📄 AIDA Agents: A Multi-Agent Translation Platform with Context-Aware Quality Control

Summary

Localization teams in industry must choose between NMT engines that need costly fine-tuning and heavy post-editing, and LLMs that translate well but do not fit existing CAT/TMS workflows. AIDA Agents, from DATAmundi, addresses this with a pipeline of specialised LLM agents. Up to three translator agents draft candidates. An Arbiter/Rater scores them on MQM-inspired criteria (accuracy, fluency, terminology, style) and picks the best. A Post-Editor Agent revises that choice, and a second rating pass produces the final output. Optional RAG adds translation memories, glossaries and style guides at every step, so the system adapts to a domain without fine-tuning. Files go in and out as XLIFF, with their structure preserved.

On WMT24++ (11 English→X pairs, no TM or terminology), AIDA scores the highest BLEU on 10 of 11 languages. On a small industrial IT benchmark, linguists judged most AIDA MultiLLM output publishable without human editing. The work shows that agent orchestration can compete with commercial MT and partly replace post-editing. The evaluation is preliminary, however, and is described in the paper as such.

Key points

• Three product tiers. AIDA LLM uses 3 calls per segment. AIDA MultiLLM uses up to 6 calls, with parallel translators, a rater, a post-editor and a re-rater. AIDA NMT uses 3 calls: an NMT engine plus an LLM post-editor and an LLM rater.
• WMT24++ results. The test set has 960 segments in 4 domains. AIDA has the top BLEU in 10/11 languages, with the largest margin on Japanese (+3.9). Swahili is the exception, where it trails GPT-4o by 1.3.
• Industrial benchmark setup. It covers 96 segments per language in 6 languages (JA, ZH, FR, KO, DE, ES) of technical IT documentation, with full TM, terminology and style guides. Three linguists per language rated output blindly on a 5-point scale.
• Industrial benchmark results. AIDA MultiLLM had the highest BLEU in all six languages. Between 69.4% and 98.0% of its segments were rated Good/Excellent.
• Comparison with human post-editing. AIDA's Good/Excellent share beat the human-post-edited competitor only for German (98.0% vs. 91.9%). The conclusion claims several language pairs, but Table 2 supports only German. The proposed 50–70% cut in post-editing effort is an estimate, not a measured result.