Research Papers 论文研究 6h ago Updated 1h ago 更新于 1小时前 46

ESTS at WMT26: Routing-Informed Expert Pruning for Model Compression ESTS在WMT26:基于路由的专家剪枝模型压缩

ESTS team submitted six compression variants to WMT26 Model Compression Shared Task for En-Zh and En-Arabic translation, all derived from GPT-OSS-20B Novel routing-informed expert pruning approach uses task-specific routing mass to rank experts and cross-lingual routing divergence to allocate retained capacity across layers Pruned models undergo recovery tuning on GPT-5.1-generated synthetic data followed by MXFP4 quantization of retained expert projection weights Final models range from 4.186B ESTS团队在WMT26模型压缩共享任务中提交6个英中/英埃阿拉伯语翻译模型,基于GPT-OSS-20B压缩至4.186B-7.770B参数 提出路由感知专家剪枝方法:利用任务特定路由质量排序专家、跨语言路由差异分配保留容量,物理移除低重要性专家 采用GPT-5.1生成合成翻译数据进行恢复微调,并结合MXFP4量化进一步压缩专家投影权重 实现完整的健壮推理系统,支持类别推理、输出验证、重试、分段回退和JSON重建 打包后模型体积仅4.55-6.33 GiB,内部xCOMET-XL评估验证了不同压缩点的质量

60
Hot 热度
72
Quality 质量
68
Impact 影响力

Analysis 深度分析

TL;DR

  • ESTS team submitted six compression variants to WMT26 Model Compression Shared Task for En-Zh and En-Arabic translation, all derived from GPT-OSS-20B
  • Novel routing-informed expert pruning approach uses task-specific routing mass to rank experts and cross-lingual routing divergence to allocate retained capacity across layers
  • Pruned models undergo recovery tuning on GPT-5.1-generated synthetic data followed by MXFP4 quantization of retained expert projection weights
  • Final models range from 4.186B to 7.770B parameters with packed artifact sizes of 4.55 to 6.33 GiB
  • Robust inference pipeline includes category inference, output validation, retries, segmented fallback, and JSON reconstruction for instruction-conditioned settings

Why It Matters

This work demonstrates a practical, scalable approach to compressing large open-source language models for production machine translation, achieving significant parameter reduction while maintaining competitive quality. The routing-informed pruning methodology offers a principled alternative to uniform pruning strategies, potentially applicable to any MoE-based architecture. The full pipeline—from expert selection through quantization to robust inference—provides a reproducible blueprint for practitioners looking to deploy compressed models in resource-constrained environments.

Technical Details

  • Expert Pruning via Routing Analysis: Uses task-specific routing mass to rank experts by importance and cross-lingual routing divergence to determine how retained capacity should be distributed across transformer layers, then physically removes low-importance experts rather than merely masking them
  • Recovery Tuning and Quantization: Post-pruning specialists are recovery-tuned on GPT-5.1-generated synthetic translation data, then further compressed via MXFP4 quantization applied specifically to retained expert projection weights
  • Three Compression Operating Points per Direction: Six total submissions covering En-Zh and En-Arabic, each with three compression levels, enabling analysis of the quality-compression tradeoff curve
  • Robust Inference System: Implements category inference, output validation, retries, segmented fallback, and source-owned JSON reconstruction to handle the instruction-conditioned WMT26 evaluation setting reliably
  • Evaluation: Internal xCOMET-XL evaluation using GPT-5.1 pseudo-references for cross-operating-point comparison; parameter counts span 4.186B–7.770B with artifact sizes of 4.55–6.33 GiB

Industry Insight

  • Routing-based expert pruning represents a significant advancement over heuristic or uniform compression methods; practitioners should consider analyzing routing patterns in their own MoE models before applying pruning to maximize quality retention
  • The combination of synthetic data recovery tuning (using GPT-5.1) with aggressive quantization (MXFP4) demonstrates that high-quality compressed models can be produced without access to large parallel corpora, lowering the barrier for specialized translation system development
  • The full inference pipeline—including validation, retries, and fallback—highlights that model compression is only one component of production readiness; robust serving infrastructure is equally critical for meeting competition and real-world quality requirements

TL;DR

  • ESTS团队在WMT26模型压缩共享任务中提交6个英中/英埃阿拉伯语翻译模型,基于GPT-OSS-20B压缩至4.186B-7.770B参数
  • 提出路由感知专家剪枝方法:利用任务特定路由质量排序专家、跨语言路由差异分配保留容量,物理移除低重要性专家
  • 采用GPT-5.1生成合成翻译数据进行恢复微调,并结合MXFP4量化进一步压缩专家投影权重
  • 实现完整的健壮推理系统,支持类别推理、输出验证、重试、分段回退和JSON重建
  • 打包后模型体积仅4.55-6.33 GiB,内部xCOMET-XL评估验证了不同压缩点的质量

为什么值得看

本文展示了大模型压缩在机器翻译领域的最新实践,为20B级模型压缩至7B以下提供了可复现的技术路径。对于需要在资源受限环境下部署翻译系统的团队,该工作提供了从剪枝、微调到量化的完整方案参考。

技术解析

  • 路由感知专家剪枝:利用MoE架构中专家的路由质量(routing mass)对专家重要性排序,通过跨语言路由差异(cross-lingual routing divergence)在不同语言对之间动态分配保留容量,实现物理剪枝而非软剪枝
  • 恢复微调策略:剪枝后使用GPT-5.1生成的合成翻译数据进行恢复微调(recovery-tuning),弥补剪枝带来的性能损失
  • MXFP4量化:对保留专家的投影权重应用MXFP4低比特量化,进一步压缩模型体积
  • 健壮推理系统:针对WMT26指令条件设置,实现类别推理、输出验证、重试机制、分段回退和源端JSON重建等工程组件
  • 压缩效果:参数量从20B降至4.186B-7.770B,打包体积4.55-6.33 GiB,覆盖三个压缩操作点

行业启示

  • MoE架构压缩路径成熟:路由感知剪枝为MoE模型的高效压缩提供了新思路,未来可推广至其他多语言任务
  • 合成数据+量化组合策略:GPT-5.1生成数据恢复微调结合MXFP4量化的组合方案,展示了"大模型生成数据辅助小模型训练"的可行路径
  • 工程系统完整性:论文不仅关注模型压缩算法,还强调推理系统的健壮性设计,对工业界部署具有直接参考价值

Disclaimer: The above content is generated by AI and is for reference only. 免责声明:以上内容由 AI 生成,仅供参考。

LLM 大模型 Research 科学研究 Inference 推理 Quantization 量化