Hyra - A scientific discovery intelligent agent launched by Tencent Hunyuan

What is Hyra?

Hyra (Hunyuan Research Agent) is a recursive, self-improving research agent launched by Tencent Hunyuan. The first version, Hyra-1.0, has been officially released. Hyra uses the lightweight Harness framework, generating inspirational contexts by maintaining an experience bank through a Context Agent. Multiple Proposal Agents asynchronously consume queues, run and score solutions in independent sandboxes, and continuously iterate to optimize solutions. Hyra supports a two-layer loop: an inner layer for optimizing solutions and an outer layer for evolutionary evaluators, covering more than ten scenarios including AI research and development, scientific discovery, and game design.

Hyra's main functions

• Recursive self-improvementBased on historical experience, we continuously explore better solutions and accumulate successes and failures into a knowledge base.
• Asynchronous producer-consumer pipelineThe Context Agent organizes the inspiration context to fill the queue, and the Proposal Agent consumes and executes it, ensuring efficient and full utilization of resources.
• Double loop mechanismThe inner loop optimizes the solution based on the evaluator; the outer loop upgrades the evaluator by combining experience history, realizing the co-evolution of evaluator and solution.
• Multi-scenario coverageSupports AI for AI (model training, GPU kernel optimization), AI for Science (mathematical discovery, drug design, astronomical prediction), and AI for Fun (game bots, 3D modeling, music arrangement).
• Sandbox Security ExecutionEach solution runs in an independent sandbox, with solve.sh as the entry point, ensuring isolation and reproducibility.

Hyra's technical principles

• Harness DesignFollowing the principle of "The Bitter Lesson," the framework is lightweight and offers ample room for maneuver. Given a task description, it continuously runs an exploration-evaluation-improvement cycle.
• Experience BankIt stores multimodal experiences such as source code, runtime logs, and evaluator feedback from past solutions.
• InspirationsContext Agent integrates diverse contexts from an experience base to inspire the creation of better solutions in the future.
• Semaphore resource control: Ensure that resources such as sandbox, model inference, and task queue are healthy and fully utilized, and that solution quality scales efficiently over time.
• Evaluator co-evolutionFor tasks without a ready-made evaluator, Hyra first designs an initial version of the evaluator, optimizes the inner loop, and then upgrades the evaluator in the outer loop to start the next round.

How to use Hyra

• Define taskProvides a task description and an optional initial evaluator.
• Start HarnessHyra automatically initializes the Context Agent and multiple Proposal Agents.
• Asynchronous executionThe Context Agent continuously generates inspiration contexts and adds them to the queue; the Proposal Agent retrieves the context and writes the solution (including the solve.sh entry point).
• Sandbox Execution RatingThe solution runs in a separate sandbox, and the results are fed back to Experience Bank.
• loop iteration: Continuously explore based on historical best practices until voluntary withdrawal or budget exhaustion.
• Obtain the optimal solutionReturns to the best historical solution and the complete experience base.

Hyra's core advantages

• Simple and effectiveThe framework is lightweight, does not rely on complex orchestration, and provides a wide range of action options for the agents.
• High versatilityThe same Harness framework spans more than ten scenarios, including training algorithms, system optimization, scientific discovery, and creative generation.
• The evaluator can evolveThe double-loop approach addresses the issue of missing or easily hackable evaluation criteria in open tasks.
• Real-world scenario verificationLearn and evolve within Tencent's product systems, real AI R&D pipelines, and industrial scenarios.
• Open source and transparentAll experimental products and results are open source, facilitating reproduction and academic collaboration.

Hyra's project address

• Project official website：https://hy.tencent.com/research/hyra

Hyra's Competitive Comparison

Hyra's application scenarios

• AI Basic Model DevelopmentAutomatically optimizes training recipes, data strategies, inference systems, and underlying GPU kernels, setting new records on NanoChat/NanoGPT/SOL-ExecBench.
• Mathematical Science DiscoveriesSolving open mathematical problems from databases such as EinsteinArena, achieving 29 new best results out of 55 problems.
• Astronomical time series predictionBy analyzing historical observational data such as sunspots, a recursive formula was discovered and a long-term prediction model was established. The out-of-sample validation over the past century has yielded high accuracy.
• Drug molecular design: To generate candidate molecules that bind stably to anticancer targets such as PARP1 and have better druggability than the marketed inhibitor olaparib.
• Quantum computing optimizationDesign a bit routing algorithm for quantum chips such as IBM Q20, which reduces two-bit gate overhead by 44.4% and noise accumulation compared to the classic SABRE algorithm.