The InnoComp 2025 study compared prompted language models on Romanian diacritic restoration, the task of recovering missing marks while preserving the rest of the text. Two 1,000-item datasets and eight evaluators produced one aggregate score per prompt configuration.

Romanian diacritic restoration looks easy when most of the characters being counted never needed restoration. In our InnoComp comparison, that innocent denominator helped an echo baseline reach 0.8100.

Publication update, January 2026. The paper appears in Springer Communications in Computer and Information Science, volume 2794—not LNCS. The published chapter and arXiv preprint are now available.

Prompting moved the ranking

The study uses eight task evaluators and aggregates them into a total average score (TAS). For each system, the maximum TAS (MTAS) selects its best prompt configuration across zero-shot through three-shot prompting. GPT-4o reached an MTAS of 0.9639, 19% above the echo baseline.

System MTAS
GPT-4o 0.9639
GPT-4 0.9350
Gemini 1.0 Pro 0.9108
Llama 3 70B 0.8735
Echo baseline 0.8100
Llama 3 8B 0.7663
RoLlama 2 7B 0.6463

Model size did not determine the outcome, and prompt choice was not a cosmetic detail: GPT-4o improved by 1.66% from the two-shot to the three-shot setup.

Where the remaining errors collect

Most Romanian characters do not need a diacritic. A system can therefore score well while missing the positions the task exists to restore. In the error analysis, misplaced â accounted for 21.3% of all errors. Separately, the authors estimate that rule-aware post-processing could eliminate up to 19% of the remaining mistakes. Sentence-initial capitals and final syllables were also recurring trouble spots.

Generative systems add another failure: they can rewrite text outside the restoration target. A fluent sentence that changes a name or non-diacritical character has still violated the contract.

What the study is good for

The comparison supplies prompted-model baselines on two 1,000-item datasets. It does not decide which system belongs in a keyboard, an OCR cleanup pipeline, or a private document workflow. Those choices also need latency, cost, privacy, unwanted-edit, and noisy-text measurements.

I presented the work at InnoComp 2025 in Cluj-Napoca on 22–24 October. GPT-4o led at 0.9639. The echo baseline’s 0.8100 changed the next experiment: diacritizable-position accuracy and unwanted-edit rate became primary measures.