New Framework Routes Agent Compute Dynamically
A research paper titled "Think Fast and Slow" introduces COGROUTER, a system that dynamically routes computational resources for LLM agents based on task complexity. The framework is inspired by the ACT-R cognitive architecture theory. It aims to optimize agent performance by allocating more resources for difficult steps and fewer for simpler ones, mimicking human cognitive adaptation.
- In benchmark tests on ALFWorld and ScienceWorld, COGROUTER using the Qwen2.5-7B model achieved an 82.3% success rate, which surpassed GPT-4o's performance by 40.3% while consuming 62% fewer tokens. - The framework trains agents to adapt their "cognitive depth" through four hierarchical levels, from simple instinctive responses to more complex strategic planning, a concept directly grounded in the ACT-R cognitive theory. - COGROUTER employs a two-stage training method: Cognition-aware Supervised Fine-tuning (CoSFT) to establish stable cognitive patterns, followed by Cognition-aware Policy Optimization (CoPO) to refine decision-making at each step. - The concept of dynamically routing LLM queries is a broader architectural pattern for efficiency, with other tools like Martian Learning's router demonstrating up to a 30% increase in success rates for multi-step financial workflows. - This dynamic routing approach contrasts with more static multi-agent orchestration frameworks like CrewAI or Microsoft's AutoGen, which primarily focus on dividing tasks among specialized agents rather than adapting the computational load of a single agent. - The underlying principle of integrating cognitive architectures like ACT-R with LLMs is an active area of research, with related projects like LLM-ACTR aiming to make AI decision-making more human-aligned and explainable, particularly in complex domains like manufacturing. - In the Chinese market, major platforms are emerging for agent deployment, such as Alibaba's DingTalk, which launched a marketplace with over 200 AI agents, signaling a growing ecosystem for consumer-facing agent products. - The AI agents market in China is projected to grow at a compound annual growth rate of 50.8% between 2026 and 2033, with major tech players like Tencent (Hunyuan), Baidu (Wenxin), and Ant Group (Lingji) all developing agent platforms.
