Meta Ai researchers present matrix: A traditional ray-traced framework for artificial agent data generation

How do you store fresh and diverse synthetic data for today's AI models without turning a single orchestration pipeline into a bottleneck? Meta Ai researchers present Ima matrixa decentralized framework in which both flow control and data flow are applied to messages traveling over distributed lines. Since LLM training is increasingly dependent on artificial intelligence, tool traces and multiple chains, many existing systems rely on a central controller or a GPU cell, adding to the coordination of data diversity. In Matrix Instead it uses peers to organize the ray Cluster and moves the top tokens to 15 to the top tokens while maintaining a comparable quality.

From central administrators to peers to peer agents

Traditional agent frameworks maintain workflow state and logical control within a centralized orchestrator. Every agent call, tool call and retry pass through that controller. This model is easy to think about, but it doesn't fit well when you need tens of thousands of communication conversations or tools.

The Matrix takes a different approach. It seralizes both control flow and data flow in an Object called Orchestrator. Orchestrator captures the state of the task, including the conversation history, intermediate results and the routing logic. Agnostic agents, launched as ray players, pull the orchestrator from the queue, use their intuitive view, update the state and send it directly to the next agent chosen by the Orchestrator. There is no intermediate schedule in the inner loop. Each task progresses independently at the line level, rather than waiting for Batch Level constraints as in Spark or Ray data.

This design reduces idle time when different trajectories have very different lengths. It also does local error handling at work. If one orchestrator fails, the batch fails.

System stack and services

Matrix runs against a collection of rays that are often introduced in slurm. Ray provides streaming players. Ray Fordas exposes Endpoints behind vllm and Sgglang, and can also route to external APIs such as Azire Openrai or Gemini through Proxy Servers.

Tool calls and other complex services run inside apptainer containers. This isolates the runtime of the agent from code execution sandboxes, HTTP tools or custom parsers. Hydra downloads the configuration of agent roles, orchestrator types, resource allocations and o or o schemas. Grafana integrates with metro metrics to track queue length, pending tasks, token usage and GPU utilization in real time.

The matrix also introduces the message payload. When the conversation history grows above the size limit, large payloads are stored in the Ray object store and only object pointers are stored in the orchestrator. This reduces the bandwidth of the cluster while allowing the agents to re-promote when needed.

Case Study 1: The gateway to collaboration

Cooperative Reason, also known as Coral, analyzes the VLM agent dialog where llm agents discuss a question, disagree when needed and reach a final answer. In the original implementation the central controller manages thousands of self collajectories. The matrix reimplements the same protocol that uses peer-to-peer orchestrators and asynchronous agents.

In 31 places of A100, using LLama 3.1 8B I teach, the matrix configures the conversation consistency as 248 gpus with 50 queries per GPU, so 12,400 parallel conversations. The coral foundation works with its 5,000 best experiences. Under the same hardware, the matrix generates about 2 billion tokens in 4 hours, while coral generates 0.62 billion tokens in about 9 hours. That's a 6.8 times increase in Token Fook with almost the same accuracy around 0.47.

Case Study 2: Natural Feb Data Calation

NaturalReasoning builds a reasoning dataset from web corpora. Matrix models with a three-agent pipeline. The filtering agent uses a small classifier model to select English passages that are likely to contain reasoning. The score agent uses a large tuned fuzzy model to provide standardized scores. A query agent issues queries, responses and discussion chains.

Of the 25 million Web documents, about 5.45 percent survived all filters, revealing Personal Answers around the steps to answer related steps. Matrix then compared different matching techniques on a 500 thousand document database. The best configuration includes data parallelism and task parallelism, with 20 data partitions and 700 parallel functions divided into each. This is about 1.61 times higher than the scale of the work.

Over a full run of 25 million documents, the matrix achieves 5,853 tokens per second, compared to 2,778 tokens per second for batch ray data with 14,000 jobs. That's like 2.1 times the follow-up that comes from peer-to-peer learning to plan for different levels, not from different models.

Study case 3, The Tau2-Bench tool uses trajectories

Tau2-Bench evaluates the communication agents who must use the tools and the database in a customer support setting. The matrix represents this environment with four agents, a user simulator, a helper, tool difference and reward calculation, and a sink that collects metrics. Tool APIS and Reward Logic are reused in Tau2 reference work and packaged.

For a cluster with 13 h100 locations and llm replicas, the matrix creates 22,800 trajectories in about 1.25 hours. Corresponding to 41,000 tokens per second. Baseline Tau2-Agent usage on a single node, configured with 500 threads, reaches approximately 2,654 tokens per trajectories and 1,519. The average reward remains almost unchanged in both systems, which ensures that the speed does not come from cutting corners in nature. In total, the matrix delivers about 15.4 high tokens in the Benchmark.

Key acquisition

• Matrix replaces peer-to-peer integrated orchestrators, a message-driven service provider that acts as a representative state machine running through virtual agents.
• The framework is built entirely on open source, slurm, Ray, vllm, sglang and apptainer, and scales to tens of thousands of multi-agent workflows, monitoring and data processing.
• In three subtle studies, cooperative reasons, matrix, bandch, matrix moves 2 times higher to 15,4 higher token than special parts under the same hardware, while rewards.
• Matrix uploads large chat logs to Ray's Object Store and keeps only simple references to Peak bandwidth records.

Editorial notes

The matrix is ​​a pragmatic systems role that takes the production of multi-agent data from bespoke scripts to runtime. By coding flow control and data flow in orchestrators, it compresses the execution of countless P2P agents in ray, split clean, LLMs to find and tools. The case swears by the reason of the cooperation, NaturalReaving and Tau2-Bench shows that the systems that fall in the name, not the new models of Model Design, are now the main lever for measuring pipelines of artificial data.

Look Paper and repo. Feel free to take a look at ours GitHub page for tutorials, code and notebooks. Also, feel free to follow us Kind of stubborn and don't forget to join ours 100K + ML Subreddit and sign up Our newsletter. Wait! Do you telegraph? Now you can join us by telegraph.

Michal Sutter is a data scientist with a Master of Science in Data Science from the University of PADOVA. With a strong foundation in statistical analysis, machine learning, and data engineering, Mikhali excels at turning complex data into actionable findings.