EVENT DOSSIER
To address the issue of insufficient inference capabilities in low-resource languages for multilingual large models, researchers proposed a training-free testing framework called Reliability-Aware Adaptive Inference (RAAI). This framework utilizes LLaMA and Qwen models for experiments and achieves dynamic allocation of computing resources through ECE routing, prompt fusion, and a risk index (RI) gating mechanism. Results show that RAAI improved accuracy by 25-37.7% in low-resource languages and reduced calibration errors by approximately 3-6%. The study confirms that while English context helps in understanding and correcting errors, redundant bilingual contexts can increase interference; RAAI effectively addresses this trade-off issue, significantly enhancing the model’s reliability in low-resource languages.
研究人员针对多语言大模型在低资源语言中推理能力不足的问题，提出了一种无需训练的测试时框架 Reliability-Aware Adaptive Inference (RAAI)。该框架利用 LLaMA 和 Qwen 模型进行实验，通过预期校准误差（ECE）路由、提示融合及中层风险指数（RI）门控机制，实现了计算资源的动态分配。结果显示，在低资源语言上 RAAI 将准确率提升了 25-37.7%，并将校准误差降低了约 3-6%。研究证实，虽然英文上下文有助于理解并纠正错误，但冗余的双语上下文会加剧干扰；RAAI 有效解决了这一权衡问题，显著提升了模型在低资源语言下的可靠性。