Performance ceiling of purely end-to-end OCR architectures

Ascertain the ultimate performance ceiling of purely end-to-end optical character recognition architectures relative to heavily optimized two-stage pipeline OCR systems by determining whether end-to-end models can fully match or surpass pipeline approaches under variations in architecture, training strategies, and data scaling laws.

Background

Qianfan-OCR is presented as a unified end-to-end model that competes strongly with pipeline systems on several benchmarks, narrowing but not entirely eliminating gaps with top pipeline methods. This raises a broader question about the fundamental limits of end-to-end designs.

The authors explicitly identify uncertainty about whether purely end-to-end architectures can ultimately match or surpass heavily engineered pipelines, calling for systematic exploration of architectural innovations, training strategies, and data scaling laws to resolve this question.

References

As a pioneering attempt at end-to-end OCR, the ultimate performance ceiling of purely end-to-end architectures remains an open question -- future work should systematically explore architectural innovations, training strategies, and data scaling laws to determine whether end-to-end models can fully match or surpass heavily optimized pipeline systems.

— Qianfan-OCR: A Unified End-to-End Model for Document Intelligence  (2603.13398 - Dong et al., 11 Mar 2026) in Section 7: Limitations and Future Work — Performance Ceiling

We hypothesize that training a VLM specifically for performing OCR on historical newspaper crops could produce a model that is both smaller, more accurate and less sensitive to token looping issues than dots.mocr for this specific task.

— Institutional Newspapers Pipeline: Deriving billions of high quality tokens from historical newspapers  (2608.18972 - Cargnelutti et al., 19 Aug 2026) in Section 14, Discussion and Future Directions