Revise MoICE reference

content/11.future_trends.md

@@ -1,4 +1,4 @@
-## Future Trends and Opportunities with Foundation Models
+<img width="1712" height="928" alt="image" src="https://github.com/user-attachments/assets/219c092e-3915-4a7f-96db-7073fd898a7c" />## Future Trends and Opportunities with Foundation Models
 
 ### A New Paradigm for Context-Adaptive Inference
 Recent advances in large-scale foundation models have fundamentally reshaped the landscape of context-adaptive inference. Trained on vast and diverse datasets with self-supervised objectives, these models internalize broad statistical regularities across language, vision, and multimodal data [@doi:10.48550/arXiv.2108.07258]. Unlike earlier approaches that relied on hand-crafted features or narrowly scoped models, foundation models can process and structure complex, high-dimensional contexts that were previously intractable.
@@ -69,6 +69,6 @@ Several new architectures exemplify how foundation models advance context-sensit
 
 **LMPriors** (Pre-Trained Language Models as Task-Specific Priors) [@doi:10.48550/arXiv.2210.12530] leverages semantic insights from pre-trained models like GPT-3 to guide tasks such as causal inference, feature selection, and reinforcement learning. This method markedly enhances decision accuracy and efficiency without requiring extensive supervised datasets. However, it necessitates careful prompt engineering to mitigate biases and ethical concerns.
 
-**Mixture of In-Context Experts** (MoICE) [@doi:10.48550/arXiv.2210.12530] introduces a dynamic routing mechanism within attention heads, utilizing multiple Rotary Position Embeddings (RoPE) angles to effectively capture token positions in sequences. MoICE significantly enhances performance on long-context sequences and retrieval-augmented generation tasks by ensuring complete contextual coverage. Efficiency is achieved through selective router training, and interpretability is improved by explicitly visualizing attention distributions, providing detailed insights into the model's reasoning process.
+**Mixture of In-Context Experts** (MoICE) [@doi:10.48550/arXiv.2406.19598] introduces a dynamic routing mechanism within attention heads, utilizing multiple Rotary Position Embeddings (RoPE) angles to effectively capture token positions in sequences. MoICE significantly enhances performance on long-context sequences and retrieval-augmented generation tasks by ensuring complete contextual coverage. Efficiency is achieved through selective router training, and interpretability is improved by explicitly visualizing attention distributions, providing detailed insights into the model's reasoning process.
 
 Collectively, these directions suggest a future in which foundation models evolve from passive representation learners into active, context-sensitive inference engines that unify adaptivity, efficiency, and interpretability within a principled framework.