cs.CLNov 3, 2026

Rewrite, Then Veto: Rewriter-Agnostic Edit Gating for LLM-Based Post-OCR Correction

Authors: Jorge Ramón Fonseca Cacho

Organizations: Department of Computer Science, University of Nevada, Las Vegas, NV, USA

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

LLMs can repair noisy OCR transcripts of historical documents, but they also change text that was already correct. We propose EditVeto, a veto gate placed after a full-context LLM rewrite. The rewrite is character-aligned to the OCR input, its differences are split into blocks, and a 0.6B-parameter gate keeps or reverts each block. An oracle analysis motivates the design: given the true error locations, per-span generation is worse than no correction, whereas vetoing a full rewrite's edits helps. Trained on rewrites by two Qwen3 models and evaluated, frozen, on the eight HIPE-OCRepair-2026 test sets, the gate improves all five rewriters tested, two of them unseen in training: weighted cMER falls by 0.0026-0.0044 and the preference score rises by 0.15-0.25 (paired bootstrap, p ~ 0.0002 each). On published competition outputs it lowers cMER for three of four teams but raises it for the winning run (0.0050 to 0.0065). In a second, disclosed test look, a retrained gate still improves LoRA-tuned rewriters (cMER -0.0011 to -0.0021), and an ensemble of their gated edits reaches a cMER of 0.0069 and a preference score of 0.928, mostly through fine-tuning and pooling rather than gating. Built only on public models and data, it is behind the winner (0.0050) and tied with the second-ranked run (0.0071), both from the only team using domain pre-training, and ahead of all other published runs. A third, secondary look shows the unchanged gate improving zero-shot Qwen3.8-27B (cMER 0.0114 to 0.0108) but only the preference score of its LoRA-tuned version.

Figures & tables

Figure 1: Overview of EditVeto on a constructed example (not from the data). The rewrite is aligned to the OCR input, each edit block is kept or vetoed, and vetoed blocks revert to the OCR text. Here the gate keeps the OCR fix and vetoes the modernized spelling and the case change.
Figure 2: Test cMER (official weighted mean, 590 units) of every single rewriter without gate, with the full gate, and with the label oracle ( ∗ reads the reference; not a system). Look 1 uses the frozen gate, looks 2 and 3 the retrained one (Sect. 8.1 ); no correction is at 0.0226. Values in Table 1 and Table 6 in the appendix.
RewriterNo gateFull gateText-onlyΔ\Delta cMER (full)
Qwen2.5-72B zs †0.0148 / 0.61070.0105 / 0.83430.0106 / 0.8324−- 0.0043 [ −- 0.0055, −- 0.0034]
LoRA-24B †0.0107 / 0.79800.0086 / 0.91690.0087 / 0.9100−- 0.0021 [ −- 0.0026, −- 0.0016]
LoRA-32B †0.0107 / 0.79610.0088 / 0.89800.0089 / 0.8913−- 0.0019 [ −- 0.0023, −- 0.0015]
LoRA-72B0.0096 / 0.84090.0085 / 0.89440.0086 / 0.8974−- 0.0011 [ −- 0.0016, −- 0.0007]
LoRA-72B-syn0.0098 / 0.90040.0080 / 0.93200.0081 / 0.9302−- 0.0018 [ −- 0.0042, −- 0.0004]
Ensemble A–0.0071 / 0.9244––
Table 1: Looks 2 and 3, test set (590 scored units): weighted cMER / preference score. Gates: the retrained gate (Sect. 8.1 ), unchanged in look 3. Δ\Delta : full gate −- no gate, paired bootstrap, 95% CI (look 2: p≈0.0002p\approx 0.0002 on cMER for all five, on preference p≤0.011p\leq 0.011 ; look 3: see text). † The gate was trained on this rewriter’s training-split outputs. LoRA-24B/32B/72B: adapters on Mistral-Small-24B , Qwen3-32B , Qwen2.5-72B ; ensembles A–C: text; zs: zero-shot, rerun with the look-1 prompt and decoding (look 1: 0.0149 / 0.6074).
Appendix figures & tables12 assets

Supplementary material from the paper’s appendix.

Appendix

SystemcMER ↓\downarrowPref ↑\uparrow
No correction0.02750.0000
Zero-shot rewrite (ZS)0.02470.4161
Oracle per-span generation (earlier design)0.0562−- 0.6247
Oracle window rewrite0.02550.2627
ZS + oracle span gate (block)0.02290.6501
ZS + oracle span gate (character)0.02360.5681
Appendix
Table 2: Development set, oracle analysis (earlier Qwen3-8B zero-shot run; see Sect. 5.3 ): weighted means over the seven official development cells (518 units). Oracle systems read error locations from the reference and are not valid systems.
 No corr.ZSZS + OSGOPGOW
CellcMERcMERPrefcMERPrefcMERcMERPref
icdar2017 -en (188)0.03430.02880.64360.02780.73400.05060.02850.7500
impresso -en (10)0.03720.03500.40000.03500.40000.06360.03350.5000
impresso -fr (10)0.01500.01130.60000.01020.80000.05270.01540.0000
impresso -de (10)0.02260.02090.40000.01690.90000.03400.02190.0000
dta19-l0 (100)0.00450.0078−- 0.53000.00450.20000.01820.00490.0000
Appendix
Table 3: Development set, oracle analysis: per-cell micro cMER and preference score (earlier Qwen3-8B zero-shot run). ZS+OSG: zero-shot rewrite with the oracle span gate (block). OPG: oracle per-span generation. OW: oracle window rewrite. Units per cell in parentheses.
 Qwen3-8BQwen3-32B
SystemcMERPrefcMERPref
No correction0.02650.00000.02650.0000
Zero-shot rewrite (ZS)0.02380.54220.02840.6533
ZS + oracle span gate (block)0.02220.70220.01780.8000
ZS + oracle span gate (char.)0.02190.62220.01690.8133
ZS + reverse span gate (block)––0.0371−- 0.4244
Appendix
Table 4: Development probe subset (100 units, 90 in official cells; fixed per-cell allocation, seed 1234), oracle analysis: weighted cMER / preference score. Not comparable to Table 2 . –: not run.
RewriterNo gateFull gateText-only gateSpan oracle ∗Label oracle ∗
Qwen3-8B0.0245 / 0.4610.0213 / 0.6770.0211 / 0.6790.0228 / 0.6560.0198 / 0.745
Qwen3-32B0.0220 / 0.6000.0160 / 0.8100.0161 / 0.8110.0186 / 0.7930.0135 / 0.859
Qwen3-8B + LoRA0.0176 / 0.6770.0143 / 0.898n/r0.0159 / 0.8530.0108 / 0.970
Mistral-Small-24B0.0251 / 0.4860.0133 / 0.8960.0138 / 0.8370.0195 / 0.7320.0110 / 0.944
Qwen2.5-72B0.0216 / 0.5630.0134 / 0.8330.0133 / 0.8150.0162 / 0.8020.0105 / 0.910
Appendix
Table 5: Development set: weighted cMER / preference score over the seven official cells (518 units). Full and text-only: the frozen pooled Qwen3-0.6B gates. Columns marked ∗ read the reference and are not valid systems. No correction: 0.0275 / 0.000. n/r: not run.
 No gateFull gateText-only gateLO ∗
RewritercMER ↓\downarrowPref ↑\uparrowcMERPrefcMERPrefcMERPref
Qwen3-8B0.01840.49460.01530.71240.01530.69500.01340.7780
Qwen3-32B0.01460.64330.01190.82280.01200.80830.00950.8844
Qwen3-8B + LoRA0.01400.70610.01140.85410.01140.85350.00900.9274
Mistral-Small-24B0.01580.60020.01160.84610.01160.82890.00860.9230
Qwen2.5-72B0.01490.60740.01060.84460.01060.82810.00770.9132
Appendix
Table 6: Test set, our rewriters: weighted micro cMER and preference score over the eight official cells (590 scored units). Full and text-only: the frozen gates. LO ∗ : label oracle (reads the reference). Lower part: paired-bootstrap difference to the ungated rewriter (95% CIs for the text-only gate in Table 10 ); p≈0.0002p\approx 0.0002 , the smallest attainable value, for all 20 differences. No correction: 0.0226 / 0.0000.
CellUnitsBaseBestQwen3-8BQwen3-32BLoRAMistralQwen2.5
dta19-l0300.400.29 a0.88 →\to 0.401.19 →\to 0.340.45 →\to 0.281.48 →\to 0.341.27 →\to 0.33
dta19-l1302.400.543.21 →\to 1.551.87 →\to 1.251.36 →\to 1.081.58 →\to 1.172.08 →\to 0.95
dta19-l2305.460.825.03 →\to 4.064.25 →\to 2.943.43 →\to 2.535.51 →\to 2.924.33 →\to 2.35
icdar2017 -en1002.600.441.71 →\to 1.591.22 →\to 0.961.26 →\to 1.091.58 →\to 0.931.65 →\to 0.79
icdar2017 -fr1001.840.401.23 →\to 1.360.87 →\to 1.101.33 →\to 1.290.89 →\to 1.110.91 →\to 1.05
impresso -de1002.960.582.47 →\to 1.872.05 →\to 1.651.94 →\to 1.472.11 →\to 1.651.93 →\to 1.55
Appendix
Table 7: Test set, per cell: micro cMER ×100\times 100 of the OCR (Base) and the best competition result per cell (Best; BnF-Mistral run 1 unless noted), both copied from Table 7 of [ 5 ] , and of each rewriter without →\rightarrow with the frozen full gate (LoRA: Qwen3-8B + LoRA). Units: scored units. Weighted Best: BnF-Mistral run 1 overall. The dta19 cells have weight 1/31/3 . Shaded: the gate raises cMER.
Run (cells)cMERPrefΔ\Delta cMER [95% CI]pp (cMER / Pref)
BnF-Mistral 1 (8)0.0050 →\to 0.00650.9000 →\to 0.9252+0.0015 [+0.0012, +0.0019]0.0002 / 0.0090
BnF-Mistral 2 (8)0.0090 →\to 0.00900.8739 →\to 0.9070−- 0.0001 [ −- 0.0004, +0.0003]0.69 / 0.0054
BnF-Mistral 3 (8)0.0071 →\to 0.00770.8737 →\to 0.9196+0.0006 [+0.0003, +0.0008]0.0002 / 0.0002
BLOCR 1 (8)0.0106 →\to 0.00960.7028 →\to 0.8596−- 0.0011 [ −- 0.0016, −- 0.0006]0.0002 / 0.0002
L3i 1 (8)0.0176 →\to 0.01570.3591 →\to 0.4902−- 0.0018 [ −- 0.0027, −- 0.0012]0.0002 / 0.0002
L3i 2 (3)0.0221 →\to 0.0199−- 0.0933 →\to 0.0167−- 0.0022 [ −- 0.0028, −- 0.0017]0.0002 / 0.0002
Appendix
Table 8: Test set, competition systems: weighted cMER and preference score of the published run [ 5 ] (reproduced with the official scorer) →\to the same outputs after the frozen text-only gate, and the paired-bootstrap cMER difference with 95% CI and pp for both metrics. Means over the cells each run covers (in parentheses). The gate leaves the no-correction baseline unchanged.
CellUnitsQwen3-8BQwen3-32BLoRA-8BMistral-24BQwen2.5-72B
dta19-l030−- 0.70 →\to 0.00−- 0.97 →\to 0.300.30 →\to 0.53−- 0.47 →\to 0.30−- 0.63 →\to 0.33
dta19-l1300.20 →\to 0.970.73 →\to 1.000.83 →\to 1.000.73 →\to 1.000.47 →\to 1.00
dta19-l2300.43 →\to 0.970.53 →\to 1.000.87 →\to 1.000.67 →\to 1.000.60 →\to 1.00
icdar2017 -en1000.70 →\to 0.760.87 →\to 0.940.91 →\to 0.930.70 →\to 0.940.78 →\to 0.98
icdar2017 -fr1000.85 →\to 0.840.96 →\to 0.940.78 →\to 0.840.94 →\to 0.980.92 →\to 0.96
impresso -de1000.42 →\to 0.800.68 →\to 0.800.52 →\to 0.870.57 →\to 0.940.61 →\to 0.81
Appendix
Table 9: Test set, per cell: preference score of each rewriter without →\rightarrow with the frozen full gate (590 scored units).
RewriterΔ\Delta cMER [95% CI]Δ\Delta Pref [95% CI]
Qwen3-8B−- 0.0030 [ −- 0.0042, −- 0.0021]+0.200 [0.154, 0.248]
Qwen3-32B−- 0.0026 [ −- 0.0031, −- 0.0022]+0.165 [0.124, 0.207]
Qwen3-8B + LoRA−- 0.0027 [ −- 0.0037, −- 0.0019]+0.147 [0.106, 0.192]
Mistral-Small-24B−- 0.0042 [ −- 0.0058, −- 0.0029]+0.229 [0.177, 0.284]
Qwen2.5-72B−- 0.0043 [ −- 0.0056, −- 0.0033]+0.221 [0.172, 0.272]
Appendix
Table 10: Test set: paired-bootstrap difference between the text-only gate and no gate (95% CI; p≈0.0002p\approx 0.0002 for all).
CellUnitsOCRLoRA-24BLoRA-32BLoRA-72B72B-synQ72Ens.
icdar2017 -en1002.600.97 →\to 0.810.97 →\to 0.830.88 →\to 0.790.83 →\to 0.791.57 →\to 0.820.61
icdar2017 -fr1001.841.30 →\to 1.171.12 →\to 1.131.05 →\to 1.090.97 →\to 1.030.89 →\to 1.000.94
impresso -en1001.700.69 →\to 0.710.73 →\to 0.700.66 →\to 0.690.64 →\to 0.720.74 →\to 0.830.63
impresso -fr1001.690.70 →\to 0.590.65 →\to 0.640.78 →\to 0.661.33 →\to 0.611.19 →\to 0.940.55
impresso -de1002.961.58 →\to 1.091.58 →\to 1.101.26 →\to 0.981.08 →\to 0.901.92 →\to 1.500.77
dta19-l0300.400.30 →\to 0.210.47 →\to 0.250.31 →\to 0.200.30 →\to 0.201.28 →\to 0.330.21
Appendix
Table 11: Look 2, test set, per cell: micro cMER ×100\times 100 of the OCR and of each rewriter without →\rightarrow with the retrained full gate, and of the four-rewriter ensemble (Ens.). Q72: zero-shot Qwen2.5-72B . Shaded: the gate raises cMER.
Run (cells)cMER (run →\to ens.)Δ\Delta cMER [95% CI], ppΔ\Delta Pref, pp
BnF-Mistral 1 (8)0.0050 →\to 0.0069+0.0019 [+0.0014, +0.0024], 0.0002+0.028, 0.08
BnF-Mistral 3 (8)0.0071 →\to 0.0069−- 0.0002 [ −- 0.0008, +0.0004], 0.49+0.054, 0.0006
BnF-Mistral 2 (8)0.0090 →\to 0.0069−- 0.0022 [ −- 0.0027, −- 0.0016], 0.0002+0.054, 0.0016
BLOCR 1 (8)0.0107 →\to 0.0069−- 0.0038 [ −- 0.0046, −- 0.0031], 0.0002+0.225, 0.0002
L3i 1 (8)0.0176 →\to 0.0069−- 0.0107 [ −- 0.0122, −- 0.0094], 0.0002+0.569, 0.0002
L3i 2 (3)0.0221 →\to 0.0065−- 0.0156 [ −- 0.0172, −- 0.0141], 0.0002+1.007, 0.0002
Appendix
Table 12: Look 2, test set: the four-rewriter ensemble minus each published run (ungated), paired bootstrap over the cells the run covers. Negative Δ\Delta cMER and positive Δ\Delta Pref favor the ensemble; preference CIs are omitted for space.
Figure 3: Test set: weighted cMER against preference score. Arrows go from the ungated output to the gated output: our five rewriters with the full gate (black) and the published competition runs with the text-only gate (blue), all from look 1. Star: the look-2 ensemble (Sect. 8.1 ). Lower right is better. Values from Tables 6 , 8 , and 1 .
More in this topic

Historical Document Processing

Historical document processing uses AI to transcribe, structure, restore, and analyze manuscripts, books, newspapers, and archival records.

See all 79 papers

Explore similar work

CardsList

  1. ICDAR 2026 HIPE-OCRepair Competition on LLM-Assisted OCR Post-Correction for Historical Documents

    Jul 9, 2026Maud Ehrmann, Emanuela Boros, Juri Opitz +3Historical Document AnalysisLLM Evaluation

    1 École Polytechnique Fédérale de Lausanne (EPFL), Switzerland · University of Zurich, Switzerland

  2. Edit-level Majority Voting Mitigates Over-Correction in LLM-based Grammatical Error Correction

    May 13, 2026Takumi Goto, Yusuke Sakai, Taro WatanabeSelf-Consistency DecodingGrammatical Error Correction

    Nara Institute of Science and Technology (NAIST)

  3. Reviser: Writing and Editing Text with Autoregressive Cursor Actions

    Jul 23, 2026Sean DiabText GenerationLanguage Model Decoding

    Independent Researcher