{
  "id": 12240029,
  "title": "Evaluating in-context learning with prompting regimes for transformer-based synthetic health tabular data generation",
  "url": "https://urgent.news/2026/10/05/evaluating-in-context-learning-with-prompting-regimes-for-transformer",
  "topic": "ai",
  "section": "AI",
  "published": "2026-10-05T00:00:00.000Z",
  "source": {
    "name": "bioRxiv",
    "slug": "biorxiv",
    "url": "https://www.biorxiv.org/content/10.64898/2026.09.29.754354v1?rss=1"
  },
  "original_language": "en",
  "account": "Tabular synthetic data generation (SDG) has the potential to aid research and development in the health sector, as access to health data is often restricted. Due to the intricate nature of such data, transformer models have emerged as a more suitable alternative to traditional SDG techniques. These models excel at capturing global spatial coherence within tabular records through the utilization of textual context.\n\nHowever, the optimal strategies for adapting transformer models to SDG tasks remain largely unexplored. This study aims to shed light on this matter by examining the fidelity and privacy aspects of three prompting regimes, namely zero shot, one shot, and few shot training, applied to tabular health SDG using compact transformer models.\n\nThe research team assessed the performance of llama 3.2 (1B and 3B parameters) and Phi (1.3B and 2.7B parameters) on a real clinical dataset. The findings revealed that zero shot learning exhibited a lower propensity for hallucinations and generally outperformed in terms of fidelity. Nevertheless, this advantage came at the expense of privacy, as compared to the alternative prompting regimes.\n\nInterestingly, the study observed no discernible differences in performance across the various model sizes and families tested. The researchers suggest that future investigations should broaden the scope of this analysis to encompass additional models and datasets, as well as investigate potential methods to mitigate the presence of hallucinations.",
  "summary": "Tabular synthetic data generation (SDG) can facilitate research and development in the health domain, as access to health data is restricted. To accommodate for its structural complexity, such as multimodal distributions and fragmentation, transformer models provides an alternative to conventional SDG methods since they are better equipped to infer global spatial coherence due to textual context…",
  "key_points": [],
  "editors_take": null,
  "illustration": null,
  "coverage": {
    "outlets": 1,
    "also_reported_by": []
  },
  "ai_generated": true,
  "disclaimer": "Summaries, key points and the editor’s take are written by software from other outlets’ reporting and may contain errors — always check the linked original."
}