When a language model turns a research paper into a social-media post, it simplifies. Some simplifications are harmless. Others are loans against the truth: "proves" in place of "suggests," a headline number with no mention of the benchmark it came from, a method called "general" when it was tested in one domain. A reader who asks a follow-up question calls in the loan. The system then has two options, and both are bad. It can contradict its own post, or it can defend the simplification and pile further errors on top of it. Current systems grade the post once, in isolation, so they never see this cost. Tools that turn papers into posts already treat factual fidelity as their main weakness (AutoPR), and models routinely turn tentative scientific findings into absolute claims (Tang et al.). Yet fidelity is still scored on a single static output.
We propose to score and train the post by what happens after it is published. For each candidate post, a simulated audience that sees only the post asks follow-up questions. A responder that rereads the paper answers them. The post is then penalized whenever a faithful answer would have to walk back what the post said. This thread-level fidelity signal is combined with engagement in a multi-objective reward and used to fine-tune the post generator. The claim is that thread-level fidelity catches a class of errors that single-turn fidelity judges miss: claims that are defensible on the surface but break under questioning. The claim also says that penalizing those errors costs little engagement, because the useful hooks and the misleading ones are mostly different sentences.
Core Claim
Single-turn fidelity scoring rewards "simplification debt": compressions that a one-shot judge accepts but that a faithful follow-up answer must contradict. Training against a thread-level reward that charges this debt back to the original post should move the trade-off between engagement and fidelity outward. Two things count as moving it outward. Posts trained this way should keep engagement while sharply lowering the rate at which faithful replies contradict them. Readers who see only the post should hold fewer misconceptions about the paper.
The interaction is deliberately multi-turn, but the thing we optimize is the turn-one artifact. This reverses the usual multi-turn hallucination fix. Work on hallucination "snowballing" (an early error compounding as it is fed back as context) shows that models favor consistency with the conversation over the evidence. This holds on MM-Snowball, where LLaVA-1.5-7B accuracy falls below 2% in turns 3–5, and on MMHalSnowball, where open models lose at least 31% under hallucinatory history. Those papers repair later turns at decoding time. We instead remove the seed of the snowball from the first turn. We treat the post as the earliest and most consequential turn of a conversation and train it with the conversation in view.
Research Program
Let $d$ be a paper and $p$ a candidate post. A reader model $A$ sees only $p$ and samples $k$ follow-up questions $q_1,\dots,q_k$. These include ordinary clarifications and presupposing questions such as "so this works for any model?" A responder $G$ answers each question using the paper $d$ and the question alone, with the post removed from its context. This history-free design copies the history-free stream that RVD blends in to resist snowballing, applied here to text. A judge then labels each pair $(p, r_i)$ as consistent, as requiring a retraction or qualification of $p$, or as unanswerable from $d$. The reward is
$$R(p) = E(p) + \lambda\, S(p) + \mu\, \frac{1}{k}\sum_{i=1}^{k} C(p, r_i),$$
where $E$ is predicted engagement, $S$ is a single-turn claim-level fidelity score of $p$ against $d$, $C(p,r_i)\in\{0,1\}$ marks that reply $r_i$ does not force a retraction of $p$, and $\lambda,\mu$ are trade-off weights. We sweep $\lambda$ and $\mu$ to trace a trade-off curve; we do not claim one best setting.
This reward has a probing structure that is close in spirit to the indirect-query finding of Do Language Models Know When They're Hallucinating References?. In that work, a fabricated citation gives inconsistent answers to repeated questions about its details, so the inconsistency itself flags it. Here, an overclaim gives answers that contradict the post under follow-up, so questioning exposes what one reading does not.
The work has three stages. The first is diagnostic. On PRBench papers, we generate posts with direct prompting and with a PRAgent-style multi-agent pipeline, run the probing loop, and measure how often faithful replies force retractions. We stratify by error type: certainty inflation, dropped scope conditions, and numbers stripped of their context. We validate the judge on a human-annotated subset in which domain experts mark retraction-forcing replies.
The second stage trains a 7–8B open model as the post generator with policy-gradient fine-tuning on $R$. The responder and judge are kept frozen. As a variant, we schedule the objectives with the bandit curriculum of DORB. DORB documents gradient domination under static weighting, and MARIO shows a single factuality reward collapsing into repetitive phrasing. Both results constrain our design, so diversity and engagement stay in the objective rather than being treated as afterthoughts.
The third stage tests reward hacking against the probe. One hack is to make posts so vague that no question can force a retraction. We monitor it with engagement, with an information-content check (whether core results and their scope are stated at all), and with a false-refusal analogue modeled on HalluLens. That benchmark shows that tuning hard against hallucination can produce heavy over-refusal: Llama-3.1-405B falsely refuses 56.77% of PreciseWikiQA queries.
This work differs from prior systems in specific ways. AutoPR builds a feed-forward multi-agent pipeline scored once on Fidelity, Engagement, and Alignment, and does no training against the downstream conversation. Engagement-driven persona prompting uses predicted engagement feedback in prompting and finds that LLMs make shallow stylistic edits. Neither looks past turn one. Co-opNet reranks summaries with discourse and salient-span factuality discriminators, which is our strongest single-turn analogue, but it never asks what the summary implies under questioning. Themis gives a reward model tools to verify claims. Our probe-and-reread loop is a verification tool aimed specifically at the post's implications. The steerability studies (Measuring Steerability, A Course Correction) motivate the work: pushing style moves other attributes as a side effect, and single-turn evaluation cannot show how those side effects pile up. We test whether one such side effect, lost fidelity under a casual, engaging style, appears only under follow-up.
Baselines
The naive reference point is direct prompting of the same base model. We use a strong instruction to write an engaging, faithful post for the target platform, with no training and no reranking. This anchors how common simplification debt is when nothing is done about it.
The strongest existing system is PRAgent from AutoPR, re-implemented on the same base model and PRBench split. It is the strongest competitor because it is the only system in the literature built for this exact task, and it already beats direct prompting by a wide margin. In RedNote trials it reports a 604% increase in total watch time and a 438% rise in likes over direct prompting, and annotators preferred its posts to human-written ones 64.8% of the time. No numeric PRBench Fidelity score is reported on Lacuna, so we will re-score PRAgent ourselves with the same judges used for every system. Beating it means equal or better PRBench Engagement and Alignment, together with a statistically significant reduction in retraction-forcing replies and in reader misconceptions.
The strongest single-turn competitor, and the most direct threat, is Best-of-N reranking with a claim-level fidelity judge in the style of Co-opNet. It uses the same $N$ samples per paper as our method's rollouts and the same judge model used for $S$. Matching total judge calls makes the compute comparable. Its RL counterpart is the same policy-gradient training with reward $E + \lambda S$ only ($\mu = 0$), tuned with the same hyperparameter budget and with or without DORB scheduling. This $\mu = 0$ run is the key ablation, because it removes exactly the thread-level term. A second ablation keeps the thread term but replaces post-conditioned reader questions with generic questions written from the paper. It tests whether it matters that questions are generated from what the post invites.
Finally, a reply-time fix pairs any baseline post with the re-grounded responder. It is compared on thread consistency and on misconceptions among readers who never see a reply.
Why not just…?
Why not just use a better single-turn fidelity judge that flags overclaims directly? This is the most serious alternative. A strong judge told to look for certainty inflation and dropped scope conditions might catch most of the debt without any probing. The $\mu = 0$ ablation and the Best-of-N baseline exist to settle this. If posts trained or reranked only on $S$ match our retraction rate and misconception rate, the thread term adds nothing, and we will report that. Our hypothesis is that the judge reliably catches false statements but misses true-but-incomplete ones, such as a correct number with its benchmark omitted. The stage-one stratified analysis tests that hypothesis before any training is spent.
Why not just re-ground each reply at reply time and leave the post alone? Re-grounded replies are already part of our pipeline, so this alternative partly coincides with our method. What it cannot do is repair the post. Most readers never ask a follow-up and never see a reply. Those who do ask see a thread in which the system contradicts itself. The snowballing literature also shows that conversational history pulls replies toward the post unless they are deliberately isolated from it (MM-Snowball). The reader-misconception test measures the remaining harm. That test has readers answer a quiz about the paper after seeing the post only.
Why not just hedge everything, or use long chain-of-thought? Uniform hedging is the degenerate optimum of a fidelity-only reward. The over-refusal pattern on HalluLens suggests it will cost engagement and information, and we measure both. Long reasoning before writing did not consistently help in AutoPR. It caused "specification drift," in which extra details crept into the post and blurred its message. We include a chain-of-thought variant of direct prompting so the reader can see this effect on our metrics.
Proposed Experiments
The primary testbed is PRBench: 512 papers paired with human-written posts for X and RedNote. We split it by paper into training and test sets. We report PRBench Fidelity, Engagement, and Alignment alongside four measures of our own. The first is the retraction rate: the share of reader questions whose faithful answer forces a qualification of the post. The second is the presupposition-acceptance rate: how often the responder, when the post is put back in its context, defends an overclaim. This is the textual analogue of the flip rates on MMHalSnowball. The third is information content. The fourth is reader misconceptions, measured with a quiz written by the paper's authors or by domain experts and given to readers who see only the post. To check that judging the conversation is reliable outside our own setup, we validate the consistency judge on MHB's multi-turn dialogues, which come with physician-validated rubrics. Claude-4-Sonnet's 29.1% hallucination rate there serves as a reference point for how hard such dialogue is.
Three results decide the case. First, the diagnostic must show that simplification debt is common in PRAgent and direct-prompting posts. It must also show that a meaningful share of it escapes the single-turn judge. If retraction-forcing posts are rare, or the judge already flags them, the project should stop at that finding. Second, on held-out papers, the trained generator must improve on the $\mu = 0$ and Best-of-N curves. Concretely, at matched Engagement it must have a lower retraction rate and fewer reader misconceptions. Third, we test whether the probe has been hacked. Gains must survive a reader model from a different model family than the one used in training, and information content must not fall. If the improvement disappears under a new questioner, the model has learned that questioner's habits, not to avoid overclaiming.
A small live deployment on RedNote or X follows the AutoPR trial design. Its only purpose is to check that engagement holds outside the predictor. The main conclusions rest on the controlled comparisons.
References
- AutoPR: Let's Automate Your Academic Promotion!
- Evaluating large language models on medical evidence summarization
- MM-Snowball: Evaluating and Mitigating Hallucination Snowballing in Multimodal Multi-Turn Dialogue
- Investigating and Mitigating the Multimodal Hallucination Snowballing in Large Vision-Language Models
- Do Language Models Know When They're Hallucinating References?
- DORB: Dynamically Optimizing Multiple Rewards with Bandits
- Tailoring Self-Rationalizers with Multi-Reward Distillation
- HalluLens: LLM Hallucination Benchmark
- Engagement-driven Persona Prompting for Rewriting News Tweets
- Cooperative Generator-Discriminator Networks for Abstractive Summarization with Narrative Flow
- Tool-Augmented Reward Modeling
- Measuring Steerability in Large Language Models
- A Course Correction in Steerability Evaluation: Revealing Miscalibration and Side Effects in LLMs
- MHB: Medical Hallucination Benchmark for Large Language Models in Complex Clinical Tasks