Yichao Wu

Forget scaling up; this study shows that the right reasoning architecture—structured memory for deterministic tasks, RAG for conversational ones—can dramatically outperform a baseline LLM even when compute is severely limited.

LLMs can significantly boost their emotional intelligence simply by role-playing conversations with themselves, iteratively refining their ability to both recognize and express emotions.

LLMs can now autonomously retrieve relevant memories from a database using specialized tools, significantly improving performance on long-term conversational question answering.

Slash RAG latency by an order of magnitude using a tiny, LoRA-adapted SLM that routes queries, achieving GPT-4o-mini level accuracy at a fraction of the cost.