101783. Minju Kim, Dongje Yoo, Yeonjun Hwang, Minseok Kang, Namyoung Kim, Minju Gwak, Beong-woo Kwak, Hyungjoo Chae, Harim Kim, Yunjoong Lee, Min Hee Kim, Dayi Jung, Kyong-Mee Chung, and Jinyoung Yeo. 2025. Can you share your story? mod- eling clientsā metacognition and openness for LLM therapist evaluation. In Findings of the Association for Computational Linguistics: ACL 2025, pages 25943ā25962, Vienna, Austria. Association for Com- putational Linguistics. Neema Kotonya and Francesca Toni. 2020. Explainable automated fact-checking for public health claims. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 7740ā7754. Parthasarathy Krishnamurthy and Ye Hu. 2026. Covid- 19 vaccine framing and acceptance among adults who are vaccine hesitant. JAMA Network Open, 9(3):e264114. Philippe Laban, Hiroaki Hayashi, Yingbo Zhou, and Jennifer Neville. 2025. Llms get lost in multi-turn conversation. arXiv preprint arXiv:2505.06120. Stephan Lewandowsky, Ullrich K. H. Ecker, Colleen M. Seifert, Norbert Schwarz, and John Cook. 2012. Misinformation and its correction: Continued influ- ence and successful debiasing. Psychological Sci- ence in the Public Interest, 13(3):106ā131. PMID: 26173286. Shuyue Stella Li, Vidhisha Balachandran, Shangbin Feng, Jonathan S. Ilgen, Emma Pierson, Pang Wei Koh, and Yulia Tsvetkov. 2024. Mediq: Question- asking llms and a benchmark for reliable interactive clinical reasoning. In Advances in Neural Informa- tion Processing Systems, volume 37, pages 28858ā 28888. Curran Associates, Inc. Xiaoyu Lin, Xinkai Yu, Ankit Aich, Salvatore Giorgi, and Lyle Ungar. 2024. Diversedialogue: A methodol- ogy for designing chatbots with human-like diversity. arXiv preprint arXiv:2409.00262. Yu Lu Liu, Hyokun Yun, Tanya Roosta, and Ziang Xiao. 2026. Synthetic users, real differences: an evaluation framework for user simulation in multi-turn conver- sations. arXiv preprint arXiv:2605.02624. S Emlen Metz. 2023. Building better beliefs through ac- tively open-minded thinking. The cognitive science of belief, pages 574ā591. Xiaoli Nan, Yuan Wang, and Kathryn Thier. 2022. Why do people believe health misinformation and who is at risk? a systematic review of individual differences in susceptibility to health misinformation. Social Science & Medicine, 314:115398. 10 Eryn J Newman, Briony Swire-Thompson, and Ull- rich KH Ecker. 2022. Misinformation and the sins of memory: False-belief formation and limits on belief revision. Minh Hao Nguyen, Ellen MA Smets, Nadine Bol, EugĆØne F Loos, Hanneke WM van Laarhoven, Debby Geijsen, Mark I van Berge Henegouwen, Kristien MAJ Tytgat, and Julia CM Van Weert. 2019. Tailored 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 Re- search, 21(10):e14407. Don Nutbeam. 2000. Health literacy as a public health goal: a challenge for contemporary health education and communication strategies into the 21st century. Health promotion international, 15(3):259ā267. Lydia Ogbadu-Oladapo, Kossi Bissadu, Heejun Kim, and Daniella LaShaun Smith. 2026. Information and health literacy: could there be any impact on health decision-making among adults?āevidence from north america.Journal of Public Health, 34(1):151ā181. Raymond L Ownby, Rosemary Davenport, and Joshua Caballero. 2026. User perceptions of individually- tailored health information in digital apps: devel- opment of a scale.Frontiers in Digital Health, 8:1731948. Kenny Peng and James Grimmelmann. 2024. Res- cuing counterspeech: A bridging-based approach to combating misinformation.arXiv preprint arXiv:2410.12699. Alireza Salemi, Sheshera Mysore, Michael Bendersky, and Hamed Zamani. 2024. Lamp: When large lan- guage models meet personalization. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 7370ā7392. Greta Arancia Sanna and David Lagnado. 2025. Belief updating in the face of misinformation: The role of source reliability. Cognition, 258:106090. Peter J Schulz and Kent Nakamoto. 2025. Understand- ing health knowledge failures: uncertainty versus misinformation. Scientific Reports, 15(1):23867. William A Scott. 1955. Reliability of content analysis: The case of nominal scale coding. Public opinion quarterly, pages 321ā325. Ivan Sekuli Ģ c, Silvia Terragni, Victor GuimarĆ£es, Nghia Khau, Bruna Guedes, Modestas Filipavicius, An- dre Ferreira Manso, and Roland Mathis. 2024. Reli- able llm-based user simulator for task-oriented dia- logue systems. In Proceedings of the 1st Workshop on Simulating Conversational Intelligence in Chat (SCI-CHAT 2024), pages 19ā35. Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, Y. K. Li, Y. Wu, and Daya Guo. 2024. Deepseekmath: Pushing the limits of mathemati- cal reasoning in open language models. Preprint, arXiv:2402.03300. Jana Siebert and Johannes Ulrich Siebert. 2023. Effec- tive mitigation of the belief perseverance bias after the retraction of misinformation: Awareness training and counter-speech. Plos one, 18(3):e0282202. Xiaoying Song, Anirban Saha Anik, Dibakar Barua, Pengcheng Luo, Junhua Ding, and Lingzi Hong. 2025. Speaking at the right level: Literacy-controlled counterspeech generation with RAG-RL. In Find- ings of the Association for Computational Linguistics: EMNLP 2025, pages 2812ā2830, Suzhou, China. As- sociation for Computational Linguistics. Kristine SĆørensen, Stephan Van den Broucke, James Ful- lam, Gerardine Doyle, JĆ¼rgen Pelikan, Zofia Slonska, Helmut Brand, and (HLS-EU) Consortium Health Literacy Project European. 2012. Health literacy and public health: A systematic review and integra- tion of definitions and models. BMC Public Health, 12(1):80. Briony Swire-Thompson, Nicholas Miklaucic, John P Wihbey, David Lazer, and Joseph DeGutis. 2022. The backfire effect after correcting misinformation is strongly associated with reliability. Journal of Experimental Psychology: General, 151(7):1655. Aman Singh Thakur, Kartik Choudhary, Venkat Srinik Ramayapally, Sankaran Vaidyanathan, and Dieuwke Hupkes. 2025. Judging the judges: Evaluating align- ment and vulnerabilities in llms-as-judges. In Pro- ceedings of the Fourth Workshop on Generation, Evaluation and Metrics (GEM 2 ), pages 404ā430. Sander Van Der Linden. 2022. Misinformation: sus- ceptibility, spread, and interventions to immunize the public. Nature medicine, 28(3):460ā467. Nathan Walter and Sheila T. Murphy. 2018. How to un- ring the bell: A meta-analytic approach to correction of misinformation. Communication Monographs, 85(3):423ā441. Yanming Wan, Jiaxing Wu, Marwa Abdulhai, Lior Shani, and Natasha Jaques. 2025. Enhancing per- sonalized multi-turn dialogue with curiosity reward. arXiv preprint arXiv:2504.03206. Kuang Wang, Xianfei Li, Shenghao Yang, Li Zhou, Feng Jiang, and Haizhou Li. 2025. Know you first and be you better: Modeling human-like user sim- ulators via implicit profiles. In Proceedings of the 63rd Annual Meeting of the Association for Compu- tational Linguistics (Volume 1: Long Papers), pages 21082ā21107. Chloe Wittenberg and Adam J Berinsky. 2020. Mis- information and its correction. Social media and democracy: The state of the field, prospects for re- form, pages 163ā198. 11 Shirley Wu, Michel Galley, Baolin Peng, Hao Cheng, Gavin Li, Yao Dou, Weixin Cai, James Zou, Jure Leskovec, and Jianfeng Gao. 2025. Collabllm: From passive responders to active collaborators. In Inter- national Conference on Machine Learning, pages 67260ā67283. PMLR. Zhenrui Yue, Huimin Zeng, Yimeng Lu, Lanyu Shang, Yang Zhang, and Dong Wang. 2024. Evidence-driven retrieval augmented response generation for online misinformation. In Proceedings of the 2024 Confer- ence of the North American Chapter of the Associ- ation for Computational Linguistics: Human Lan- guage Technologies (Volume 1: Long Papers), pages 5628ā5643. Michael JQ Zhang and Eunsol Choi. 2025. Clarify when necessary: Resolving ambiguity through in- teraction with lms. In Findings of the Association for Computational Linguistics: NAACL 2025, pages 5526ā5543. Yifei Zhou and Andrea Zanette. 2024. Archer: training language model agents via hierarchical multi-turn rl. In Proceedings of the 41st International Conference on Machine Learning, pages 62178ā62209. Jiayuan Zhu, Jiazhen Pan, Yuyuan Liu, Fenglin Liu, and Junde Wu. 2025. Ask patients with patience: En- abling llms for human-centric medical dialogue with grounded reasoning. In Proceedings of the 2025 Con- ference on Empirical Methods in Natural Language Processing, pages 2846ā2857. 12 A User Modeling A.1 Health literacy Functional health literacy means basic read- ing/writing and understanding skills sufficient to function effectively in everyday health situations, such as understanding health information and ser- vice instructions (Nutbeam, 2000). Interactive health literacy is more advanced per- sonal, communicative, and social skills that enable independent action on health knowledge (Nutbeam, 2000). Critical health literacy This is the most advanced level. Nutbeam (2000) defines it in terms of ad- vanced cognitive and social skills used to critically analyze information and act on broader determi- nants of health at individual and community levels. A.2 Belief Commitment Strong belief commitment can be understood as a state in which individuals accept misinformation as true, allow it to continue shaping subsequent at- titudes and judgments (Newman et al., 2022), and show difficulty revising that belief even after cor- rective information is provided (Wittenberg and Berinsky, 2020). In stronger cases, correction may even increase belief in the original misconception, indicating an especially resistant form of commit- ment (Swire-Thompson et al., 2022). Hesitant belief commitment is a state in which a person gives a misinformation claim partial ac- ceptance, but does so with noticeable uncertainty or limited confidence (Guigon et al., 2026; Schulz and Nakamoto, 2025), so the belief is not fully con- solidated and remains comparatively more open to revision than a strong misinformation belief (Kemp et al., 2024). Open belief commitment refers to a revisable form of belief commitment in which individuals may initially accept a claim, but remain willing to consider alternative evidence, evaluate source cred- ibility, and update their beliefs when credible cor- rective information becomes available (Jongman- Sereno et al., 2023; Metz, 2023). This form of commitment is better understood as an orienta- tion toward revisability and evidence-based updat- ing, rather than simply as weak belief (Sanna and Lagnado, 2025). A.3 User Simulation We simulate user replies with a profile-conditioned LLM environment. For the main results, each dia- logue is paired with a fixed latent user profile u = (health_literacy, belief_commitment), and the same profile is kept fixed throughout the entire interaction. We evaluate all posts under the full3 Ć 3grid of user types: health literacy in Functional, Interactive, Critical and belief com- mitment in Open, Hesitant, Strong. The profile dimensions are defined behaviorally rather than demographically. Referring to the defi- nition of Nutbeam (2000), we specify that Health literacy controls how complex the userās reason- ing sounds: Functional-literacy users prefer sim- ple language and rely more on stories or surface cues; Interactive-literacy users show some inter- est 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, we detail that Belief com- mitment controls how resistant the user is to cor- rection: 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 assistantās latest question, and generates one short reply. In our main setting, the simulator is implemented with gpt-4o-miniat temperature0.3. The prompt ex- plicitly conditions on the latent user state but in- structs the model not to reveal it directly. Its struc- ture is in Figure 6. The prompt further specifies style and behav- ioral constraints. It requires casual, brief replies in 1ā2 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 skepti- cal of correction, hesitant users to sound torn and uncertain, and open users to sound receptive and non-defensive. 13 Strong believer Defends the claim and hard to persuade Hesitant Expresses concern and uncertainty Open Willing to revise views with credible evidence Critical-Strong believer āI donāt think that study proves anything.ā There must be something wrong with their methods or agenda.ā Critical-Hesitant āIām not fully convinced yet. Can you show me more high-quality evidence or explain the mechanism in more detail?ā Critical-Open āThat makes sense. If the study is well-designed and from a reliable source, Iām willing to change my mind.ā Critical Uses complex reasoning; refers to evidence quality, mechanisms, study design, or source credibility. Interactive-Strong believer āIāve heard differently from others I trust.ā Iāl need more than that to believe it.ā InteractiveāHesitant āHmm, Iām not sure... Itās possible, but Iād like to see more sources or hear other opinions.ā InteractiveāOpen āI see your point! If it comes from a trustworthy source, Iām open to it.ā Interactive Some interest in proof or source credibility without much technical detail. Functional-Strong believer āI just know itās true because Iāve seen it for myself.ā Functional-Hesitant āIām a bit worried, but Iām not sure what to think yet.ā Functional-Open āOh, I didnāt know that. Thanks for explaining in simple words.ā Functional Relies on simple language, personal stories, or surface cues. Health Literac y Belief Commitment Figure 3: 3Ć3 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 settings SFT5Ć 10 ā5 4182048LoRAr = 16,Ī± = 32, dropout0.05; bf16; 100 warmup steps; seed 42 Offline GRPO 5Ć 10 ā6 2182048 Ī² KL = 0.05; SFT CE weight0.2; warmup ratio 0.03; max grad norm 1.0; seed 42 Table 5: Hyperparameters for SFT and GRPO configuration To reduce role drift, we combine several safe- guards. 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 con- straints: short replies, no belief-state summary, no multiple questions, and no unnecessary factual ex- position. 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 pro- file is hidden from the policy and only affects the environment-side reply generation. The assistant must infer how to adapt from the userās language and reactions, rather than from direct access to the profile metadata. B Training Dataset Construction We construct training data to match the reward de- sign at the level of turn-level decisions, rather than treating each dialogue as a single terminal exam- ple. The starting point is a collection of scored multi-turn rollouts generated for each misinforma- tion post under each of the nine user-profile con- ditions. For a fixed postāprofile 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 immedi- 14 ScorePersona AccuracyPersona Consistency 1The 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 unex- plained shifts in health literacy or belief commitment. 3The simulated user partially matches the assigned pro- file. Both dimensions are somewhat recognizable, but the dialogue is vague, generic, or only weakly aligned. The user is mostly stable, but there are noticeable in- consistencies or minor contradictions across turns. 4 The simulated user largely matches the assigned health literacy and belief commitment, with only minor imper- fections. The user maintains the assigned profile across most turns, and any change in belief strength is mostly justi- fied by the conversation. 5The simulated user strongly matches the assigned pro- file. 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. MetricPersona AccuracyPersona Consistency Mean Pairwise Agreement0.830.85 Mean Pairwise Scottās Ļ0.700.68 All-Three Agreement0.740.78 Multi-Rater Scottās Ļ0.700.68 Table 7: Inter-annotator agreement among three eval- uators on persona accuracy and persona consistency. Both pairwise and multi-rater statistics indicate substan- tial agreement. MetricScore 1Score 2Score 3Score 4Score 5Mean Persona Accuracy0%3%21%62%14%3.87 Persona Consistency0%0%3%30%67%4.64 Table 8: Distribution of human ratings for persona accu- racy and persona consistency. ately, while others continue probing and reveal the value of obtaining additional user information. From each trajectory, we extract prefix-level de- cision statesh t from the first four turns. Each state is converted into a structured assistant tar- get consisting of a latent belief estimate, an action token (PROBE or RESPOND), and the associated natural-language content. Crucially, the rollout trace provides both the value of responding un- der the current information state and the down- stream outcomes that become reachable if the dia- logue continues. We therefore build two kinds of supervision. A RESPOND record uses the candi- date response available at stateh t and is assigned the immediate stopping valueR now . A PROBE record uses the clarification question asked at that state and is assigned a forward-looking continua- tion value based on the best downstream outcome DimensionScottās ĻMean Pairwise AgreementAll-three Exact Agreement Audience 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. DimScottās ĻPercent AgreementMAEBiasWithinĀ±0.5WithinĀ±1.0 Audience 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. 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 tra- jectories that correspond to closely matched de- cision contexts, allowing PROBE and RESPOND to be compared locally rather than across unre- lated 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 which RESPOND is supervised by the current stopping value, while PROBE is supervised by future-sampled continuation value. This organi- zation makes the subsequent optimization directly 15 aligned with the intended probe-and-respond be- havior. For SFT, we select the higher Quality and fac- tually clean trajectory for each post/profile group. The selected trajectories are converted into step- level supervision examples: given the visible con- versation prefix, the model learns to output a struc- tured assistant action, either probe with a clarifica- tion question or respond with a final response. For GRPO, we use scored rollouts from all nine user-profile combinations. Each rollout is decom- posed 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 re- ward variation or without both probe and respond alternatives are filtered out. The remaining grouped examples provide normalized advantages for of- fline GRPO, initialized from the SFT adapter. C CounterHealth Dataset Collection We collected Reddit posts and comments on COVID-19, influenza, and HIV via the PRAW API 7 , using health-related keywords (e.g., āvac- cines,ā āCOVID-19,ā āalternative medicineā) to retrieve 4,968 posts and 17,223 comments from high-engagement subreddits. To identify misinfor- mation, 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 an- notations (F1= 0.76) and applied it to the remain- ing posts, surfacing 831 additional candidates. A final filtering pass using GPT-5-mini 8 with web- search-assisted human review produced 769 high- quality health-misinformation posts. On a 100-post validation sample, three annotators showed substan- tial agreement with the final labels (mean pairwise agreement= 87.4%, CohenāsĪŗā„ 0.67), confirm- ing the reliability of the filtering pipeline. D Human Validation D.1 LLM Simulator Evaluation We recruit three PhD students in health informatics to evaluate the quality of the LLM-based user sim- ulator along two dimensions: persona accuracy and persona consistency. Each annotator independently 7 https://praw.readthedocs.io/ 8 https://platform.openai.com/docs/models/ gpt-5-mini DimensionPearsonSpearmanLLM MeanHuman Mean Audience Alignment0.500.663.763.62 Personalized Grounding0.510.673.503.37 Tailored Actionability0.720.813.933.86 Quality0.570.723.733.62 Table 11: Correlation between the LLM judge and hu- man consensus under 0.5-point discretization. Figure 4: Human pairwise preference comparison be- tween RO-PnR and Fixed-Q across different users. De- tailed usersā background are provided in Table 13. rates 100 sampled simulated dialogues, spanning all 9 user profiles, on a 5-point Likert scale accord- ing to the rubric in Table 6. 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 commit- ment. User utterances are explicitly marked in the dialogue to encourage annotators to focus on whether the simulated userās replies match the as- signed profile. Before the formal annotation, an- notators are trained with a small set of example dialogues and discuss the rating criteria to ensure a shared understanding of the two persona dimen- sions. 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 pro- vided informed consent prior to the task. To assess annotation reliability, we compute pair- wise agreement and ScottāsĻamong the three an- notators on both dimensions (Table 7). Annota- 16 ItemQuestionRating BHow accurate do you think this claim is?0 = Definitely false / 10 = Definitely true CHow certain are you about your answer above?0 = Not at all / 10 = Completely OIf 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 CategoryOverall BC total BC Category User 12.07Communicative2.03Hesitant User 21.93Functional6.83Strong User 32.71Communicative1.20Open User 43.00Critical1.20Open User 54.00Critical0.40Open User 63.64Critical0.35Open Table 13: Health Literacy and Belief Commitment Cat- egories. 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 cate- gories are mapped fromBC total :BC total < 1.5= Open, 1.5ā¤ BC total < 5= Hesitant, andBC total ā„ 5= Strong. tors show substantial agreement on persona accu- racy (mean pairwise agreement= 0.83, multi-rater ScottāsĻ = 0.70) and persona consistency (mean pairwise agreement= 0.85, multi-rater Scottās Ļ = 0.68), indicating that the simulator can reli- ably instantiate the intended user profiles and main- tain them across the dialogue. D.2 Human-judge agreement We recruit three PhD students in health informatics to assess the reliability of the LLM judge. Fol- lowing Thakur et al. (2025), we quantify align- ment between human annotations and LLM rat- ings using percent agreement and ScottāsĻcoeffi- cient (Scott, 1955). We sample 100 samples from the results, stratified across (a) the three models (Llama / Gemma / Qwen), (b) all baseline and fine- tuning 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 adjudica- tion 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), with ScottāsĻranging from 0.63to0.78under 0.5-point discretization, moder- ate 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): on the aggregated Quality score, ScottāsĻ = 0.55with 73%exact agreement after binning, MAE= 0.23, and87%of samples falling withinĀ±0.5points. Rank-order correlation is also substantial (Pearson = 0.57, Spearman= 0.72; Table 11), indicating that the LLM judge and the human consensus rank responses similarly even when their exact scores differ slightly. Agreement is strongest on Tailored Actionability and overall Quality, while Personal- ized Grounding is 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.3 Human pairwise preference Due to budget constraints, we recruited six users with varying educational backgrounds: two with middle school education, two with high school edu- cation, and two PhD students. This sampling strat- egy was informed by prior evidence showing a pos- itive association between educational attainment and health literacy (Ishikawa et al., 2008). Each par- ticipant spent approximately one hour completing the screening questionnaire and the pairwise pref- erence 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 Func- tional, Communicative, and Critical Health Liter- acy (FCCHL) scale (Ishikawa et al., 2008). The FCCHL scale aligns with Nutbeamās categorization of health literacy into functional, communicative, and critical dimensions (Nutbeam, 2000), see ques- tionnaire in Figure 5. Based on the participantsā FCCHL scores, we used the sample median to cat- egorize them into lower and higher health literacy groups. In terms of the belief commitment screening measure, given that there are no established meth- ods available for this purpose, we define belief 17 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 compo- nents using three 0ā10 slider items for each of the three target misinformation claims, as shown in Table 12. For each claimi, we compute a 0-10 commitment score that gates acceptance(B i )by a strength multiplier formed from certainty(C i )and the complement of openness (O i ): commitment i = B i Ć C i + (10ā O i ) 20 . If the user does not accept the claim, meaning B 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: BC total = 1 3 3 X i=1 commitment i where BC total ranges from 0 to 10. The final screening results are shown in Table 13. Among the six users, three were classified as hav- ing a critical HL and open BC background, one as having a communicative HL and open BC back- ground, one as having a communicative HL and hesitant BC background, and one as having a func- tional HL and strong-believer BC background. We sample 50 pairs from one of the best base- lines (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, with- out knowing whether the responses are generated by RO-PnR or Fixed-Q. E Use of AI Assistants We used AI-assisted tools for support with code de- velopment and language refinement. All core ideas, methodological decisions, analyses, interpretations, and conclusions were developed by the authors. 18 Figure 5: Questionnaire for health literacy screen 19 System 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 the user role. Do not act as an expert, educator, or assistant. Respond naturally as a Reddit user with personal opinions and emotions. Keep replies short (1ā2 sentences). Current User State: State.to_dict() User State Dimension Guide: guide_text Interpret CURRENT_USER_STATE using the guide above and stay aligned with it. Conversation So Far: history_text if history_text else ā(no prior conversation)ā Assistant Message: question Generation Instruction: Generate exactly one realistic Reddit-style user reply. User Prompt / Output Constraints (JSON Reply Formatting): Guidelines: ā¢ Speak casually like a real Reddit user. ā¢ Do not summarize the belief state. ā¢ Do not explain medical facts unless the user would naturally know them. ā¢ Do not ask multiple questions. ā¢ Keep the reply brief. ā¢ 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: āfinal_responseā Evaluation Instructions: ā¢ Output Label: 1 if the response is factually correct and does not contain false or misleading claims. ā¢ Output Label: 0 if the response contains any false, misleading, or unsupported health-related claim. ā¢ Output Label: UNCERTAIN if the factual accuracy cannot be determined after verification. ā¢ Provide a brief reason for your judgment in 1ā3 sentences. Output Format: Label: 0, 1, or UNCERTAIN Reason: brief explanation of the judgment Figure 7: Prompt for evaluating the Factual Error. 20 Audience Alignment Whether the agent frames the correction in a way that is appropriate for the userās inferred health literacy and belief commitment. Score Range 1.0ā2.0: unclear, misleading, badly framed, or likely to provoke rejection. Score Range 2.1ā3.2: directionally appropriate, but weak, generic, indirect, or poorly matched to the userās literacy or belief commitment. Score Range 3.3ā3.8: reasonably clear and credible, with some fit to the userās stance, literacy, and comprehension. Score Range 3.9ā4.4: clearly frames the correction in a way that is well adapted to the userās stance, literacy, and belief commitment. Score Range 4.5ā5.0: exceptionally clear, credible, well-framed, and highly responsive to the userās literacy, comprehension, and belief commitment. Personalized Grounding Whether the agent gives useful, user-comprehensible evidence, mechanisms, or reasoning behind the correction. Score Range 1.0ā2.0: missing, inaccurate, vague, or mismatched to the userās comprehension. Score Range 2.1ā3.2: includes some reason or evidence, but it is shallow, generic, weakly connected, or not well matched to the userās understanding. Score Range 3.3ā3.8: provides useful reason, evidence, or mechanism that is moderately tailored to the userās comprehension. Score Range 3.9ā4.4: provides clear, relevant, and understandable evidence or reasoning that is well matched to the userās level and stance. Score Range 4.5ā5.0: provides especially sharp, credible, and well-translated evidence, mechanisms, or reasoning that strongly improves the userās understanding. Tailored Actionability Whether the agent provides safe, concrete, and user-appropriate actionable guidance when needed. Score Range 1.0ā2.0: no meaningful next step, unsafe, unrealistic, not actionable, or mismatched to the userās situation. Score Range 2.1ā3.2: offers some direction, but it is vague, generic, awkward, or only weakly matched to the userās needs. Score Range 3.3ā3.8: provides a reasonable, safe next step that the user could follow, with some attention to the userās situation. Score Range 3.9ā4.4: provides clear, practical, safe, and user-appropriate guidance that supports informed decision-making. Score Range 4.5ā5.0: provides unusually strong, natural, highly actionable, and well-tailored guidance, such as appropriate next steps or trusted information sources that fit the userās 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. 21 Health Misinformation Post: hm_post Dialogue History: history Latest User Reply: latest_user_reply Task: Decide between the following two actions: ā¢probe: ask one additional clarification question if the userās concern, evidence basis, or interpretation remains unclear and one more question would likely help improve the final response. ā¢ respond: stop asking questions if the current dialogue already provides enough information to produce an effective final response. Decision Guidance: ā¢ Choose probe when an important uncertainty remains unresolved. ā¢ Choose respond when the userās concern is already sufficiently clear and another question would likely be redundant. ā¢ Ask at most one question. ā¢ Do not repeat or paraphrase a previous question. ā¢ If you choose probe, the question must be brief, natural, and directly useful for improving the final counterspeech. ā¢ If you choose respond, do not ask any additional question. ā¢ If you choose respond, the counterspeech should follow the A/PG/TA principle: ā A (Audience Alignment): Frame the correction appropriately for the userās inferred health literacy and belief commitment. ā PG (Personalized Grounding): Provide clear, user-comprehensible evidence or reasoning to support the correction. ā TA (Tailored Actionability): Offer safe, concrete next steps suited to the user when needed. ā¢ Use only the visible conversation context. Do not mention or assume any hidden user state. Output Constraints: ā¢ Return only the required tagged fields. ā¢ Do not include explanations, reasoning, or extra text outside the tags. ā¢ If the decision is probe, fill the <question> field and leave <response> empty. ā¢ If the decision is respond, 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 userās 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. 22 Health Misinformation Post: hm_post Dialogue History: history Latest User Reply: latest_user_reply Task: First infer the userās likely state from the observed dialogue. Then decide between the following two actions: ā¢ 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. ā¢ respond: stop asking questions if your confidence in the inferred user state is high enough to produce an effective final counterspeech response. Decision Guidance: ā¢ Choose probe when your estimate of the userās state is still uncertain. ā¢ Choose respond when the userās likely state is sufficiently clear and another question is unlikely to substantially improve the final counterspeech. ā¢ Ask at most one question. ā¢ Do not repeat or paraphrase a previous question. ā¢If you choose probe, the question must be brief, natural, and directly useful for reducing uncertainty about the userās likely state or concern. ā¢ If you choose respond, do not ask any additional question. ā¢ If you choose respond, the counterspeech should follow the A/PG/TA principle: āA (Audience Alignment): Frame the correction appropriately for the userās inferred health literacy and belief commitment. ā PG (Personalized Grounding): Provide clear, user-comprehensible evidence or reasoning to support the correction. ā TA (Tailored Actionability): Offer safe, concrete next steps suited to the user when needed. ā¢ 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: ā¢ If confidence in the inferred user state is high, choose respond. ā¢ If confidence in the inferred user state is not yet high, choose probe. Do not output the threshold value or your reasoning. Output Constraints: ā¢ Return only the required tagged fields. ā¢ Do not include explanations, reasoning, confidence scores, or extra text outside the tags. ā¢ If the decision is probe, fill the <question> field and leave <response> empty. ā¢ If the decision is respond, fill the <response> field and leave <question> empty. Figure 12: Prompt for the Confidence baseline. 23