The ESTS team participated in the WMT26 Model Compression Shared Task, submitting six different compression operating points for English--Simplified Chinese and English--Egyptian Arabic. These submissions utilized a GPT-OSS-20B base model. The core technique involved routing-informed expert pruning, leveraging task-specific routing mass to rank experts and cross-lingual routing divergence to determine capacity allocation. Subsequently, low-importance experts were physically removed. The resulting specialist models were recovery-tuned using synthetic translation data generated by GPT-5.1 and further compressed through MXFP4 quantization of the expert projection weights.
Across these six submissions, the achieved parameter counts ranged from 4.186 billion to 7.770 billion, while the packed artifact sizes spanned from 4.55 GiB to 6.33 GiB. Internal evaluations, conducted using GPT-5.1 pseudo-references via xCOMET-XL, provided a comparative assessment of the different compression operating points. This allowed for a structured evaluation of the compression strategy’s impact on model performance.
The team implemented a robust inference system designed for the instruction-conditioned WMT26 setting. This system included category inference, output validation, retry mechanisms, segmented fallback strategies, and source-owned JSON reconstruction. This comprehensive system facilitated reliable and accurate translation outputs.
The results demonstrate a viable approach to model compression, achieving significant reductions in model size while maintaining reasonable performance. The combination of expert pruning, recovery tuning, and quantization represents a practical strategy for deploying large language models in resource-constrained environments. Source: https://arxiv.org/abs/2609.12310
