RSI-router is a framework that improves the cost-performance trade-off for LLM agents by assigning different models to different subtasks within a long-horizon task. It uses : first mining subtask definitions from training trajectories, then evolving routing strategies and model-specific skills through comparing routed versus large-model-only trajectories. In tests across five benchmarks, routing between a large model (DeepSeek-V4.1-Flash) and a small model (Qwen3.5-9B) consistently outperformed using only the large model while cutting inference cost to about half (48.3%). On three benchmarks (ALFWorld, ScienceWorld, WebShop), it actually improved performance while cutting cost by 74.7-82.2%; on Terminal-Bench 2.0 it improved performance by 16.7% at 18.0% lower cost. The resulting routers also form a better Pareto frontier than nine alternative routing methods.
Subtask routing halves agent inference cost while improving performance
RSI-Router learns when to assign subtasks to large or small language models, along with reusable model-specific execution guidance.
Academic
Hao Li · Hangfan Zhang · Zhiyao Cui · Chunjiang Mu · Yiqun Zhang · Bo Zhang · +2 more
Shanghai Artificial Intelligence Laboratory
Research Digest··2 min read
In one line
RSI-router achieves cheaper, better agent performance by routing subtasks to different LLMs and evolving skills.
Before you read this
Three ideas this paper builds onmodel routingDirecting each query or subtask to the most suitable model to optimize cost, accuracy, or quality.89 papers →recursive self-improvementAn AI system repeatedly uses its own outputs to enhance its capabilities, creating a feedback loop for autonomous improvement.66 papers →skill evolutionThe process of developing, refining, or optimizing reusable capabilities in LLM agents through learning, adaptation, or experience accumulation.42 papers →
Read it as
Li and colleagues developed RSI-Router, which iteratively learns subtask definitions, model assignments and execution skills from accumulated agent trajectories.
The authors built a router that assigns each emerging subtask within a long agent trajectory to either DeepSeek-V4.1-Flash or Qwen3.5-9B.
How it works
before and afterThe Daily Edition
The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.