PSK at WMT 2026 MIST: Task-Specialized QLoRA Adapters for Multilingual Summarization and Question Answering

# PSK at WMT 2026 MIST: Task-Specialized QLoRA Adapters for Multilingual Summarization and Question Answering
Source: [https://arxiv.org/html/2608.20757](https://arxiv.org/html/2608.20757)
###### Abstract

We describe the PSK submission to the WMT 2026 Multilingual Instruction Shared Task\. Our system uses the 3\.35B\-parameter Tiny Aya Global model with three QLoRA adapters, one for each task\. The adapters are trained on multilingual document–summary pairs, passage\-based question answering, and filtered standalone question answering\. The summarization data also includes scientific papers with their author\-written abstracts\. On our held\-out split, the context and summarization adapters perform better than our multitask adapter, which was trained only on data supplied by the organizers\. Results for open QA are mixed and vary with answer length and evaluation method\. We therefore submit three systems with the same context and summarization adapters but different open\-QA adapters\.

## 1Introduction

The WMT 2026 Multilingual Instruction Shared Task \(MIST\) evaluates models under a 10B\-parameter limit on three tasks: context\-based question answering, summarization from a document in languageXXto languageYY, and open\-ended generation[18](https://arxiv.org/html/2608.20757#bib.bib1)\. The test set covers 24 languages and contains both same\-language and cross\-lingual generation\. MIST follows the broader multilingual evaluation effort introduced in WMT 2025, which found substantial variation across tasks and languages and also showed that automatic metrics are imperfect for open generation[9](https://arxiv.org/html/2608.20757#bib.bib2)\.

We use one multilingual backbone with three task\-specific adapters\. Tiny Aya Global[16](https://arxiv.org/html/2608.20757#bib.bib3)is first adapted to the provided data and then continued separately for each task\. This design preserves a single 3\.35B backbone while allowing the supervision and decoding policy to follow the output structure of each task\.

## 2Task and System Overview

MIST provides a task label at inference time\. We use it to select one of three LoRA adapters and the corresponding decoding configuration \(Figure[1](https://arxiv.org/html/2608.20757#S2.F1)\); there is no language\-specific routing\.

Prompt \+ task labelTask routerSummary adapterContext\-QA adapterOpen\-QA adapterTask\-specific greedy decodingand repetition controlOutput

Figure 1:The three routes share one frozen Tiny Aya Global backbone\. The known task label selects an adapter and its decoding policy; language does not change the route\.### 2\.1Base Model and Initial Adaptation

We useCohereLabs/tiny\-aya\-global, a 3\.35B\-parameter model covering 70 languages[16](https://arxiv.org/html/2608.20757#bib.bib3)\. Its 8K context window supports the longer task inputs\. An initial multitask adapter is trained on 20,293 provided examples using a deterministic 85/15 split; the resulting 3,626 held\-out examples are reused for model selection\. The task\-specific adapters are initialized from this model\.

## 3Training Data

Table[1](https://arxiv.org/html/2608.20757#S3.T1)summarizes the task\-specific training data\. Validation examples and exact test prompts are excluded from training\.

Table 1:Supervised data used by the selected task adapters\. Each is initialized from the multitask adapter\.### 3\.1Summarization

The 12,000\-example mixture is divided evenly between general and scientific text\. General examples combine provided data with CrossSum[2](https://arxiv.org/html/2608.20757#bib.bib4), WikiLingua[10](https://arxiv.org/html/2608.20757#bib.bib5), and UPDESH[4](https://arxiv.org/html/2608.20757#bib.bib6)\. The scientific portion uses full ACL Anthology papers[3](https://arxiv.org/html/2608.20757#bib.bib8)with author\-written abstracts\. No Language Left Behind \(NLLB\)[13](https://arxiv.org/html/2608.20757#bib.bib7)supplies non\-English targets, which are filtered for language, length, repetition, and semantic consistency\. We remove test overlaps and placeholder\-heavy papers\. We also screen the provided data for likely misalignment by comparing each summary with chunks of its source document and discarding a pair only when both its best LaBSE similarity and best chrF score fall below fixed thresholds\.

### 3\.2Context Question Answering

The context\-QA training set combines Belebele[1](https://arxiv.org/html/2608.20757#bib.bib9), answerable TyDi QA[5](https://arxiv.org/html/2608.20757#bib.bib10), MLQA[11](https://arxiv.org/html/2608.20757#bib.bib11), MCIF[14](https://arxiv.org/html/2608.20757#bib.bib12), UPDESH, and small Czech and Yoruba Aya subsets\. MLQA supplies parallel and cross\-lingual examples\.

### 3\.3Open\-Ended Question Answering

Open QA combines provided examples, WMT25 MIST questions, and the Aya Dataset and Collection[17](https://arxiv.org/html/2608.20757#bib.bib13)\. We retain short factual and explanatory QA while excluding passage\-dependent, multiple\-choice, creative\-writing, code, translation, and continuation tasks\. We also translate a small set of Hindi questions and answers into Bhojpuri with NLLB\.

## 4Training and Inference

We use QLoRA[6](https://arxiv.org/html/2608.20757#bib.bib14)with 4\-bit NF4 quantization and BF16 computation\. LoRA[8](https://arxiv.org/html/2608.20757#bib.bib15)is applied to the attention and feed\-forward projections with rank 16, alpha 32, and dropout 0\.05\. Each adapter is trained for one epoch with effective batch size 16; learning rates are1×10−41\\times 10^\{\-4\}for summarization,8×10−58\\times 10^\{\-5\}for open QA, and2×10−42\\times 10^\{\-4\}for context QA\.

Inference uses greedy decoding without sampling\. After observing repetitive outputs on the development set, we add repetition penalties of 1\.05 for open QA and 1\.03 for summarization\. We also block repeated 4\-grams in summarization outputs\.

## 5Development Results

We report exact match \(EM\), chrF[15](https://arxiv.org/html/2608.20757#bib.bib16), ROUGE\-L[12](https://arxiv.org/html/2608.20757#bib.bib17), and LaBSE cosine similarity[7](https://arxiv.org/html/2608.20757#bib.bib18)\. For open QA, we use chrF, ROUGE\-L, and LaBSE together with manual inspection of validation outputs\.

### 5\.1Task\-Level Results

Table[2](https://arxiv.org/html/2608.20757#S5.T2)reports development scores for the evaluated adapters\. The 8\.5k context mixture and the 12k summary mixture are selected for submission\.

Table 2:Internal model\-selection results on identical held\-out rows\. EM is a percentage and LaBSE is cosine similarity\.The targeted context continuation is marginally higher on ROUGE\-L and LaBSE, whereas the 8\.5k context mixture is higher on EM and chrF\. We therefore select the latter\.

### 5\.2Scientific Summarization

Table[3](https://arxiv.org/html/2608.20757#S5.T3)reports results on a 336\-example scientific validation set\. Gold targets are author abstracts; non\-English targets are translated and quality filtered\. The 12k mixture is strongest across all three metrics\.

Table 3:Results on the scientific summarization validation set \(n=336n=336\)\. The summary mixtures are evenly divided between general and scientific data\.Blocking repeated 4\-grams removes the observed summary loops\.

### 5\.3Open QA and Submitted Systems

Best\-score QA has the strongest aggregate automatic scores\. Long\-form QA is stronger on a manually identified set of long\-form prompts and reaches the generation limit less often\. Our main submission uses Long\-form QA\. A second submission uses Best\-score QA\. Table[4](https://arxiv.org/html/2608.20757#S5.T4)summarizes the three submitted routes\.

Table 4:Submitted routes\. The open\-QA variants are named for their selection criterion\.

## 6Analysis

Separate adapters improve context QA and summarization on our development set\. Scientific summarization also improves after adding papers and their abstracts to the training data\. Open QA has no clear winner\. Best\-score QA performs better on automatic metrics, while Long\-form QA works better on the longer questions we checked manually and is less likely to hit the output limit\.

The shared summarization route improves chrF over the multitask adapter in all 28 represented languages\. The shared context\-QA route improves exact match in 19 of 25 languages, ties in four, and declines in two\. Open QA remains mixed: the primary route improves chrF in 9 of 19 languages, and variant A improves it in 10 of 19\. Appendix[A](https://arxiv.org/html/2608.20757#A1)gives the complete language\-level changes and cell sizes\.

## 7Conclusion

We present a routed Tiny Aya Global system with one QLoRA adapter per MIST task\. Development results select the 8\.5k context mixture and the 12k summary mixture, while the three submissions vary the less stable open\-QA route\. Official evaluation will provide the appropriate shared\-task comparison\.

## Limitations

Reference metrics do not directly measure factuality and are particularly limited for open QA\. Translated scientific targets may retain artifacts\. Results use one backbone and one development split\. The router also assumes that the task label is available\.

## Ethics Statement

Public and challenge\-provided datasets are used under their respective terms\. Machine translation may reproduce source and model biases, and open\-ended outputs may contain incorrect or harmful claims\. The system should not be treated as a factual authority without verification\.

## References

- Bandarkaret al\.\(2024\)L\. Bandarkar, D\. Liang, B\. Muller, M\. Artetxe, S\. N\. Shukla, D\. Husa, N\. Goyal, A\. Krishnan, L\. Zettlemoyer, and M\. KhabsaThe Belebele benchmark: a parallel reading comprehension dataset in 122 language variants\.InProceedings of the 62nd Annual Meeting of the Association for Computational Linguistics \(Volume 1: Long Papers\),Bangkok, Thailand,pp\. 749–775\.External Links:[Document](https://dx.doi.org/10.18653/v1/2024.acl-long.44),[Link](https://aclanthology.org/2024.acl-long.44/)Cited by:[§3\.2](https://arxiv.org/html/2608.20757#S3.SS2.p1.1)\.
- Bhattacharjeeet al\.\(2023\)A\. Bhattacharjee, T\. Hasan, W\. U\. Ahmad, Y\. Li, Y\. Kang, and R\. ShahriyarCrossSum: beyond English\-centric cross\-lingual summarization for 1,500\+ language pairs\.InProceedings of the 61st Annual Meeting of the Association for Computational Linguistics \(Volume 1: Long Papers\),Toronto, Canada,pp\. 2541–2564\.External Links:[Document](https://dx.doi.org/10.18653/v1/2023.acl-long.143),[Link](https://aclanthology.org/2023.acl-long.143/)Cited by:[§3\.1](https://arxiv.org/html/2608.20757#S3.SS1.p1.1)\.
- Birdet al\.\(2008\)S\. Bird, R\. Dale, B\. Dorr, B\. Gibson, M\. Joseph, M\. Kan, D\. Lee, B\. Powley, D\. Radev, and Y\. F\. TanThe ACL Anthology reference corpus: a reference dataset for bibliographic research in computational linguistics\.InProceedings of the Sixth International Conference on Language Resources and Evaluation \(LREC’08\),Marrakech, Morocco\.External Links:[Link](https://aclanthology.org/L08-1005/)Cited by:[§3\.1](https://arxiv.org/html/2608.20757#S3.SS1.p1.1)\.
- Chitaleet al\.\(2026\)P\. A\. Chitale, V\. Gumma, S\. Ahuja, P\. Kodali, M\. Uppadhyay, D\. Sudharsan, and S\. SitaramUPDESH: synthesizing grounded instruction tuning data for 13 Indic languages\.InProceedings of the 64th Annual Meeting of the Association for Computational Linguistics \(Volume 1: Long Papers\),San Diego, California, United States,pp\. 37997–38041\.External Links:[Document](https://dx.doi.org/10.18653/v1/2026.acl-long.1763),[Link](https://aclanthology.org/2026.acl-long.1763/)Cited by:[§3\.1](https://arxiv.org/html/2608.20757#S3.SS1.p1.1)\.
- Clarket al\.\(2020\)J\. H\. Clark, E\. Choi, M\. Collins, D\. Garrette, T\. Kwiatkowski, V\. Nikolaev, and J\. PalomakiTyDi QA: a benchmark for information\-seeking question answering in typologically diverse languages\.Transactions of the Association for Computational Linguistics8,pp\. 454–470\.External Links:[Document](https://dx.doi.org/10.1162/tacl%5Fa%5F00317),[Link](https://aclanthology.org/2020.tacl-1.30/)Cited by:[§3\.2](https://arxiv.org/html/2608.20757#S3.SS2.p1.1)\.
- Dettmerset al\.\(2023\)T\. Dettmers, A\. Pagnoni, A\. Holtzman, and L\. ZettlemoyerQLoRA: efficient finetuning of quantized LLMs\.InAdvances in Neural Information Processing Systems,Vol\.36,New Orleans, Louisiana, USA,pp\. 10088–10115\.External Links:[Document](https://dx.doi.org/10.52202/075280-0441),[Link](https://proceedings.neurips.cc/paper_files/paper/2023/hash/1feb87871436031bdc0f2beaa62a049b-Abstract-Conference.html)Cited by:[§4](https://arxiv.org/html/2608.20757#S4.p1.1)\.
- Fenget al\.\(2022\)F\. Feng, Y\. Yang, D\. Cer, N\. Arivazhagan, and W\. WangLanguage\-agnostic BERT sentence embedding\.InProceedings of the 60th Annual Meeting of the Association for Computational Linguistics \(Volume 1: Long Papers\),Dublin, Ireland,pp\. 878–891\.External Links:[Document](https://dx.doi.org/10.18653/v1/2022.acl-long.62),[Link](https://aclanthology.org/2022.acl-long.62/)Cited by:[§5](https://arxiv.org/html/2608.20757#S5.p1.1)\.
- Huet al\.\(2022\)E\. J\. Hu, Y\. Shen, P\. Wallis, Z\. Allen\-Zhu, Y\. Li, S\. Wang, L\. Wang, and W\. ChenLoRA: low\-rank adaptation of large language models\.InInternational Conference on Learning Representations,External Links:[Link](https://openreview.net/forum?id=nZeVKeeFYf9)Cited by:[§4](https://arxiv.org/html/2608.20757#S4.p1.1)\.
- Kocmiet al\.\(2025\)T\. Kocmi, S\. Agrawal, E\. Artemova, E\. Avramidis, E\. Briakou, P\. Chen, M\. Fadaee, M\. Freitag, R\. Grundkiewicz, Y\. Hou, P\. Koehn, J\. Kreutzer, S\. Mansour, S\. Perrella, L\. Proietti, P\. Riley, E\. Sánchez, P\. Schmidtova, M\. Shmatova, and V\. ZouharFindings of the WMT25 multilingual instruction shared task: persistent hurdles in reasoning, generation, and evaluation\.InProceedings of the Tenth Conference on Machine Translation,Suzhou, China,pp\. 414–435\.External Links:[Document](https://dx.doi.org/10.18653/v1/2025.wmt-1.23),[Link](https://aclanthology.org/2025.wmt-1.23/)Cited by:[§1](https://arxiv.org/html/2608.20757#S1.p1.1)\.
- Ladhaket al\.\(2020\)F\. Ladhak, E\. Durmus, C\. Cardie, and K\. McKeownWikiLingua: a new benchmark dataset for cross\-lingual abstractive summarization\.InFindings of the Association for Computational Linguistics: EMNLP 2020,Online,pp\. 4034–4048\.External Links:[Document](https://dx.doi.org/10.18653/v1/2020.findings-emnlp.360),[Link](https://aclanthology.org/2020.findings-emnlp.360/)Cited by:[§3\.1](https://arxiv.org/html/2608.20757#S3.SS1.p1.1)\.
- Lewiset al\.\(2020\)P\. Lewis, B\. Oguz, R\. Rinott, S\. Riedel, and H\. SchwenkMLQA: evaluating cross\-lingual extractive question answering\.InProceedings of the 58th Annual Meeting of the Association for Computational Linguistics,Online,pp\. 7315–7330\.External Links:[Document](https://dx.doi.org/10.18653/v1/2020.acl-main.653),[Link](https://aclanthology.org/2020.acl-main.653/)Cited by:[§3\.2](https://arxiv.org/html/2608.20757#S3.SS2.p1.1)\.
- Lin \(2004\)C\. LinROUGE: a package for automatic evaluation of summaries\.InText Summarization Branches Out,Barcelona, Spain,pp\. 74–81\.External Links:[Link](https://aclanthology.org/W04-1013/)Cited by:[§5](https://arxiv.org/html/2608.20757#S5.p1.1)\.
- NLLB Team \(2024\)NLLB TeamScaling neural machine translation to 200 languages\.Nature630\(8018\),pp\. 841–846\.External Links:[Document](https://dx.doi.org/10.1038/s41586-024-07335-x),[Link](https://www.nature.com/articles/s41586-024-07335-x)Cited by:[§3\.1](https://arxiv.org/html/2608.20757#S3.SS1.p1.1)\.
- Papiet al\.\(2026\)S\. Papi, M\. Züfle, M\. Gaido, B\. Savoldi, D\. Liu, I\. Douros, L\. Bentivogli, and J\. NiehuesMCIF: multimodal crosslingual instruction\-following benchmark from scientific talks\.InThe Fourteenth International Conference on Learning Representations,External Links:[Link](https://openreview.net/forum?id=PtPYZYfa0h)Cited by:[§3\.2](https://arxiv.org/html/2608.20757#S3.SS2.p1.1)\.
- Popović \(2015\)M\. PopovićChrF: character n\-gram F\-score for automatic MT evaluation\.InProceedings of the Tenth Workshop on Statistical Machine Translation,Lisbon, Portugal,pp\. 392–395\.External Links:[Document](https://dx.doi.org/10.18653/v1/W15-3049),[Link](https://aclanthology.org/W15-3049/)Cited by:[§5](https://arxiv.org/html/2608.20757#S5.p1.1)\.
- Salamancaet al\.\(2026\)A\. R\. Salamanca, D\. Abagyan, D\. D’souza, A\. Khairi, D\. Mora, S\. Dash, V\. Aryabumi, S\. Rajaee, M\. Mofakhami, A\. Sahu, T\. Euyang, B\. Prince, M\. Smith, H\. Lin, A\. Locatelli, S\. Hooker, T\. Kocmi, A\. Gomez, I\. Zhang, P\. Blunsom, N\. Frosst, J\. Pineau, B\. Ermis, A\. Üstün, J\. Kreutzer, and M\. FadaeeTiny Aya: bridging scale and multilingual depth\.External Links:2603\.11510,[Document](https://dx.doi.org/10.48550/arXiv.2603.11510),[Link](https://arxiv.org/abs/2603.11510)Cited by:[§1](https://arxiv.org/html/2608.20757#S1.p2.1),[§2\.1](https://arxiv.org/html/2608.20757#S2.SS1.p1.1)\.
- Singhet al\.\(2024\)S\. Singh, F\. Vargus, D\. D’souza, B\. F\. Karlsson, A\. Mahendiran, W\. Ko, H\. Shandilya, J\. Patel, D\. Mataciunas, L\. O’Mahony, M\. Zhang, R\. Hettiarachchi, J\. Wilson, M\. Machado, L\. Moura, D\. Krzemiński, H\. Fadaei, I\. Ergun, I\. Okoh, A\. Alaagib, O\. Mudannayake, Z\. Alyafeai, V\. Chien, S\. Ruder, S\. Guthikonda, E\. Alghamdi, S\. Gehrmann, N\. Muennighoff, M\. Bartolo, J\. Kreutzer, A\. Üstün, M\. Fadaee, and S\. HookerAya dataset: an open\-access collection for multilingual instruction tuning\.InProceedings of the 62nd Annual Meeting of the Association for Computational Linguistics \(Volume 1: Long Papers\),Bangkok, Thailand,pp\. 11521–11567\.External Links:[Document](https://dx.doi.org/10.18653/v1/2024.acl-long.620),[Link](https://aclanthology.org/2024.acl-long.620/)Cited by:[§3\.3](https://arxiv.org/html/2608.20757#S3.SS3.p1.1)\.
- WMT 2026 MIST Organizers \(2026\)WMT 2026 MIST OrganizersMultilingual instruction shared task\.Note:[https://www2\.statmt\.org/wmt26/multilingual\-instruction\.html](https://www2.statmt.org/wmt26/multilingual-instruction.html)Accessed 7 August 2026Cited by:[§1](https://arxiv.org/html/2608.20757#S1.p1.1)\.

## Appendix ALanguage\-Level Changes

Tables[5](https://arxiv.org/html/2608.20757#A1.T5)and[6](https://arxiv.org/html/2608.20757#A1.T6)report changes from the multitask adapter on identical held\-out examples\. All submitted systems share the summary and context routes\. The primary system uses Long\-form QA, variant A uses Best\-score QA, and variant B retains the multitask open\-QA route, for which every open\-QA change is zero\. Context\-QA changes are exact\-match percentage points; summary and open\-QA changes are chrF points\. These are descriptive results, and some language cells are small\.

Table 5:Changes for the shared routes\.SSdenotes summarization \(Δ\\DeltachrF\), andCCdenotes context QA \(Δ\\DeltaEM in percentage points\)\.Table 6:Open\-QA chrF changes\. The primary system uses Long\-form QA; variant A uses Best\-score QA\. Variant B is the multitask reference and has zero change\.