Abstract: Open-domain question answering (ODQA) typically involves retrieving multiple relevant documents from large-scale corpora and leveraging large language models (LLMs) to comprehend the content and generate answers. However, in low-resource languages such as Burmese and Lao, the retrieved documents often contain irrelevant noise, and LLMs exhibit limited understanding capabilities and generate more errors. To address these challenges, this paper proposes a multi-stage ODQA framework for low-resource languages based on multi-dimensional answer filtering. Instead of generating answers directly from documents, our approach first extracts diverse candidate answers, followed by a multi-dimensional filtering process. This includes global discourse-level validation, local evidence verification, and inter-answer relevance ranking to identify the optimal answer. Experiments on four Southeast Asian low-resource ODQA datasets demonstrate that our method, built upon LLMs such as GPT-4o-mini and DeepSeek-v3, outperforms state-of-the-art techniques including chain-of-thought reasoning and summary verification.