Tencent's CogRouter Varies Agent Reasoning Depth
Tencent has introduced CogRouter, a framework that enables AI agents to dynamically adapt their cognitive depth, from instinctive to strategic thinking, based on task complexity. The model reportedly achieved 82.3% success on benchmarks while using 62% fewer tokens than competing frameworks. This approach aims to improve both performance and efficiency in agentic systems.
- The framework's design is grounded in the Adaptive Control of Thought-Rational (ACT-R) theory, a cognitive architecture that models human cognition. CogRouter implements four hierarchical cognitive levels, ranging from simple, instinctive responses (ℒ1) to complex, strategic planning (ℒ4). - CogRouter employs a two-stage training methodology: Cognition-aware Supervised Fine-tuning (CoSFT) to establish distinct reasoning patterns for each cognitive level, followed by Cognition-aware Policy Optimization (CoPO), a reinforcement learning algorithm that teaches the agent when to apply each level. - On the ALFWorld and ScienceWorld benchmarks, a 7B-parameter model using the CogRouter framework achieved an 82.3% average success rate. This performance reportedly surpassed larger models like GPT-4o by 40.3% and OpenAI's o3 by 18.3%. - The efficiency gains are a key feature, with the model using 62% fewer tokens compared to the GRPO baseline method on the same benchmarks. The model learns to apply deeper, more token-intensive reasoning only when task complexity increases. - The underlying base model for these published benchmarks was Qwen2.5-7B, with similar performance gains observed when using Llama3.1-8B. - This research aligns with Tencent's broader AI strategy of creating "good to use AI" that is deeply integrated into its products and cloud offerings. The company is focused on accelerating the adoption of AI agents and has also launched an agent development platform.
