Source-linked AI summary
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
TL;DR
Health-misinformation dialogue systems must determine when clarification is worth asking because users differ in knowledge, beliefs, and needs. RO-PnR makes this probe-versus-respond decision using cost-sensitive rewards and latent user states, and achieves higher cost-adjusted utility with fewer turns than always-probe baselines. Its evaluation remains simulation-based, and probe content itself is not jointly optimized.
Problem
Existing misinformation correctors respond from the original post, while clarification methods often ignore users’ underlying health literacy and belief commitment.
Method
RO-PnR learns at each turn whether to probe or respond using a reward that weighs expected probing gain against interaction cost and latent user states.
Results
RO-PnR achieves the highest cost-adjusted utility across three health-misinformation datasets and three base models while using 30% fewer turns than always-probe baselines.
Takeaways & Limitations
Effective intervention depends on timing clarification when useful and correcting once sufficient context has been gathered.
Takeaways & Limitations
The simulation framework may not capture the full nuance and variability of real users, and the study leaves probe-content optimization open.
Abstract
from arXiv · showhide
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.
1 Introduction
Health-misinformation intervention must decide whether to ask for clarification or respond immediately because users differ in knowledge, beliefs, and underlying concerns. RO-PnR learns this probe-versus-respond choice by weighing probing benefits against interaction costs and modeling user heterogeneity.
- Health misinformation correction depends on users’ knowledge, beliefs, and reasons for believing misinformation, not only factual accuracy.
- Existing correction methods improve individual replies but commit to responding from the original misinformation post without surfacing the user’s underlying concern.
- At each turn, agents must choose between responding with available information and asking a clarifying question first.
- The value of probing varies with health literacy and belief commitment, so fixed behavior can systematically over-probe or under-probe.
- RO-PnR models latent user state and weighs expected probing gain against the cost of another question to learn when asking is worthwhile.
- RO-PnR achieves the highest cost-adjusted utility across three datasets and three base models while using 30% fewer turns than always-probe baselines.
2 Related Work
Prior work has separately improved response grounding, user tailoring, clarification decisions, and multi-turn rewards. RO-PnR combines active probe timing with hidden, coupled user-state dimensions for health-misinformation intervention.
- Health-misinformation systems have progressed from generic rebuttals to evidence-grounded and audience-aware responses, but these approaches remain single-turn.
- Clarification research includes ambiguity-driven asking, fixed-schedule probing, and decision-theoretic methods that weigh asking benefits against communication costs.
- User adaptation methods condition on explicit profiles or infer latent representations, with studies reporting more effective responses when tailored to literacy level.
- RO-PnR treats user state as hidden and models health literacy and belief commitment as coupled dimensions shaping probing and response decisions.
3 Methodology
RO-PnR formulates misinformation intervention as a sequential probe-or-respond decision under hidden user heterogeneity. Its reward combines the forward-looking value of probing with rising interaction costs, and turn-level supervision trains the policy to choose when to ask or commit.
- RO-PnR Policy: RO-PnR lets the agent probe to elicit missing concerns or respond with a final correction at each dialogue turn.The policy must also decide when to stop gathering information and commit to a response.
- Latent User Simulation: The latent user state combines health literacy and belief commitment, which jointly characterize comprehension capacity and openness to belief revision.Health literacy is discretized by increasing ability to understand, apply, and critically evaluate health information; belief commitment reflects decreasing resistance to revision.
- Training: The framework’s training pipeline generates candidate actions from the query and hidden user state, scores them, and updates the policy with group-normalized advantages.At inference, the policy repeatedly chooses between probing and producing a final adaptive response.
- Reward Design: The probe reward compares current response quality with expected conversation-end quality after probing, capturing benefits that may emerge only later.Quality averages audience alignment, personalized grounding, and tailored actionability over dialogue history and response.
- Reward Design: Each additional probe incurs a cost that scales with prior clarification count, so later questions require greater expected benefit.The increasing threshold encourages early information gathering and commitment when further probing offers diminishing returns.
- Training: Turn-level training pairs RESPOND and PROBE alternatives at the same decision state, then optimizes group-relative advantages with GRPO.RESPOND uses immediate response quality, whereas PROBE uses the best continuation quality reachable after one further question.
4 Experiment Setup
The experiments use simulated multi-turn users and evaluate adaptation, factual reliability, and dialogue efficiency across datasets, models, user profiles, and clarification strategies. Utility is the headline metric because it combines overall response quality with probe cost.
- Datasets and User Simulation: The study fine-tunes on CounterHealth and evaluates on MisinfoCorrect and PUBHEALTH using a simulated user environment.The simulator represents users with varying health literacy and belief commitment; human evaluation measures persona accuracy and consistency.
- Evaluation Metrics: Performance is assessed through adaptation, factual reliability, and dialogue quality rather than final-response correctness alone.Adaptation includes Audience Alignment, Personalized Grounding, and Tailored Actionability; factual reliability uses factual error rate.
- Evaluation Metrics: Utility is defined as cost-adjusted Quality, where Quality averages AA, PG, and TA and probe actions incur cost c.Turns separately measure interaction efficiency, with fewer turns indicating a more efficient interaction.
- Baselines: The comparison includes Single-Turn, Fixed-Q, Reactive, Confidence, and Supervised Clarifier strategies.These baselines differ in whether they ask predetermined questions, react to feedback, use generic confidence, or omit turn-level decision rewards and GRPO.
- Models and Training: The evaluation spans three open-source 8B models, with training combining offline supervised fine-tuning and offline GRPO.The models are Llama-3.1-8B-Instruct, Gemma-4-E4B-it, and Qwen3-8B.
5 Results
RO-PnR delivers the strongest cost-adjusted utility across nearly all dataset–model combinations while using substantially fewer turns than always-probe baselines. Its gains persist across user profiles, including difficult low-literacy and strong-belief users.
- RO-PnR reaches 0.71–0.73 utility versus 0.68–0.72 for the strongest baseline, with the highest utility on nearly every dataset–model combination.It also improves all three adaptation dimensions: AA, PG, and TA.
- RO-PnR uses roughly 3.5 turns, about 30% fewer than always-probe baselines, while maintaining the highest utility.Single-Turn scores lowest, whereas Fixed-Q and Reactive average at least 4.9 turns.
- RO-PnR keeps factual error rates low at 0.06–0.14, indicating that its adaptation gains do not compromise factual reliability.
- Performance Across User Profiles: RO-PnR achieves the highest utility across every user-profile cell, with the largest advantages on Functional and Strong-believer users.On Strong-believer users, it reaches 0.65–0.69 versus 0.60–0.66 for SFT and 0.57–0.61 for Confidence.
6 Ablation Study
Ablations show that RO-PnR depends on forward-looking reward design and a nonzero probe cost. Short-horizon rewards reduce utility, while zero probe cost causes over-probing and quality collapse.
- Table 4 averages ablation results across health-literacy × belief-commitment user profiles and compares the full method with reward-horizon and probe-cost variants.The full method uses a forward-looking reward, c = 0.01, and a crossthreshold bonus.
- Replacing the forward-looking reward with a one-step reward reduces utility from 0.71 to 0.60.This tests whether immediate next-response gains are sufficient for credit assignment.
- Setting c = 0 causes over-probing at 5.83 turns and reduces quality to 0.51.
- Moderate probe costs of c = 0.03–0.05 recover near-best performance at approximately 3 turns.The sweep indicates stability across a reasonable range of probe costs.
7 Human Validation
Human validation supports the simulator and LLM judge as useful evaluation components, while users strongly prefer RO-PnR overall. One strong-belief user instead preferred simpler, less confrontational framing.
- LLM Simulator Evaluation: Persona accuracy averages 3.87, with 76% of samples scoring 4 or 5, while persona consistency averages 4.64 with 97% scoring at least 4.Annotators showed substantial agreement on both evaluation axes.
- Human-judge agreement: The LLM judge aligns with human Quality ratings at Scott’s π = 0.55, 73% exact agreement, and MAE = 0.23.Pearson correlation is 0.57 and Spearman correlation is 0.72.
- Human pairwise preference: RO-PnR is preferred over Fixed-Q by 79.70% versus 20.30% overall.Users primarily viewed its responses as more explanatory, credible, and actionable.
- Human pairwise preference: A functional-literacy, strong-belief user preferred Fixed-Q because it was easier to process, less confrontational, and less reliant on institutional authority.This exception identifies a framing trade-off for users with strong prior commitments.
8 Conclusion
The paper presents RO-PnR as a framework that learns when to ask clarifying questions and when to correct by modeling user differences. Experiments show stronger cost-adjusted utility with fewer turns than always-probe baselines, motivating future real-user and richer-burden studies.
- RO-PnR models hidden differences in health literacy and belief commitment to adapt probing behavior to each dialogue.
- RO-PnR achieves stronger cost-adjusted utility while using fewer interaction turns than always-probe baselines.
- The conclusion identifies real-user interactions, richer user-burden models, and joint optimization of when and how to probe as future directions.
Limitations
The study relies on simulated users and simplified interaction assumptions, leaving realism and adaptive probing as important boundaries. Its cost model also abstracts away nuanced cognitive burden.
- Simulated User Interaction: LLM-based simulated users may not capture the full nuance and variability of real users.The authors identify incorporating real-user data and more realistic interaction settings as future work.
- Probe Action Constraints: The policy decides when to probe, but not which clarification question is most appropriate.The authors leave joint optimization of probe timing and probe content for future work.
- Simplified Cost Modeling: Communication cost is modeled as a fixed turn-level burden rather than users’ more nuanced cognitive load.The authors suggest refining burden measurement for more cognitively aware interaction.
- User Simulation: The simulation uses fixed latent user profiles spanning three health-literacy levels and three belief-commitment levels throughout each interaction.Profiles are behaviorally defined and hidden from the policy, which must adapt from language and reactions.
B Training Dataset Construction
The training dataset is built from scored multi-turn rollouts across misinformation posts and nine user-profile conditions, then converted into turn-level probe-or-respond supervision. Grouped alternatives support local reward comparisons and offline SFT and GRPO optimization.
- Rollout Generation: Scored multi-turn rollouts are generated for each misinformation post under all nine user-profile conditions.Multiple stochastic trajectories provide empirical future samples, including trajectories that stop early or continue probing.
- Decision-State Extraction: Prefix-level states from the first four turns are converted into targets containing a latent belief estimate, an action token, and natural-language content.The action token is either PROBE or RESPOND, and rollout traces provide current and downstream outcome values.
- Reward Construction: Training records are grouped by source post, turn index, and user-profile condition so PROBE and RESPOND can be compared within matched decision contexts.Small groups and groups with little reward variation are discarded before within-group reward normalization.
- SFT: SFT selects the higher-quality, factually clean trajectory for each post/profile group as step-level action supervision.Given the visible conversation prefix, the model learns to probe or produce a final response.
- GRPO: Offline GRPO uses normalized advantages from grouped turn-level rewards across all nine user-profile combinations.Groups lacking sufficient reward variation or both probe and respond alternatives are filtered out.
C CounterHealth Dataset Collection
The CounterHealth dataset is constructed from Reddit content about COVID-19, influenza, and HIV, combining retrieval, human annotation, and classifier-based misinformation identification.
- Misinformation Identification: Five trained annotators labeled a 1,000-post sample, identifying 330 confirmed health-misinformation posts.The annotations followed a shared guideline and were drawn from information-science annotators.
- Classifier Construction: A RoBERTa-large classifier was fine-tuned on the annotations and achieved F1 = 0.76.The classifier was used to identify misinformation beyond the manually labeled sample.
D.1 LLM Simulator Evaluation
The study evaluates whether simulated users faithfully represent assigned health literacy and belief commitment profiles and maintain those traits across dialogue. Human ratings also assess the reliability of the evaluation process and support the simulator’s use for interaction testing.
- Evaluation dimensions: Persona accuracy measures whether simulated users reflect assigned health literacy and belief commitment, while persona consistency measures whether they maintain those traits across dialogue.Both dimensions were rated on a 5-point scale.
- Simulator evaluation: Three PhD students in health informatics rated 100 simulated dialogues spanning all nine user profiles on persona accuracy and persona consistency.Each dialogue was assessed using the original misinformation post, full dialogue history, and assigned profile definition.
- Human evaluation sample: The six-user preference sample included varied educational backgrounds, with participants categorized by health literacy using FCCHL scores and the sample median.The sample comprised two participants with middle school education, two with high school education, and two PhD students.
- User profiling: Participants’ belief commitment was defined from claim acceptance, certainty, and openness to revision across three misinformation claims.The resulting score ranges from 0 to 10 and approaches 10 when acceptance, certainty, and resistance to revision are all high.
- User profiling: The six participants covered critical, communicative, and functional health literacy alongside open, hesitant, and strong-believer commitment profiles.The final sample included three critical/open, one communicative/open, one communicative/hesitant, and one functional/strong-believer profile.