ARASH Boosts Tabular Prediction Efficiency with Adaptive Shot Selection

Samirasadat Jamalidinan, Yue Xu, Kazem Cheshmi· August 19, 2026 View original

Key takeaways

  • TFMs are powerful but resource-intensive for tabular prediction.
  • Few-shot prompting is efficient but shot selection is challenging for tabular data.
  • ARASH adaptively selects optimal shots based on local neighborhood analysis.
  • ARASH significantly reduces TFM prompt length and memory usage with comparable accuracy.

Who benefits

FinanceHealthcareRetailLogisticsManufacturing

Summary

This paper introduces ARASH, a method that enhances the efficiency of Tabular Foundation Models (TFMs) like TabPFN by adaptively selecting optimal few-shot examples based on local neighborhood analysis. ARASH significantly reduces prompt length and memory usage while maintaining comparable accuracy.

Tabular data prediction is a crucial task across many applications, and while specialized Tabular Foundation Models (TFMs) like TabPFN have shown promise, they often demand substantial computational resources and frequent retraining. In-context learning (ICL) through few-shot prompting offers a more resource-efficient alternative, but effectively selecting the most relevant "shots" for tabular data remains a challenge. This research presents ARASH (Adaptive, query-specific Retrieval And Shot selection), a novel method designed to improve TFM efficiency. ARASH works by intelligently selecting optimal few-shot examples based on a local neighborhood analysis within the training dataset. The results demonstrate that ARASH can dramatically reduce the prompt length and memory footprint of TFMs, specifically TabPFN, by factors of 1261.5x and 2.56x respectively, all while maintaining competitive predictive accuracy.

Why it matters

For professionals working with tabular data and deploying AI models, ARASH provides a way to significantly reduce the computational cost and memory requirements of powerful TFMs, making them more practical and scalable for real-world applications without sacrificing performance.

How to implement this in your domain

  1. 1Integrate ARASH: Explore implementing ARASH or similar adaptive shot selection techniques for few-shot prompting in tabular prediction tasks.
  2. 2Optimize TFM deployment: Apply ARASH to reduce the resource footprint of Tabular Foundation Models, enabling more efficient deployment and inference.
  3. 3Enhance in-context learning: Develop strategies for query-specific retrieval of relevant examples to improve the performance and efficiency of in-context learning for tabular data.
  4. 4Reduce operational costs: Leverage ARASH's efficiency gains to lower the computational and memory costs associated with tabular AI model usage.

Original post by Samirasadat Jamalidinan, Yue Xu, Kazem Cheshmi

"arXiv:2608.17856v1 Announce Type: new Abstract: Tabular prediction is a critical task across numerous applications. The recent success of large language models has sparked various approaches for adapting them to the tabular domain. A prevalent strategy involves training or fine-t…"

View on X

Originally posted by Samirasadat Jamalidinan, Yue Xu, Kazem Cheshmi on X · view source

Want to go deeper?

Turn these trends into skills with Learnijoy's hands-on AI & tech courses.

Explore courses

More in AI Engineering & DevTools