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.
Figures & tables
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.
Adapter
Rows
Context
Main sources and construction
Multitask adapter
20,293
8,192
Provided MIST sample training split across all three tasks
Summarization
12,000
8,192
6,000 general examples from the provided data, CrossSum, WikiLingua, UPDESH, and targeted language pools; 6,000 ACL papers paired with author abstracts
Context QA
8,512
4,096
Belebele, answerable TyDi QA, MLQA, MCIF, UPDESH cultural multihop data, and small Czech/Yoruba Aya pools
Open QA
6,400
2,048
Provided open-QA examples plus filtered Aya and WMT25 MIST examples; includes Hindi questions and answers translated into Bhojpuri
Table 1: Supervised data used by the selected task adapters. Each is initialized from the multitask adapter.
Task
Training variant
n
EM
chrF
ROUGE-L
LaBSE
Summarization
Multitask adapter
1,651
0.00
22.97
0.1571
0.6681
12k general/scientific mix
1,651
0.00
26.81
0.1780
0.6919
Context QA
Multitask adapter
1,517
66.25
77.07
0.6834
0.8813
8.5k context mixture
1,517
68.89
78.58
0.6959
0.8897
+4.8k targeted continuation
1,517
68.36
78.36
0.6966
0.8906
Open QA
Multitask adapter
794
3.90
28.34
0.2287
0.6524
Table 2: Internal model-selection results on identical held-out rows. EM is a percentage and LaBSE is cosine similarity.
Model
chrF
ROUGE-L
LaBSE
Multitask adapter
13.27
0.0742
0.6556
8k summary mix
25.28
0.1812
0.6959
12k summary mix
29.76
0.2082
0.7535
Table 3: Results on the scientific summarization validation set ( n=336 ). The summary mixtures are evenly divided between general and scientific data.
System
Summary
Context
Open
PSK-primary
12k mix
8.5k mix
Long-form
PSK-variant_A
12k mix
8.5k mix
Best-score
PSK-variant_B
12k mix
8.5k mix
Multitask
Table 4: Submitted routes. The open-QA variants are named for their selection criterion.
Appendix figures & tables2 assets
Supplementary material from the paper’s appendix.
Appendix
Language
nS
ΔS
nC
ΔC
Arabic
85
+4.39
90
+4.44
Bengali
66
+8.18
20
+10.00
Bhojpuri
14
+26.89
–
–
Czech
14
+12.80
–
–
Central Kurdish
14
+9.23
45
-2.22
German
34
+5.97
93
+3.23
Appendix
Table 5: Changes for the shared routes. S denotes summarization ( Δ chrF), and C denotes context QA ( Δ EM in percentage points).
Language
n
Primary Δ chrF
Var. A Δ chrF
Arabic
52
+0.22
-0.16
Bengali
11
+0.15
-0.72
Czech
11
+1.53
+1.62
German
43
-2.02
-1.34
English
11
-0.18
+1.82
French
45
-1.89
-0.79
Appendix
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.