Hyra - Tencent Hunyuan's AI agent for scientific discovery

What is Hyra

Hyra (Hunyuan Research Agent) is a recursive self-improving research agent launched by Tencent Hunyuan, and its first version, Hyra-1.0, has been officially released. Hyra uses a lightweight Harness framework, in which a Context Agent maintains an Experience Bank to generate inspiration contexts, while multiple Proposal Agents asynchronously consume the queue, run in independent sandboxes, and score results, continuously iterating and optimizing solutions. Hyra supports a two-level loop, with the inner loop optimizing solutions and the outer loop evolving the evaluator, covering more than ten scenarios including AI R&D, scientific discovery, and game design.

Hyra's main features

• Recursive self-improvement: Continuously explores better solutions based on historical experience, distilling successes and failures into an Experience Bank.
• Asynchronous producer-consumer pipeline: The Context Agent organizes inspiration contexts to fill the queue, and Proposal Agents consume and execute them, keeping resources efficiently saturated.
• Two-level loop mechanism: The inner loop optimizes solutions based on the evaluator; the outer loop upgrades the evaluator by combining experience history, enabling co-evolution of eval and solution.
• Multi-scenario coverage: Supports 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 orchestration).
• Safe sandbox execution: Each solution runs in an independent sandbox, with solve.sh as the entry point, ensuring isolation and reproducibility.

Hyra's technical principles

• Harness design: Follows the principle of The Bitter Lesson, with a lightweight framework and a broad action space. After a task description is given, it continuously runs the explore-evaluate-improve loop.
• Experience Bank: Stores multimodal experience such as source code, run logs, and evaluator feedback from past solutions.
• Inspiration contexts (Inspirations): The Context Agent integrates diverse contexts from the Experience Bank to inspire the creation of better future solutions.
• Semaphore resource control: Ensures that resources such as sandboxes, model inference, and task queues remain healthily saturated, so solution quality scales efficiently over time.
• Evaluator co-evolution: For tasks without an existing evaluator, Hyra first designs an initial evaluator, and after the inner loop optimizes it, the outer loop upgrades the evaluator again to start the next round.

How to use Hyra

• Define the task: Provide a task description and an optional initial evaluator.
• Start the Harness: Hyra automatically initializes the Context Agent and multiple Proposal Agents.
• Asynchronous operation: The Context Agent continuously generates inspiration contexts and fills the queue; Proposal Agents claim contexts and write solutions (including the solve.sh entry point).
• Sandbox execution and scoring: Solutions run in independent sandboxes, and results are fed back to the Experience Bank.
• Loop iteration: Continuously explores based on the best historical experience until actively exiting or the budget is exhausted.
• Get the best solution: Returns the best historical solution and the complete Experience Bank.

Hyra's core advantages

• Simple and effective: The framework is lightweight and does not rely on complex orchestration, while the agent's action space is broad.
• Strong generality: The same Harness spans more than ten scenarios, including training algorithms, system optimization, scientific discovery, and creative generation.
• Evolvable evaluator: A two-layer loop solves the problem of missing or easily hacked evaluation criteria in open tasks.
• Real-world scenario validation: Learns and evolves in Tencent product systems, real AI R&D pipelines, and industrial scenarios.
• Open source and transparent: All experimental artifacts and results are open sourced, facilitating reproduction and academic collaboration.

Hyra project address

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

Comparison of Hyra with similar competing products

Hyra application scenarios

• AI foundation model R&D: Automatically optimizes training recipes, data strategies, inference systems, and low-level GPU kernels, setting new records on NanoChat/NanoGPT/SOL-ExecBench.
• Mathematical scientific discovery: Solves open mathematical problems from databases such as EinsteinArena, setting new historical best results for 29 of 55 problems.
• Astronomical time series forecasting: Analyzes historical observation data such as sunspots, discovers recurrence formulas and builds long-term forecasting models, with high out-of-sample validation accuracy over nearly a century.
• Drug molecule design: For anticancer targets such as PARP1, generates candidate molecules with stable binding and better druggability than the marketed inhibitor olaparib.
• Quantum computing optimization: Designs bit routing algorithms for quantum chips such as IBM Q20, reducing two-qubit gate overhead by 44.4% compared with the classical SABRE and lowering noise accumulation.