Automated systems now turn a long document, such as a research paper, into short posts tailored to each social platform. Each post is optimized and judged on its own. Nobody checks whether the Twitter post and the RedNote post for the same paper say the same thing: the same headline result, the same numbers, the same caveats, in a recognizably shared voice. Marketing research says this matters. In a survey of fashion-brand customers, perceived consistency across touchpoints predicted affection for the brand, both directly and by making the brand seem more authentic (From Runway to Click-way). Meanwhile, the leading automated repurposing system names factual fidelity as its main unsolved weakness (AutoPR). This proposal treats the set of posts produced from one source as the object to evaluate and control.
The plan has two parts. First, define a set-level consistency measure and use it on an existing benchmark to find out whether today's per-platform pipelines actually contradict themselves across platforms. Second, test one simple fix, which borrows a lesson from cross-device software: store the invariant content once, as an explicit "ledger" of claims, numbers and voice rules, and generate every platform's post from that ledger, rather than re-deriving the message from the source separately for each platform.
Core Claim
The claim has two parts, and the first is the one that could be wrong.
The first part is empirical. Pipelines that adapt content to each platform independently will drift across platforms. They will pick different headline claims, report numbers inconsistently, and drop caveats in some versions but not others. This drift will not show up in per-post fidelity scores, because each post can be loosely faithful to the source while the posts still disagree with each other.
The second part is about method. A shared, explicit ledger of invariants is extracted once and verified against every output. It should reduce cross-post inconsistency without giving up the platform-native adaptation that drives engagement.
The second part only matters if the first holds. It also only matters if the obvious cheaper fix, generating all platform posts in one prompt so the model sees them together, fails to solve consistency at a competitive engagement level.
Positioning Against Prior Work
The closest work is AutoPR's PRAgent pipeline (AutoPR). It splits a paper into text hierarchies and figures. Four specialized agents then draft content, and a final orchestration agent adds platform conventions for Twitter or RedNote. It supplies our strongest baseline and our main benchmark, PRBench. PRBench contains 512 papers paired with human-written promotional posts and is scored on three axes: fidelity, engagement, and alignment with the platform.
PRAgent produces posts one platform at a time and scores them in isolation. We keep its extraction stage but change two things: what is shared across platforms, and what is evaluated.
Persona-prompted rewriting of Dutch news tweets (Engagement-driven Persona Prompting) motivates the voice side of the ledger. That work found that casual rewrites, which scored highest on predicted engagement, had significantly lower human-rated style adherence than formal or factual rewrites. In other words, chasing engagement erodes a consistent house voice. That study covered one platform and one post at a time.
The Pop Mart case study (Bao & Gao) shows that real brands deliberately assign different roles to Weibo, Douyin, Xiaohongshu and Bilibili while keeping one brand identity. It also reports homogenization fatigue across influencers. It is descriptive and offers no mechanism or measurement, so it motivates the problem but cannot be a baseline.
The "store state once, render per device" idea comes from systems work. That work kept user-interface layouts and speech settings in a cloud store and pushed them to phones, tablets and PCs (Park et al.; E-commu-book). Those systems synchronize structured, deterministic settings. Our ledger applies the same separation to unstructured claims, and it needs a verifier because a language model, unlike a renderer, can silently deviate from the stored state.
Two papers keep a model's persona stable: persona-aware contrastive training (PCL) and inference-time steering with activation vectors, meaning directions added to the model's internal activations (PERSONA). Both work on personality in dialogue, not on factual claims. We explain below why we do not make steering the core mechanism.
Research Program
The first deliverable is a set-level consistency measure. For a source document with platform posts $p_1, \dots, p_m$, we extract the atomic claims in each post, meaning headline result, numbers, method description and limitations. We then compute
$$C = \frac{1}{\binom{m}{2}} \sum_{i<j} \mathrm{contra}(p_i, p_j) \;+\; \lambda \cdot \frac{1}{m}\sum_{i} \big(1 - \mathrm{cov}(p_i, K)\big).$$
The symbols are as follows. $\mathrm{contra}(p_i, p_j)$ is the fraction of claim pairs, matched by topic across two posts, that a natural-language-inference judge labels as contradictory. Numeric mismatches are counted deterministically. $K$ is a small set of core claims that every post must carry, which human annotators define per paper on a subset. $\mathrm{cov}(p_i, K)$ is the fraction of $K$ that post $p_i$ expresses. $\lambda$ weights omission against contradiction; we report both terms separately as well as combined.
Voice consistency is measured separately. A judge model rates each post against the same written list of voice attributes, and we compute the spread of those ratings across platforms. The judges are validated against human labels on a held-out sample before any comparison is reported.
The second deliverable is the ledger pipeline. An extraction pass produces a structured record from the paper. It contains a headline claim, typed numbers with units and precision, required caveats, figure references, and three to five voice attributes. A human can edit this record. Platform renderers, which keep the same agent structure and platform conventions as PRAgent, receive the ledger together with a platform brief, not the raw paper alone. Each renderer must mark which ledger entries it uses.
A verifier then checks each post against the ledger. It looks for unmarked numbers, contradicted entries and missing core claims, and returns violations for at most two revision rounds. The verifier compares against the ledger, not against other posts. That keeps the cost linear in the number of platforms and makes the ledger the single source of truth.
Baselines
The naive reference point is the direct single-post prompt that AutoPR compares against: the paper plus "write a post for platform X," run independently per platform. We also report PRBench's human-authored posts on the per-post axes, while noting that human posts may not exist for every platform.
The strongest existing method is PRAgent (AutoPR). It reports that annotators preferred its posts over human-authored ones 64.8% of the time. In a field test on RedNote it reports a 604% increase in watch time and a 438% rise in likes over direct prompting. We have no reported cross-platform consistency number for it; no reported result was found on Lacuna for any set-level consistency metric on PRBench. We will re-run PRAgent with the same backbone model, the same papers, and a matched token budget, so that revision rounds in our method are counted against it.
The joint baseline is a single prompt that writes all platform posts at once with a brand style guide included. It is cheap and directly targets consistency, so it is the most dangerous competitor. It gets the same backbone and budget.
A source-verifier ablation keeps PRAgent's per-platform generation but adds a verifier that checks each post against the full paper rather than a ledger. This isolates the key ingredient. If it matches the ledger pipeline on $C$, then the shared structured state adds nothing beyond verification.
A no-verifier ablation renders from the ledger but skips the checking loop. This separates the value of conditioning on the ledger from the value of enforcing it.
Beating PRAgent means three things. The pipeline must lower $C$ by a margin larger than the variation across random seeds. It must do so while staying within PRAgent's per-post confidence interval on PRBench engagement and alignment. It must also not lose to PRAgent in blinded pairwise preference.
Why not just…?
Why not just generate all platform posts in one prompt? Seeing every post at once should, in principle, keep them consistent. This is a real threat. On the other side, AutoPR's ablations found that single-prompt pipelines lose information and follow platform norms poorly, which predicts lower engagement and alignment scores. Experiment one settles the question. If the joint prompt reaches the ledger pipeline's $C$ at comparable PRBench engagement and alignment, we drop the generation method as a contribution. The work then becomes the consistency measure and the finding that joint generation is sufficient.
Why not just fact-check each post against the paper? If every post is faithful, contradictions should be rare. But faithfulness is not agreement. Two faithful posts can still lead with different results, round a number differently, or include the caveat in only one of them. The source-verifier ablation tests this directly, and $C$'s coverage term is designed to catch differences in emphasis that per-post fidelity misses.
Why not just steer a brand-voice vector? PERSONA reports scores comparable to fine-tuning on PersonalityBench (9.60 vs. 9.61) at about 5.5% extra latency (PERSONA), so steering voice is attractive. However, the documented failure in repurposing is about claims and numbers (AutoPR), and a personality direction does not carry facts. We include steering only as an optional add-on for the voice-spread metric, never as a replacement for the ledger.
Why not just maximize engagement and ignore consistency? Pillai et al. show that engagement-driven casual rewriting lowers style adherence. Ahmad et al. show that consistency predicts brand love (direct path β = 0.244, plus β = 0.157 through perceived authenticity). However, that evidence comes from a survey about one apparel brand, not from readers of generated posts. Our human study tests whether set inconsistency actually lowers trust in this setting. If it does not, the practical case weakens and we will say so.
Proposed Experiments
Experiment one runs all five systems on PRBench papers. Each system produces posts for Twitter, RedNote, and a long-form thread format. We report $C$ and its two components, voice spread, and PRBench fidelity, engagement and alignment. The decisive check comes first: do PRAgent and direct prompting show measurable cross-post contradiction and omission of core claims at all? If their contradiction rate is near the judge's false-positive rate, which we measure on post pairs that are identical in meaning, the problem is too small to matter and we stop.
Experiment two tests transfer of the voice side on the Dutch news tweet corpus (Engagement-driven Persona Prompting). The formal, factual and casual registers act as stand-in channels for one newsroom voice. We compare voice spread and content preservation with and without the ledger's voice attributes. The reference figures are the paper's human ratings: an average of 4.335 for content preservation and 4.135 for style adherence.
Experiment three is a preregistered human study. Participants see a simulated audience journey: two or three posts about the same paper, as they would appear on different platforms. The posts come either from the most consistent or the least consistent system, matched on per-post quality ratings. Participants rate credibility and whether they would read the paper, and they flag any disagreements they notice. This is the test that connects the computed measure $C$ to the authenticity effect reported in the marketing survey.
References
- From Runway to Click-way: Unraveling the Power of Authenticity in Digital Marketing
- AutoPR: Let's Automate Your Academic Promotion!
- Engagement-driven Persona Prompting for Rewriting News Tweets
- Research on the Dissemination Phenomenon of Pop Mart Co-branded IP Products on Social Media
- An Efficient Interface Providing Methodology Applying Multi-Sensors in Mobile Cloud Platform
- E-commu-book: an assistive technology for users with speech impairments
- Enhancing Persona Consistency for LLMs' Role-Playing using Persona-Aware Contrastive Learning
- PERSONA: Dynamic and Compositional Inference-Time Personality Control via Activation Vector Algebra