Sakana Conductor is a 7 billion parameter language model developed by Sakana AI that exemplifies AI-managing-AI approaches through dynamic task routing and orchestration of frontier model pools via natural language instructions. The system represents an advancement in test-time scaling methodologies, where computational resources are allocated adaptively based on task complexity rather than applying uniform model capacity across all inference requests.
Sakana Conductor operates as a routing and orchestration system trained with reinforcement learning (RL) to manage multiple frontier models as a coordinated pool. Rather than relying on a single monolithic model to handle all inference tasks, the system accepts natural language specifications and dynamically determines optimal task allocation across available models. This approach addresses fundamental constraints in model scaling by leveraging test-time computation—the idea that reasoning resources can be allocated dynamically during inference rather than being fixed at training time.
The 7B parameter scale of Sakana Conductor enables it to serve as a lightweight orchestration layer, making decisions about which downstream models to invoke and how to manage context exposure across the model pool. The system learns through RL to optimize for both performance metrics and computational efficiency, balancing solution quality against resource consumption.
Sakana Conductor demonstrates significant performance improvements through orchestrated inference. On LiveCodeBench, a benchmark designed to evaluate code generation capabilities across realistic programming tasks, the system achieves 83.9% performance—a result that exceeds the performance of individual frontier models evaluated independently on the same benchmark 1).
On GPQA-Diamond, a challenging graduate-level physics and science benchmark, Sakana Conductor reaches 87.5% accuracy, again surpassing individual model performance baselines. These results indicate that dynamic routing and coordinated inference across a model pool can achieve emergent capabilities that exceed what any single model in the pool provides independently.
The development of Sakana Conductor reflects broader research into test-time scaling in large language models. Rather than increasing model parameters—which requires substantial training and inference infrastructure—test-time scaling allocates additional computation during the inference phase itself. This can take multiple forms, including chain-of-thought prompting 2)—where models generate intermediate reasoning steps—or through ensemble approaches that combine multiple model outputs.
Sakana Conductor implements a more sophisticated variant of test-time scaling by training a routing model to dynamically select which specialized models to invoke and how to structure their collective computation. This approach allows systems to allocate maximum compute to the most challenging problems while using lightweight models for straightforward tasks, effectively creating a dynamic inference budget that adapts to task difficulty.
Sakana Conductor exemplifies an emerging paradigm where AI systems actively manage and coordinate other AI systems. Rather than humans manually deciding which model to invoke for a given task, the system makes these decisions autonomously based on natural language specifications and learned routing policies. This automation of model selection has several implications:
- Scalability: Systems can efficiently manage larger pools of specialized models without requiring proportional increases in human oversight - Adaptive Optimization: Routing decisions can adapt to variations in input complexity, model availability, or performance requirements - Emergent Capabilities: Coordination across diverse models can produce capabilities exceeding individual model performance
This approach connects to broader research in AI orchestration and hierarchical reasoning, where higher-level AI systems guide the behavior of lower-level specialist systems 3).
The dynamic routing and task orchestration capabilities of Sakana Conductor suggest applications across domains requiring adaptive computational allocation:
- Code generation and software development: Where task complexity varies significantly between different programming challenges - Scientific question answering: Complex reasoning tasks that may benefit from specialized models trained on domain-specific data - Multi-step problem solving: Where different phases of a problem may best utilize different models - Resource-constrained deployment: Where inference budgets are fixed and must be allocated optimally across variable-complexity tasks
Sakana Conductor demonstrates that frontier-class performance on challenging benchmarks can be achieved through intelligent orchestration of existing models rather than solely through scaling individual model parameters. As the computational cost of training increasingly large models continues to rise, approaches that optimize inference-time allocation of compute across model ensembles or specialized pools may become increasingly important for cost-effective deployment.
The system's performance improvements over individual models suggest that the coordination problem—determining optimal task allocation across a heterogeneous model pool—represents a tractable optimization target. Future developments in this area may focus on expanding the diversity of models in managed pools, improving the natural language interface for task specification, and extending these orchestration principles to multimodal tasks beyond code and language understanding.