摘要：RAG 系统中使用的语言模型和嵌入模型，通常被认为需要大规模预训练以及显式的事实归因监督。我们提出 B1ade，这是一种高效 RAG 架构，由两个专门构建的组件组成：紧凑型嵌入模型和专用 SLM。B1ade-embed 是一个拥有 3.35 亿参数的检索模型，通过对五个预训练编码器进行无参数融合构建而成，无需任何额外训练，便在参数量低于 5 亿的模型中取得领先的 MTEB 分数。B1ade-1B 是一个 SLM，使用低成本 GPU，基于 7.23 亿 token（220 万个样本）的精选上下文—问题对，通过 Group Relative Policy Optimization（GRPO）训练；其奖励仅优化答案相似度。我们的核心发现是归因行为的涌现：尽管没有接受针对来源引用的显式监督，B1ade-1B 仍会在 42.4% 的回答中引用检索到的段落，较其训练数据分布中的归因率高出 5.5 个百分点。这表明，在 RL 训练中，事实落地行为可作为一种最大化准确率的策略自然涌现，而无需显式的奖励工程。在标准问答基准上，B1ade-1B 在 PopQA 上达到 81.82%，在 PubMedQA 上达到 65.8%，在 FEVER 上达到 51.09%。在端到端 RAG 评估中，B1ade-1B 在正确性、完整性、连贯性和忠实性维度上的平均得分达到 0.654，较 SFT 提升 10.8%，并缩小了与其参数量 1.5 倍模型之间的差距。这些结果表明，无需大规模预训练，战略性的模型组合与奖励设计便足以实现资源高效的 RAG。

Abstract:Language and embedding models used in RAG systems are conventionally assumed to require large-scale pretraining and explicit grounding supervision. We present B1ade, an efficient RAG architecture comprising two purpose-built components: a compact embedding model and a purpose-built SLM. B1ade-embed, a 335M parameter retrieval model constructed via parameter-free fusion of five pretrained encoders achieves top MTEB scores among sub-500M models with zero additional training, and B1ade-1B, an SLM trained on low-cost GPUs using Group Relative Policy Optimization (GRPO) on 723M tokens (2.2M examples) of curated context-question pairs with rewards that optimize only answer similarity. Our central finding is emergent attribution: despite receiving no explicit supervision for source citation, B1ade-1B cites retrieved passages in 42.4% of responses, exceeding the attribution rate of its training distribution by 5.5 percentage points. This demonstrates that grounding behavior can emerge as an accuracy-maximizing strategy under RL training, without explicit reward engineering. On standard QA benchmarks, B1ade-1B achieves 81.82% on PopQA, 65.8% on PubMedQA, and 51.09% on FEVER. In end-to-end RAG evaluation, B1ade-1B achieves an average score of 0.654 across correctness, completeness, coherence, and faithfulness, a 10.8% improvement over the SFT, while closing the gap with models 1.5x its size. These results show that strategic model composition and reward design suffice for resource-efficient RAG, without large-scale pretraining.