Tencent has unveiled Hyra, a general-purpose autonomous research system engineered to translate computational intelligence into measurable outcomes across demanding domains. The platform operates on a continuous feedback loop in which AI agents generate, execute, and iteratively refine solutions by drawing on accumulated experience, with applications spanning artificial intelligence research, scientific discovery, and creative content generation. Hyra enters a field where several labs—including Sakana AI with its AI Scientist and various academic groups—have been exploring autonomous agents capable of conducting research tasks with minimal human oversight, but Tencent's emphasis on a bilevel evaluator-evolution mechanism and its breadth of demonstrated domains distinguish the system's scope.

Architecture and Core Design

Hyra's architecture is built around a lightweight, asynchronous harness designed to keep the framework minimal while granting agents substantial autonomy. A central Context Agent maintains an Experience Bank—a repository storing execution logs, evaluator feedback, source code, and artifacts from prior solution attempts. This agent synthesizes "inspirations" from the bank and routes them into a task queue.

Multiple Proposal Agents then consume these contextual inputs, each generating a new solution equipped with a standardized solve.sh entry point. Solutions are executed within isolated sandboxes, scored, and their resulting artifacts are returned to the Experience Bank for future reference. Semaphores govern resource allocation across sandboxes, model inference, and the task queue, enabling output quality to scale efficiently with available compute.

For tasks that lack predefined evaluators, Hyra employs a bilevel optimization loop. An inner cycle refines solutions against the current evaluator, while an outer cycle evolves the evaluator itself based on accumulated experience. This dual-layered approach is designed to reduce reward-hacking risks—a known failure mode in reinforcement learning where agents exploit evaluator flaws rather than solving the underlying problem—and create headroom for substantive progress.

Introducing Hyra-1.0, the first version of Hunyuan Research Agent. Built to recursively improve solutions for performance-driven research and engineering tasks. Explore our demos in AI4AI, AI4Science, and AI4Fun: pic.twitter.com/DgKkba1EEY — Tencent Hy (@TencentHunyuan) July 21, 2026

Benchmark Performance

Tencent evaluated Hyra against three benchmarks from Recursive's automated AI research system: NanoChat Autoresearch, NanoGPT Speedrun, and SOL-ExecBench, which cover model training, training acceleration, and GPU kernel optimization, respectively. NanoGPT, the open-source minimal GPT implementation widely used in the AI research community as a training-speed benchmark, and NanoChat represent compact-scale tasks where iterative optimization can yield measurable gains. Using identical task definitions and evaluation protocols, Hyra-1.0 surpassed Recursive's reported results on all three tasks. It reduced NanoChat's validation bits-per-byte (BPB) to 0.9015, cut NanoGPT's time-to-target-loss to 76.4 seconds, and achieved a mean score of 0.771 on SOL-ExecBench.

Tencent observed that enhanced search capabilities also surfaced evaluator vulnerabilities—such as solutions exploiting bidirectional attention leakage or empty kernel caching—underscoring the importance of evolving evaluation mechanisms in tandem with solution search.

Scientific and Creative Achievements

Beyond AI infrastructure, Hyra demonstrated capabilities across scientific and creative domains:

• Mathematics: The system set new best-known results on 29 of 55 open mathematical problems and designed a 15-parameter Transformer for 10-digit addition, reducing parameter count by 58.3% compared to the previous record.
• Astronomy: Hyra discovered a recurrence relation for sunspot prediction that showed strong out-of-sample accuracy across nearly a century of historical data.
• Quantum Computing: The system produced a qubit-routing algorithm that improved efficiency on IBM Q20 by 44.4% over the established SABRE method, a widely cited routing algorithm introduced in 2019 that has served as a standard baseline in quantum circuit compilation.
• Drug Design: Hyra generated a PARP1 inhibitor candidate that outperformed the approved drug olaparib—marketed by AstraZeneca and Merck as Lynparza, used to treat certain breast, ovarian, and pancreatic cancers—on combined binding and drug-likeness metrics. Tencent emphasized that this result stems from early-stage simulation and requires further validation.
• Game AI: An Othello bot reached third place among 730 entries on Botzone, evolving through self-play from minimax search to a hybrid AlphaZero-style approach, echoing the self-play methodologies pioneered by DeepMind's AlphaZero.
• 3D Modeling: The system generated 3D models from 2D reference images, evaluated using vision-language model rubrics.
• Music: Multi-instrument musical arrangements were refined through seven rounds of evaluator evolution based on harmonic, rhythmic, and aesthetic criteria.

Future Outlook

Tencent described Hyra as the foundation for a broader evolutionary loop in which improved scaffolds produce higher-quality data and experience, leading to stronger models that in turn enhance future scaffolds. The company indicated that upcoming generations of its Hunyuan model family—Tencent's proprietary large language and multimodal models—will co-evolve with the Hyra framework, and it invited researchers to collaborate through an open form for academic partnerships, internships, and feedback.