Back to papers

Cut Costs, Not Accuracy: LLM-Powered Data Processing with Guarantees

Summary: Targets LLM cascade cost-quality tradeoffs by provable selection of when to use cheaper LLMs for record processing, addressing weak quality estimation in prior confidence-based cascades. BARGAIN uses adaptive sampling and statistical estimation tuned to data/task to give tight theoretical guarantees (accuracy/precision/recall) and empirically reduces cost up to 86% vs. state-of-the-art. (summarized by gpt-5-mini on Feb 11 2026)

Paper ID

7373

Venue

SIGMOD

Year

2026

Pagerank

4.1905499e-05

Overall Rank

10,064 | 30.06%

DOI

10.1145/3769776