Research map / Large language models / Retrieval and knowledge
Exposing the Cracks: Vulnerabilities of Retrieval-Augmented LLM-based Machine Translation
Yanming Sun, Runzhe Zhan, Chi Seng Cheang, Han Wu, Xuebo Liu, Yuyao Niu, Fengying Ye, Kaixin Lan, Lidia S. Chao, Derek F. Wong
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace gold meaning with opposite-meaning retrieved context | Qwen2.5-7B-Instruct · Idiom translation avg 10 pairs | Fidelity: 2.3 -> 0.8 | Hurt |
| replace gold meaning with structure-perturbed gold meaning | Qwen2.5-7B-Instruct · Idiom translation avg 10 pairs | Fidelity: 2.3 -> 2.1 | No clear effect |
| add literal-translation noise instead of no context | Qwen2.5-7B-Instruct · Idiom translation avg 10 pairs | Fidelity: 1.2 -> 1.4 | No clear effect |
| use thinking mode instead of non-thinking | Qwen3-8B · Idiom translation, opposite-meaning noise | Fidelity: 0.8 -> 0.4 | Hurt |
| use thinking mode instead of non-thinking | Qwen3-8B · Idiom translation, semantic-perturbed noise | Fidelity: 1.1 -> 0.9 | Hurt |
| add ALI fine-tuning (opposite-meaning noise training) | Qwen2.5-7B-Instruct · Fr→En, opposite noise | Fidelity: 0.7 -> 1.8 | Improved |
| add ALI fine-tuning | Qwen2.5-7B-Instruct · Hi→En, opposite noise | Fidelity: 0.3 -> 0.8 | Improved |
| add ALI fine-tuning | Qwen2.5-7B-Instruct · Hi→En, gold meaning | Fidelity: 2.1 -> 1.7 | Hurt |
| add CK-PLUG training-free decoding | Qwen2.5-7B-Instruct · Fr→En, opposite noise | Fidelity: 0.7 -> 1.3 | Improved |
| add CK-PLUG training-free decoding | Qwen2.5-7B-Instruct · Hi→En, gold meaning | Fidelity: 2.1 -> 1.5 | Hurt |
| use CDA instead of ALI fine-tuning (ablation) | Qwen2.5-7B-Instruct · Hi→En, opposite noise | Fidelity: 0.7 vs 0.8 | No clear effect |
| use ALI fine-tuning, evaluated on syntactic noise | Qwen2.5-7B-Instruct · Fr→En, Nstruct | Fidelity: 2.3 -> 2.2 | No clear effect |
Closest papers by meaning
- A Paradigm Shift in Machine Translation: Boosting Translation Performance of Large Language Models (ICLR 2024 · similarity 0.91)
- Are Large Reasoning Models Good Translation Evaluators? Analysis and Performance Boost (NeurIPS 2025 · similarity 0.89)
- Long-Context LLMs Meet RAG: Overcoming Challenges for Long Inputs in RAG (ICLR 2025 · similarity 0.89)
- DelTA: An Online Document-Level Translation Agent Based on Multi-Level Memory (ICLR 2025 · similarity 0.89)
- Language Ranker: A Metric for Quantifying LLM Performance Across High and Low-Resource Languages (AAAI 2025 · similarity 0.89)
- Can LLMs Really Learn to Translate a Low-Resource Language from One Grammar Book? (ICLR 2025 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
