A new AI content-generation system produced much larger gains in a model’s predicted social-media influence than a competing method on two Twitter/X audience datasets, according to an arXiv preprint. Diffusion Gain—the relative increase in predicted influence over the input—was 16.51% for DOCG-AS on Movie, compared with 4.49% for Designed2Spread. On SpaceX, the figures were 40.70% and 12.01%.
Those figures describe a model forecast, not a tally of real-world reposts. The influence indicator was learned from observed retweet behavior, but the study’s main outcome was the relative increase in predicted diffusion influence over the input content.
How the system works
DOCG-AS first optimizes a hidden propagation feature, using a model trained on observed retweet behavior while discouraging the feature from drifting outside a realistic range. It then decodes that feature into eight audience-susceptible dimensions covering theme, structure, rhetoric and emotion for text and images, generates multiple multimodal candidates and selects the candidate with the highest predicted influence.
The evaluation used constructed Movie and SpaceX interest datasets. The Movie set contained 124,336 users and 2,910,600 tweets, including 1,699,675 with images. The SpaceX set contained 89,663 users and 2,217,930 tweets, including 620,025 with images.
GPT-5 generated 200 evaluation advertisements. Designed2Spread used 160 for training, while 40 formed a shared test set. The decoder corpus contained 75,000 annotated posts, and the system used six starting points for Movie generation and 10 for SpaceX generation.
The study compared DOCG-AS with four baseline methods under shared text-generation and image-generation models and identical inference settings.
The gain came with a trade-off
The benchmark also tracked Consistency, a cosine-similarity measure of how closely generated and input feature representations match. DOCG-AS scored 0.7783 on Movie and 0.7756 on SpaceX, compared with 0.8648 and 0.8580 for Designed2Spread.
The higher predicted diffusion score therefore came with less similarity to the starting material in these tests. The paper also gives 0.8722 as an example Movie Consistency score for LLM-Zero-shot, above DOCG-AS’s 0.7783.
The table values are reported as means ± standard deviations over three runs, so the ± figures describe run-to-run variation in the benchmark.
People preferred the rewrites in a small test
A separate study asked 50 people who actively engage with movie-related Twitter/X content to compare generated advertisements. Participants assessed about 100 randomized pairs each. The mean participant-level Repost Preference Rate was 64.27%, above the 50% chance threshold; the authors report a one-sided one-sample t-test with p = 6.94 × 10−7.
That result measures stated preference, not whether participants actually shared the advertisements. It also covers the Movie audience only, so the supplied evidence does not show that the preference would persist for SpaceX or other audiences.
What the study does not show
Removing parts of the pipeline reduced the reported Diffusion Gain on both datasets. Without feature decoding, the gain fell to 7.96% on Movie and 23.84% on SpaceX; removing multi-start generation reduced it to 11.57% and 34.00%, respectively. Removing vMF regularization or feature optimization also lowered the gains.
The evidence is limited to predicted influence on two interest-defined Twitter/X audiences and GPT-5-generated advertisements, plus the Movie preference study. It does not establish that DOCG-AS causes higher reposting, commenting, sharing or wider information spread, and it does not establish safety, fairness or the absence of manipulative content.
The paper is an arXiv version 1 preprint dated 20 August 2026.
Paper data and sources
Original title: From Latent Influence to Language: Diffusion-Oriented Content Generation via Audience-Susceptible Features
Authors: Jiaying Lei, Shengqi Dang, Runqian Bai et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-20
DOI: Not available
Original paper · Full text