Beyond Translation: Why Multilingual RAG Requires Bilingual Reasoning Architectures

Retrieval-Augmented Generation has become the dominant paradigm for grounding large language models in external knowledge. Yet most RAG architectures harbor a subtle but critical assumption: that reasoning is best conducted in English, with other languages treated as secondary inputs requiring translation before processing. This assumption collapses under the complexity of multi-hop question answering, where questions require synthesizing information across multiple documents through chains of inference.

In their paper "DaPT: A Dual-Path Framework for Multilingual Multi-hop Question Answering," Wang et al. expose the fragility of the standard translate then retrieve pipeline and propose an alternative that treats source language and English as parallel reasoning channels. Their work reveals that current state of the art RAG systems suffer significant performance degradation when handling multilingual queries, with accuracy drops particularly severe for low resource languages. By generating sub question decomposition graphs in parallel for both the source query and its English translation, then fusing these structures before retrieval, DaPT achieves an 18.3% relative improvement in Exact Match scores on the challenging MuSiQue benchmark.

The Decomposition Gap: Why Structure Does Not Survive Translation

The central insight of DaPT lies in recognizing that semantic decomposition is language specific. When a complex question requires multiple hops of reasoning, the way that question breaks down into sub questions depends on linguistic structures, cultural knowledge frames, and lexical granularity specific to that language. A question in Thai might decompose into three distinct sub queries based on classifiers and honorifics that have no direct English equivalent, while a German compound noun might compress what English expresses as a multi hop relationship.

DaPT addresses this through its Parallel Planning stage. Rather than translating the query and then decomposing, the framework uses the LLM to generate two separate directed acyclic graphs simultaneously: one for the source language query and one for its English translation. Each graph G = (V, E) represents sub questions as nodes and logical dependencies as edges. The innovation comes in the Node Fusion stage, where embeddings of sub question nodes from both graphs are compared for semantic similarity. Nodes with high cosine similarity are merged into bilingual nodes, while divergent nodes are preserved, creating a fused graph that captures the unique reasoning pathways available in each language.

This approach acknowledges that translation is not a lossless compression of meaning but a transformation that can alter epistemic structure. By maintaining both graphs through the decomposition phase, DaPT preserves language specific reasoning chains that would be erased by naive translation.

Benchmarking Multilingual Multi Hop Reality

Wang et al. identify a critical gap in the evaluation landscape: the absence of standardized benchmarks for multilingual multi hop QA. To remedy this, they constructed three new evaluation sets by translating HotpotQA, 2WikiMultiHopQA, and MuSiQue into five languages spanning diverse typological families: Swahili, Thai, German, Spanish, and Chinese. This selection intentionally includes both high resource languages (German, Spanish) and low resource languages with complex morphosyntax (Thai, Swahili).

The results reveal a stark performance imbalance in existing RAG systems. Advanced systems like HippoRAG2, which perform robustly in English, show significant degradation when processing queries in Thai or Swahili. On MuSiQue, the most challenging benchmark featuring adversarial distractors and requiring up to four hops of reasoning, DaPT achieves a 4.3 F1 point improvement over HippoRAG2 specifically for Thai queries. This gap underscores how current systems' reliance on English centric semantic understanding creates a two tiered user experience, where non English speakers receive inferior reasoning quality.

The Exact Match (EM) metric improvements are particularly telling. DaPT's 18.3% relative EM improvement on MuSiQue indicates not just better retrieval but more precise answer formulation. The bilingual retrieval strategy, which gathers