SRAT: RL Framework Teaches LLMs When to Retrieve for Better Machine Translation
_reachsumit · x · 2026-10-15
A new arXiv paper introduces SRAT (Selective Retrieval-Augmented Translation), an agentic reinforcement learning framework for machine translation.
Key insight: translation operates in a "partial-evidence" regime—LLM parametric knowledge suffices for many inputs, while external evidence (idiom interpretations, terminology) only helps a subset and typically supports rather than directly provides the translation. Existing retrieval-augmented MT systems apply uniform predefined retrieval strategies to all inputs.
Method highlights:
- RL learns three decisions: when to translate directly, what to retrieve, and how to integrate evidence—optimizing final translation quality end-to-end
- A new retriever-training approach aligns evidence retrieval with downstream translation quality rather than semantic similarity, via joint Plackett–Luce optimization and translation-quality contrastive training
- Validated on idiom, knowledge-intensive, and document-level translation with offline and online retrieval
More from Research
- kalomaze argues ensembled synthetic data can cover data-scarce domains — kalomaze · 2026-10-15
- Axiom claims proof of generalized quasi-Riemann hypothesis for all Dedekind zeta functions — basedjensen · 2026-10-15
- TorchLean expands into a verified tensor programming framework with einsum and GPU execution in Lean — burny_tech · 2026-10-15
- StreamErase hits 26 FPS real-time video object erasing on a single GPU, 200x faster than prior methods — Zhuobai Dong · 2026-10-15
- Huawei's SwarmRSI: evolving agent swarms achieve recursive self-improvement in science — Qianru Meng · 2026-10-15
- Glance Forcing cuts video distillation trajectory cost from thousands of GPU-hours to one — Zhuobai Dong · 2026-10-15