# CIKM Short Paper: Knowing When to Correct
**Venue:** ACM International Conference on Information and Knowledge Management (CIKM)
**Authors:** Stergios Konstantinidis, Hayman Lotfy, Michalis Vlachos
**Affiliation:** Department of Information Systems (DESI), HEC Lausanne
**GitHub Public:** [CIKM\_public](https://github.com/Stergios-Konstantinidis/CIKM_public)
**Overleaf Project:** `https://git.overleaf.com/6a1d78e803cbdf7def32a839`
**Core Codebase:** `/Users/stergios/Documents/GitHub/Optimizing-LLM-based-OCR-Corrections`
---
## 1. Problem Statement
At multi-million page scale, LLM-based OCR correction is constrained by cost:
- Full LLM correction costs hundreds of thousands of dollars on national archival collections.
- Many archival passages are already accurate; invoking LLMs on them introduces latency, cost, and risk of hallucinations.
**Core Research Questions:**
- **RQ1:** How closely does selective routing match the theoretical oracle ceiling, and does it outperform naive confidence thresholding?
- **RQ2:** How well does the regression-estimated improvement *\\hat{Δ}\_i* correlate with the true realized improvement *Δ*i**?
- **RQ3:** What is the actual operational cost saved when routing only segments predicted to benefit across various frontier LLMs?
---
## Problem Visual Illustration
[](https://wiki.stergios.ch/uploads/images/gallery/2026-09/cikm-ocr-example.png "Click to view full-resolution image")