Think Fast and Slow: Step-Level Cognitive Depth Adaptation for LLM Agents
This new AI agent framework learns when to think fast or slow, crushing benchmarks.
Deep Dive
Researchers introduced CogRouter, a framework that trains LLM agents to dynamically adapt their 'cognitive depth' for each step in a task, instead of using a fixed reasoning pattern. Grounded in ACT-R theory, it uses a two-stage training process. In tests, a 7B-parameter model achieved an 82.3% success rate, outperforming GPT-4o by 40.3% and OpenAI's o3 model by 18.3%, while using 62% fewer tokens on ALFWorld and ScienceWorld benchmarks.
Why It Matters
It makes AI agents vastly more efficient and capable, a major step toward practical, autonomous AI assistants.