arxiv arXiv cs.CL · 2d ago · src: 6d ago · research

The Token Tax of Epistemic Accuracy in Document-Grounded AI

from English

A study compares retrieval-augmented generation (RAG) and long-context prompting in document-grounded AI. Long-context prompting achieves higher epistemic accuracy—73.1% vs. 65.4%—but at 26 times the per-query token cost, highlighting a significant token tax for broader evidentiary access.

Importance 2/3 arXiv cs.CL Evaluation & benchmarks Retrieval & RAG

Benchmarks

Benchmark	Model	Score
SWE-bench Verified	three machines, two small language models, and three retrieval/in-context prompting approaches	73.1%

Read original