Gensyn Unveils RL Swarm Framework for Team-Based Reinforcement Learning, Targeting Testnet Launch in March

Network for machine intelligence, Gensyn Gensyn has introduced RL Swarm, a decentralized peer-to-peer framework aimed at fostering collaborative reinforcement learning online. Next month, a testnet is expected to launch, allowing broader engagement in the evolution of open machine intelligence.

RL Swarm is a completely open-source initiative that allows reinforcement learning models to train collaboratively across distributed networks. This system provides a real-time showcase of research findings that demonstrate how models utilizing RL can enhance their learning effectiveness when trained collaboratively instead of individually.

By operating a swarm node, participants can either create a new swarm or link to an existing one via a public address. Inside each swarm, models participate in reinforcement learning collectively, employing a decentralized communication protocol grounded in Hivemind for knowledge exchange and enhancement of the models. By using the provided client software, individuals can engage in a swarm, observe collective updates, and locally train models while harnessing collective intelligence. In the future, new experiments will be rolled out, inviting more participants to advance this technology.

Everyone is welcome to join RL Swarm and discover how the system works firsthand. Accessing it is straightforward, whether through regular consumer-grade hardware or high-performance cloud-based GPU resources.

How RL Swarm Works? 

Gensyn Gensyn has long envisioned a world where machine learning operates in a decentralized manner, distributed across a comprehensive network of devices. Instead of depending on large, centralized models, this strategy breaks models into smaller, interconnected units that work together harmoniously. In pursuing this vision, Gensyn has investigated various routes to decentralized learning and discovered that reinforcement learning (RL) post-training proves particularly effective when models communicate and share feedback.

Research indicates that RL models experience enhanced learning efficiency when they are trained within a collaborative swarm rather than working in isolation.

In this framework, each swarm node utilizes the Qwen 2.5 1.5B model and tackles mathematical challenges (GSM8K) through a structured, three-stage process. Initially, each model independently attempts to solve the problem, documenting its thinking and answer in a specified format. During the second stage, these models evaluate the feedback provided by their peers. Finally, each model casts a vote on the response it believes the majority will regard as the best answer, refining its solution accordingly. Through these iterative exchanges, the models collectively bolster their ability to solve problems.

Preliminary results suggest that this technique accelerates the learning journey, enabling models to produce more accurate answers on previously unseen test data with fewer training cycles.

Data visualizations via TensorBoard highlight significant patterns seen in a participating swarm node. These graphs display cyclical trends caused by periodic “resets” between rounds of joint training. The x-axis in all graphs represents the elapsed time since the node joined the swarm, while the y-axis indicates various performance indicators. From left to right, the graphs represent: Consensus Correctness Reward, measuring how often a model correctly formats its output and produces a mathematically accurate answer; Total Reward, a composite score reflecting rule-based evaluations (including formatting, accuracy, and logical consistency); Training Loss, which indicates how the model adjusts based on reward signals to refine its learning; and Response Completion Length, which monitors the token count in responses—showing that models become more succinct as they receive peer feedback.