slopster
Log inGet started

Research basis

What the engine is built on.

Findings gathered through the Consensus API on 2 September 2026 that shape Slopster. 31 papers, grouped by the question they answer. Numbers in brackets refer to the list at the end.

What makes short-form content perform

  • Curiosity and pain-point hooks raise engagement over direct informational openings, holding video length, subtitles and face visibility constant. [1]

    Mechanism. Hook style is a first-class attribute. Generators must cover the taxonomy; judges score hook arousal explicitly.

  • Follower count is the strongest virality predictor on TikTok, but close-up/medium shot scale, on-screen text and point of view also carry signal. [2]

    Mechanism. Shot lists specify framing; every script carries on-screen text beats.

  • Number of shots and image complexity have inverted-U effects on engagement; vertical format raises comments and shares; faster speech rate raises shares. [3]

    Mechanism. The pacing planner targets a shots-per-second band, word counts are fitted to a speech rate, and all video is 9:16.

  • Humor and emotion raise engagement; informative content lowers it alone but raises it when combined with personality. Narrative content has the strongest effect on purchase intent. [9] [10] [11]

    Mechanism. The judge rubric scores a personality × information blend and narrative pull. Sales briefs are rewritten to carry both.

  • Muted viewing on mobile barely reduces ad effectiveness when visual storytelling carries the message. Shorter, portrait videos performed best in a 3.3M-impression field test. [19] [20]

    Mechanism. On-screen text is mandatory in every beat; packaging defaults to short vertical cuts.

  • Posting consistency relates positively to follower engagement; post frequency alone does not grow followers. [16] [17]

    Mechanism. Campaigns define a cadence and the scheduler fills it with the best available candidates instead of posting everything.

How to generate and judge with language models

  • Multi-agent debate improves factuality and reasoning over single-model output. Self-reflection suffers a degeneration of thought that a judge-adjudicated debate fixes. [4] [5]

    Mechanism. Refinement is an author-versus-critic debate, adjudicated by a judge from a third model family.

  • Heterogeneous role-specialized agents with learned consensus beat homogeneous debate by 4 to 6 points and cut factual errors by over 30%. [6]

    Mechanism. Generation uses role-specialized authors (hook-smith, storyteller, direct-response, contrarian) across model families.

  • Strong judges reach about 80% agreement with humans but show position, verbosity and self-enhancement bias. Self-preference is driven by perplexity: models rate familiar text higher. [7] [8] [12]

    Mechanism. Judges never share a family with the author they score. Every pairing is judged in both positions and disagreements are discounted.

  • Generative AI improves individual creativity but collapses collective diversity. [13] [14] [15]

    Mechanism. Embedding-distance constraints at the angle and population stages; near-duplicates are culled before judging.

  • AI-generated imagery beat human stock imagery on click-through by up to 50% in a 173k-impression field study. [21]

    Mechanism. Finalists are packaged with generated imagery, voiceover and optional video.

Optimization and learning

  • Batched Thompson sampling beats test-then-rollout for headlines by 3.69% clicks. Batch-updated bandits handle delayed rewards and delivered +6% CTR and +16% CVR in production. [22] [23] [24]

    Mechanism. Each brand keeps Beta priors per attribute value, updated in batches from real metrics and sampled during generation.

  • Pre-posting virality screening reaches Precision@1% of 0.75 even though exact count prediction is not feasible. [25]

    Mechanism. Slopster never predicts view counts. It ranks, then calibrates judge weights against realized performance per brand.

  • Multi-source burst detection gives about 36 hours of lead over native trending lists. Early detection is feasible at about 75% AUC before trending. [26] [27]

    Mechanism. A trend scanner runs on a schedule, scores topics by velocity, and injects relevant signals into briefs.

Trust and disclosure

  • AI disclosure lowers ad credibility and intent, and the penalty is worst for high-involvement products. Human-likeness moderates the effect and persuasion knowledge can partially restore trust. [28] [29] [30]

    Mechanism. Output is written to sound like a person with a named brand voice. Disclosure is a per-workspace setting so customers can comply with platform policy.

Papers

  1. 1.Phyoe (2026). The Impact of Opening Hook Style on Engagement Rate in Short-form Motivational Videos on TikTok
  2. 2.Ling et al. (2021). Slapping Cats, Bopping Heads, and Oreo Shakes: Understanding Indicators of Virality in TikTok Short Videos
  3. 3.Xiao et al. (2024). Exploring user engagement behavior with short-form video advertising: a visual-audio perspective
  4. 4.Du et al. (2023). Improving Factuality and Reasoning in Language Models through Multiagent Debate
  5. 5.Liang et al. (2023). Encouraging Divergent Thinking in Large Language Models through Multi-Agent Debate
  6. 6.Zhou et al. (2025). Adaptive heterogeneous multi-agent debate for enhanced educational and factual reasoning in LLMs
  7. 7.Wataoka et al. (2024). Self-Preference Bias in LLM-as-a-Judge
  8. 8.Zheng et al. (2023). Judging LLM-as-a-judge with MT-Bench and Chatbot Arena
  9. 9.Ouyang et al. (2025). The Influence of Social Media Marketing on Consumers' Purchase Intention
  10. 10.Lee et al. (2018). Advertising Content and Consumer Engagement on Social Media: Evidence from Facebook
  11. 11.S. K. V. et al. (2021). Social media advertisements and their influence on consumer purchase intention
  12. 12.Shi et al. (2024). Judging the Judges: A Systematic Study of Position Bias in LLM-as-a-Judge
  13. 13.Doshi et al. (2023). Generative AI enhances individual creativity but reduces the collective diversity of novel content
  14. 14.Anderson et al. (2024). Homogenization Effects of Large Language Models on Human Creative Ideation
  15. 15.Sourati et al. (2025). The Homogenizing Effect of Large Language Models on Human Expression and Thought
  16. 16.Tafesse et al. (2026). Steady posts, stronger bonds? Posting consistency and follower engagement
  17. 17.Staley et al. (2025). Engagement and Follower Growth: Social Media Strategies of Swiss Athletes at the Paris Olympics
  18. 18.Bongulwar et al. (2025). Machine Learning in Social Media Account Analytics
  19. 19.Bellman et al. (2021). Can muted video advertising be as effective as video advertising with sound?
  20. 20.Pascual-Ferrá et al. (2023). Assessing Message Deployment During Public Health Emergencies Through Social Media
  21. 21.Hartmann et al. (2024). The power of generative marketing: Can generative AI create superhuman visual marketing content?
  22. 22.Mao et al. (2018). A Batched Multi-Armed Bandit Approach to News Headline Testing
  23. 23.Xiang et al. (2021). Adaptively Optimize Content Recommendation Using Multi Armed Bandit Algorithms in E-commerce
  24. 24.Xiang et al. (2022). Multi Armed Bandit vs. A/B Tests in E-commerce
  25. 25.Az-Zahrah et al. (2026). Predicting Social Media Post Engagement and Virality Using Graph Neural Network Approaches
  26. 26.Bhagyashree et al. (2026). Early Detection and Forecasting of Emerging Trends Using Explainable ML and Multi-Source Social Media Analytics
  27. 27.Varol et al. (2017). Early detection of promoted campaigns on social media
  28. 28.Baek et al. (2024). Effect of disclosing AI-generated content on prosocial advertising evaluation
  29. 29.Israfilzade (2025). AI-generated versus human-created advertising: Effects on consumer trust and purchase intent
  30. 30.Koning et al. (2025). Disclaimer! This Content Is AI-Generated: How AI-Disclosures Influence Trust
  31. 31.Kapoor et al. (2025). Frontiers: Generative AI and Personalized Video Advertisements