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.
Figures & tables
Figure 1: Different strategies on the same misinformation post. RO-PnR weighs benefit vs. cost before acting.
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.
Dataset
Method
Llama-3.1-8B-Instruct
Qwen3-8B
Gemma-4-E4B-it
AA
PG
TA
Quality
FER ↓
Turns ↓
Utility
AA
PG
TA
Quality
FER ↓
Turns ↓
Utility
AA
PG
TA
Quality
FER ↓
Turns ↓
Utility
CounterHealth
Single-Turn
0.58
0.53
0.61
0.57
0.18
5.00
0.56
0.58
0.52
0.61
0.57
0.15
5.00
0.56
0.64
0.59
0.67
0.63
0.04
5.00
0.62
Fixed-Q
0.72
0.65
0.76
0.71
0.11
5.00
0.70
0.71
0.64
0.73
0.69
0.08
5.00
0.68
0.73
0.66
0.74
0.71
0.02
5.00
0.70
Reactive
0.72
0.64
0.76
0.70
0.11
5.00
0.69
0.71
0.64
0.72
0.69
0.09
5.00
0.68
0.72
0.65
0.74
0.71
0.02
3.92
0.70
Confidence
0.71
0.64
0.74
0.69
0.13
1.26
0.69
0.67
0.61
0.70
0.66
0.06
1.42
0.66
0.72
0.64
0.76
0.71
0.02
1.85
0.71
SFT
0.70
0.64
0.76
0.70
0.07
4.38
0.69
0.70
0.64
0.76
0.70
0.05
4.22
0.69
0.70
0.64
0.76
0.70
0.09
4.61
0.69
Table 1: Average results across all health-literacy and belief-commitment configurations, for each method, dataset, and base model. AA , PG , TA are the Audience Alignment (AA), Personalized Grounding(PG), and Tailored Actionability (TA) scores (normalized to [0,1] ); Quality is the overall quality of AA, PG and TA (normalized to [0,1] ); FER is the factual error rate ( ↓ ); Turns is the average number of dialogue turns ( ↓ ); and Utility is the cost-adjusted Quality at c=0.01 , our headline metric (normalized to [0,1] ).
CounterHealth
Misinfo
Pubhealth
Method
Model
F
I
C
Ov.
F
I
C
Q.
F
I
C
Ov.
Single-Turn
Llama
0.51
0.60
0.57
0.56
0.49
0.57
0.55
0.53
0.57
0.64
0.62
0.61
Qwen
0.50
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
Llama
0.66
0.72
0.72
0.70
0.62
0.71
0.70
0.68
0.67
0.73
0.73
0.71
Qwen
0.62
0.71
0.72
0.68
0.63
0.71
0.70
0.68
0.67
0.72
0.73
0.71
Table 2: Utility by Health Literacy Level (normalized to [0,1] ). Higher-value cells are highlighted with light tints: green ( ≥0.70 ), mint ( ≥0.73 ), and blue ( ≥0.75 ). F, I, C, Ov., denotes Functional, Interactive, Critical, and all populations.
CounterHealth
Misinfo
Pubhealth
Method
Model
S
H
O
Ov.
S
H
O
Ov.
S
H
O
Ov.
Single-Turn
Llama
0.50
0.54
0.64
0.56
0.48
0.52
0.61
0.53
0.55
0.57
0.71
0.61
Qwen
0.50
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
0.62
0.71
0.64
Fixed-Q
Llama
0.63
0.71
0.75
0.70
0.60
0.69
0.74
0.68
0.65
0.72
0.76
0.71
Qwen
0.61
0.71
0.73
0.68
0.60
0.70
0.73
0.68
0.65
0.72
0.75
0.71
Table 3: Utility by Belief Commitment. Higher-value cells are highlighted with light tints: green ( ≥0.70 ), mint ( ≥0.73 ), and blue ( ≥0.75 ); lower values are left uncolored. S, H, O, Ov., denotes Strong, Hesitant, Open, and all populations.
Variant
AA ↑
PG ↑
TA ↑
Quality ↑
FER ↓
Turns ↓
Utility ↑
RO-PnR (full)
0.72
0.66
0.77
0.72
0.06
3.56
0.71
(a) Reward-horizon ablation
immediate next-gain only
0.62
0.54
0.65
0.60
0.12
2.00
0.60
(b) Probe-cost ( c ) sweep
probe cost =0.00
0.50
0.46
0.57
0.51
0.11
5.83
0.51
probe cost =0.03
0.71
0.65
0.77
0.71
0.08
3.00
0.70
Table 4: Ablation study results, averaged over all (health literacy × belief commitment) user profiles. RO-PnR (full) is our complete method with forward-looking long-horizon reward, probe cost c=0.01 , and cross-threshold bonus. (a) replaces the forward-looking reward with the immediate next-turn gain; (b) sweeps the per-turn probe cost c .
Appendix figures & tables19 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 3: 3 × 3 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.
βKL=0.05 ; SFT CE weight 0.2 ; warmup ratio 0.03 ; max grad norm 1.0 ; seed 42
Appendix
Table 5: Hyperparameters for SFT and GRPO configuration
Score
Persona Accuracy
Persona Consistency
1
The 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.
2
The simulated user weakly matches the assigned profile. One dimension may be partially reflected, but the other is missing or incorrect.
The user shows unstable behavior, with multiple unexplained shifts in health literacy or belief commitment.
3
The simulated user partially matches the assigned profile. Both dimensions are somewhat recognizable, but the dialogue is vague, generic, or only weakly aligned.
The user is mostly stable, but there are noticeable inconsistencies or minor contradictions across turns.
4
The simulated user largely matches the assigned health literacy and belief commitment, with only minor imperfections.
The user maintains the assigned profile across most turns, and any change in belief strength is mostly justified by the conversation.
5
The simulated user strongly matches the assigned profile. Health literacy and belief commitment are both clearly and accurately expressed.
The user consistently maintains the assigned health literacy and belief commitment throughout the dialogue, with any change being natural and well-supported by the interaction.
Appendix
Table 6: Human evaluation rubric for assessing user simulation quality.
Metric
Persona Accuracy
Persona Consistency
Mean Pairwise Agreement
0.83
0.85
Mean Pairwise Scott’s π
0.70
0.68
All-Three Agreement
0.74
0.78
Multi-Rater Scott’s π
0.70
0.68
Appendix
Table 7: Inter-annotator agreement among three evaluators on persona accuracy and persona consistency . Both pairwise and multi-rater statistics indicate substantial agreement.
Metric
Score 1
Score 2
Score 3
Score 4
Score 5
Mean
Persona Accuracy
0%
3%
21%
62%
14%
3.87
Persona Consistency
0%
0%
3%
30%
67%
4.64
Appendix
Table 8: Distribution of human ratings for persona accuracy and persona consistency.
Dimension
Scott’s π
Mean Pairwise Agreement
All-three Exact Agreement
Audience Alignment
0.76
84.70%
78.00%
Personalized Grounding
0.63
76.00%
65.00%
Tailored Actionability
0.77
88.70%
83.00%
Quality
0.78
86.70%
81.00%
Appendix
Table 9: Human inter-annotator agreement under 0.5-point discretization.
Dim
Scott’s π
Percent Agreement
MAE
Bias
Within ±0.5
Within ±1.0
Audience Alignment
0.51
69.00%
0.29
-0.14
86.00%
94.00%
Personalized Grounding
0.31
51.00%
0.28
-0.14
85.00%
96.00%
Tailored Actionability
0.48
72.00%
0.14
-0.07
95.00%
99.00%
Quality
0.55
73.00%
0.23
-0.11
87.00%
97.00%
Appendix
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.
Dimension
Pearson
Spearman
LLM Mean
Human Mean
Audience Alignment
0.50
0.66
3.76
3.62
Personalized Grounding
0.51
0.67
3.50
3.37
Tailored Actionability
0.72
0.81
3.93
3.86
Quality
0.57
0.72
3.73
3.62
Appendix
Table 11: Correlation between the LLM judge and human consensus under 0.5-point discretization.
Item
Question
Rating
B
How accurate do you think this claim is?
0 = Definitely false / 10 = Definitely true
C
How certain are you about your answer above?
0 = Not at all / 10 = Completely
O
If 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
Appendix
Table 12: Belief Commitment Determinants. B denotes belief, C denotes certainty, and O denotes openness to change.
User
Total FCCHL Mean
HL Category
Overall BCtotal
BC Category
User 1
2.07
Communicative
2.03
Hesitant
User 2
1.93
Functional
6.83
Strong
User 3
2.71
Communicative
1.20
Open
User 4
3.00
Critical
1.20
Open
User 5
4.00
Critical
0.40
Open
User 6
3.64
Critical
0.35
Open
Appendix
Table 13: Health Literacy and Belief Commitment Categories. HL categories are mapped from the Total FCCHL mean: 1.00-1.99 = Functional, 2.00- 2.99 = Communicative, and 3.00-4.00 = Critical. BC categories are mapped from BCtotal : BCtotal<1.5 = Open, 1.5≤BCtotal<5 = Hesitant, and BCtotal≥5 = Strong.
Figure 4: Human pairwise preference comparison between RO-PnR and Fixed-Q across different users. Detailed users’ background are provided in Table 13 .
Figure 5: Questionnaire for health literacy screen
Figure 6: System prompt and user prompt for realistic health misinformation persona simulation.
Figure 7: Prompt for evaluating the Factual Error.
Figure 8: Evaluation rubric for Audience Alignment, Personalized Grounding, and Tailored Actionability.
Figure 9: System prompt for the Reactive baseline.
Figure 10: Prompt for the Reactive baseline.
Figure 11: System prompt for the Confidence baseline.
LLM decision-making studies how language models choose, act, seek information, and allocate resources under goals, uncertainty, constraints, and social pressures.
Large Language Model (LLM)-augmented Community Notes offer a scalable path for timely, evidence-grounded correction of health misinformation on social platforms. However, they still reset at every post, leaving useful correction experience from prior cases unused. We introduce EvoNote, an agentic framework that enables health Community Notes generation to self-evolve through an evolving experience memory of prior misinformation correction episodes. Its core is fine-grained credit assignment: EvoNote grounds trajectory-level feedback in health-specific note qualities and distills it into action-level memory for claim analysis, evidence acquisition, and note writing. We evaluate EvoNote on MM-HealthCN, a 1.2K-instance multimodal benchmark of user-flagged health posts with human-written Community Notes and crowd-derived helpfulness labels. Under a human-validated hierarchical utility judge, EvoNote-generated notes are preferred over corresponding human-written notes in 89.6% of cases; on a separate set of Needs More Ratings posts without a crowd helpfulness verdict, EvoNote produces helpful notes for 82.0% of cases. It also reduces time to an available candidate correction from multi-hour post-to-note delays to under 2 minutes on average. Analyses link these quality gains to stronger evidence use and reusable correction strategies, positioning self-evolving note generation as a promising paradigm for scalable health misinformation governance.
Zihang Fu, Fanxiao Li, Jianyang Gu +5
National University of Singapore · Yunnan University · The Ohio State University +1
Large language models (LLMs) have made substantial progress on medical question-answering, yet effective medical dialogue also requires learning to ask questions that uncover relevant patient information. To train such dialogue policies, a common pipeline combines supervised fine-tuning with reinforcement learning (RL) based on final diagnostic correctness. However, this outcome-based supervision does not directly distinguish the contributions of individual questions and provides no question-level feedback for unexecuted alternatives. To address this gap, we introduce PCQC (Privileged Counterfactual Question Credit), which uses privileged patient information during training to learn from questions never asked. During training, PCQC makes alternative questions directly comparable at the same dialogue state by using privileged patient facts to construct their answers. A frozen diagnostic scorer evaluates the diagnostic utility of each resulting question-answer pair by how strongly it supports the correct diagnosis. PCQC turns these comparisons into relative question credit that teaches the policy which questions to favor, directly supervising both executed and unexecuted questions alongside outcome-based RL without requiring complete rollouts for the unexecuted alternatives. Extensive experiments across four medical benchmarks demonstrate that PCQC achieves 63.10% mean diagnostic accuracy, outperforming GRPO and ATPO by 4.38 and 4.21 percentage points, respectively. These gains are achieved with 33.1% fewer inquiry turns than GRPO.
Chenxuan Li, Jiayi Wan, Xinrong Chen +3
Peking University, Beijing, China · Department of Medical Bioinformatics, School of Basic Medical Sciences, Peking University, Beijing, China
Misinformation harms public health and democracy, and generative artificial intelligence now writes falsehoods as persuasive as people's. Defences range from fact-checking individual claims to training people to resist falsehoods. Yet many claims are too new to check, and interventions that teach doubt can reduce belief in truth too. We propose a defence that targets not the facts of each claim but the persuasion strategies that recur across them. Here we show that learning misinformation's persuasion playbook, mapped from 214,646 fact-checked claims in 51 languages, can sharpen human and machine judgement of claims they know little about, without making either more cynical about true claims. The playbook spans social influence, information manipulation and emotional appeals, combined around a core of fear, conspiracy and fabrication. Briefed with it, plus a reminder that persuasion is not proof of falsehood, a language model better distinguished true from machine-generated false claims about events after its training data. In a pre-registered experiment, the same briefing improved people's discernment of claims they reported as unfamiliar and left their belief in true claims unchanged. Averaged over all claims, the gain was not statistically significant. Because it targets what recurs, this defence need not wait for claim-specific fact-checks.
Tong Guo, Teng Ye, Yutong Xie +1
School of Information and Department of Electrical Engineering and Computer Science, University of Michigan