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
EMNLP 2026
cs.AI
2026-08-22
RO-PnR asks only when long-horizon gain beats probe cost, topping utility on 3 health-misinfo datasets and 3 8B models at ~3.5 turns, ~30% fewer than always-probe.
Correcting health misinformation in dialogue is more than emitting a factually correct rebuttal. Users differ in two ways that matter: whether they can follow the evidence (health literacy), and how tightly they hold the misconception (belief commitment). Prior systems sit at two extremes. They either answer from the original post alone, or they ask a fixed battery of questions. Answering immediately misses the real concern. Asking on every turn burns attention even when the context is already enough.
Whether one more question is worth it depends on the quality of the eventual correction, weighed against the cost of asking. That trade-off is user-dependent. A low-literacy user who is willing to revise may need a single well-aimed probe. A high-literacy user who already stated the concern may need none. A strongly committed user may need the probe precisely so the later correction does not land as a confrontation.
RO-PnR (Reward-Optimized Probe-and-Respond) treats each turn as a binary choice: Probe (ask one more question) or Respond (commit to a final correction). The policy sees only the dialogue history. Profile labels are hidden from it; they only drive how the simulated user replies.
The latent user state is a \(3\times 3\) grid: health literacy {functional, interactive, critical} crossed with belief commitment {strong, hesitant, open}. The simulator is GPT-4o-mini, constrained to short in-character replies.
The reward has two layers. Respond scores the current reply quality \(R(ht)\), the mean of Audience Alignment (does the framing match inferred literacy and stance), Personalized Grounding (is the evidence intelligible), and Tailored Actionability (are next steps concrete and safe). Probe scores expected quality at the end of the conversation after asking, minus a cost \(c\cdot nt\) that grows with how many questions have already been asked. Default \(c=0.01\). The agent asks only when the long-horizon gain exceeds cumulative interaction cost, and later questions face a higher bar.
Training starts with SFT on the highest-quality, factually clean trajectory in each post-by-profile group (LoRA \(r=16\), \(\alpha=32\), 4 epochs), then offline GRPO on turn-level Probe/Respond pairs (2 epochs, \(\beta{\mathrm{KL}}=0.05\)). GRPO compares relative advantages inside a decision group, so no separate value network is required.
Evaluation uses three 8B models: Llama-3.1-8B-Instruct, Qwen3-8B, and Gemma-4-E4B-it. Fine-tuning data is CounterHealth, 769 Reddit posts on COVID, influenza, and HIV after annotation and filtering. MisinfoCorrect and PUBHEALTH are evaluation-only. Adaptation is judged by mistral-large-2512; factual error rate by GPT-5-mini with web verification.
Baselines: single-turn reply, fixed-question probing, reactive clarification from user feedback, confidence-gated asking, and an SFT clarifier without the turn-level decision reward.
The headline metric is cost-adjusted Utility. RO-PnR is highest on almost every dataset-by-model cell, at 0.71–0.73, against 0.68–0.72 for the strongest baseline (SFT). It uses about 3.5 turns, roughly 30% fewer than always-probe Fixed-Q / Reactive (about 4.9–5.0). Single-turn is worst (0.53–0.64). Confidence-gated asking barely questions (1.1–2.0 turns) and drops quality the most, down to 0.58–0.62 Utility on MisinfoCorrect.
| Method | Quality | Turns | Utility |
| Single-turn | 0.53–0.64 | no probe | 0.53–0.64 |
| Fixed-Q / Reactive | 0.68–0.73 | \(\ge 4.9\) | 0.67–0.72 |
| Confidence | 0.58–0.74 | 1.1–2.0 | 0.58–0.74 |
| SFT | 0.70–0.72 | 3.9–4.6 | 0.69–0.72 |
| RO-PnR | 0.72–0.74 | 3.5 | 0.71–0.73 |
Gains concentrate on hard users. On functional literacy, RO-PnR scores 0.65–0.69 against SFT at 0.62–0.67. On strong believers, 0.65–0.69 against SFT 0.60–0.66 and Confidence 0.57–0.61. Open users are easy for everyone, and the gap shrinks to 0.76–0.77 vs 0.73–0.76.
Ablations are clean. Scoring only the next utterance drops Utility from 0.71 to 0.60. Setting \(c=0\) over-probes to 5.83 turns and collapses quality to 0.51. Moderate costs \(c=0.03\)–\(0.05\) recover near-best quality at about 3 turns.
On a small human study, simulated persona accuracy averages 3.87 (76% rated 4 or 5) and consistency 4.64. The LLM judge vs human Quality has Scott's \(\pi=0.55\), 73% exact agreement, MAE 0.23. In a 50-post pairwise test with six people, RO-PnR beats Fixed-Q 79.7% to 20.3%. The one reversal is a functional, strongly committed user who found Fixed-Q easier and less confrontational. Factual error rates stay low; RO-PnR sits at 0.06–0.14, so the adaptation gain is not paid for in invented facts.
The hard decision in multi-turn correction is often whether to ask at all, not how to polish the same fact. Casting probing as value of information with an explicit cost, then training that choice with GRPO at the turn level, is closer to a real conversation constraint than a fixed question list or a confidence gate.
The setting that might actually use this is a health help-desk that needs to surface the misconception before advising. An 8B model plus LoRA is enough, and the policy is not handed the user label. Almost all of the numbers come from a simulator, and Utility is itself an LLM judgment. Reading this as "ready to correct real users in production" overclaims what was measured. It is a decision-framework result.
The paper flags three limits: users are GPT-4o-mini role-play, with humans only on persona checks, judge agreement, and a 6-person preference study; the policy learns when to ask, not what to ask; cost is a fixed per-turn \(c\), not cognitive load.
A few more discounts are warranted. Profile labels are hidden from the policy but visible to the adaptation judge, which makes the "aligned with the user" scores easier. CounterHealth is Reddit vaccine and COVID posts, a narrow slice. Gemma's FER on PUBHEALTH reaches 0.14, worse than some baselines; the utility lead is mostly adaptation and fewer turns, not fewer falsehoods. The strongly committed user in the preference study picking Fixed-Q means the 79.7% aggregate does not cover the hardest cell.