Three months ago, Tencent Hunyuan introduced the minimalist Agent loop system Hyra1.0 for general scientific research problems. Driven by its open-source model Hy3, Hyra successfully solved a core problem that had remained unresolved for over fifty years in addition algebra—the optimal exponent for sums and differences. This marked the first time an domestic open-source model fully automated the solution of a mathematical open problem. Subsequently, international AI giants have also begun competing in the fields of mathematics and theoretical computer science.
Since the path of “AI problem-solving” has been taken, where should automation in scientific research move next? How can we make AI more economical and sustainable in solving complex scientific challenges? In response to these industry considerations, Tencent Hundun suggests upgrading tools. In the past, AI4Science evolved from models such as AlphaFold, AlphaEarth, and AlphaGeometry that addressed specific problems, to a paradigm where general-purpose agents directly solve scientific issues. The recently released Hyra2.0 aims to take another step forward, enabling general-purpose agents to deeply participate in the development and upgrade of scientific tools and models. This allows the results of one development effort to be reused in subsequent research, creating intelligent compound benefits.
The new Harness architecture designed for tool upgrades
The development tools place much higher requirements on the underlying Harness than just solving problems. Training a model often requires a process of scheme design, code implementation, training, evaluation, and repeated modifications, with the entire cycle lasting several days or even weeks. To enable the Agent to adapt to a real industrial-level development environment, Hyra2.0 integrates Tencent’s internal actual development pipeline: aligning requirements with researchers via Enterprise WeChat, submitting and managing experiments on a computing platform, implementing and verifying changes in a code repository, and properly preserving the experimental process and results.
In terms of division of labor and collaboration, Hyra2.0 adopts a three-layer innovative architecture consisting of Overall, Mission, and Worker.
Overall (Overall coordination): Responsible for communicating with human researchers, managing project priorities, and coordinating computing resources and key decisions.
Mission (Mission Commander): Long-term responsibility for an independent research issue, deciding on research paths, designing experiments, organizing Workers, and being responsible for the final conclusions.
Worker (executing worker bee): Reads code independently, conducts experiments, analyzes reasons for failure, and proposes new ideas. The system supports different Workers to choose different models, providing multi-angle analysis and review suggestions.
To support long-term, parallel complex research, the system introduces persistent queues and a Research Board, and designs an independent review mechanism (Review Worker) to strictly control the quality of conclusions. Meanwhile, in order to avoid redundant experimentation, Hyra2.0 stores project status, Agent working memory, and reusable experiences separately, and deposits validated effective methods into the shared knowledge base Books. These are ultimately organized into transferable skills that can be repeatedly called upon and continuously iterated for subsequent tasks.
Initial achievements and hardcore breakthroughs in seven cutting-edge fields
Currently, Hyra2.0 has completed the development and upgrade of a large number of cutting-edge tools and models, and these achievements are continuously published in the open-source community. It covers seven core areas.
HyInv (Protein Reverse Folding): To address the reverse folding challenge in drug design, Hyra2.0 trained a graph neural network HyInv with only 1.96 million parameters from scratch. It uses a two-step sampling method of “draft first, then rewrite,” and compared to the industry-standard ProteinMPNN, the generated sequences have a better match with the target structure, with a generation speed that is 57 times faster.
HyPLM (Protein Language Model): The efficient protein language model HyPLM-Flash, trained from scratch, has 650 million parameters and uses a total of 726 billion training tokens (only one-fifth to one-eighth of similar models). It demonstrates the fastest inference speed among its peers and ranks 3rd on average across 35 evaluation tasks.
HyFF (atomic-level base model): An atomic-level base model for large-scale molecular simulations. HyFF offers five scales (2.3 million to 1.2 billion parameters) and supports 83 elements. Its actual simulations have successfully supported 100 nanoseconds simulations of a water channel protein membrane system with nearly 60,000 atoms, as well as tracking of complete bond breaking and bond forming reactions in a 46,000原子 enzyme system.
HyProver (Mathematical Proof Workflow): A agent workflow designed for pure mathematical proofs, which has strong exploration capabilities while maintaining high token efficiency. It has already collaborated with mathematicians to verify and organize the proofs of Mr. Shing-Tung Yau’s circumference action conjecture and the five-dimensional Petty isometry point conjecture.
HyQEC (Quantum Error Correction System): A system for the construction and verification of quantum low-density parity check codes (qLDPC). The designed Split Pauli-CPM algorithm successfully constructed and rigorously verified the quantum codes [488, 123, 22]. The encoding efficiency metric improved by approximately 48% compared to the original leader; in terms of code distance verification, the GPU-accelerated search throughput is 8 to 55 times that of the original CPU tool, and accurate code distance certification was achieved for 178 out of 187 entries on the ranking list.
Elegant hand body design: Try to extend exploration into the robot body structure. The agent independently discovered a “spiral hand” design with five fingers arranged around the palm and varying lengths. With the same fine-tuning budget for strategy, the average rotation speed reached 3.277 rad/s, which is 28.6% faster than the optimal design, and the number of target completions increased by 21%.
Xiangshan Processor Optimization: Performance tuning and correctness verification were carried out around the open-source RISC-V processor “Xiangshan”. By fixing several issues, the SPEC CPU2006 benchmark score per GHz increased by 5.79% (equivalent to the results achieved through 2 to 3 months of official iteration); meanwhile, through new designed fuzz tests and formal verification, 10 undisclosed bugs and one example of multi-core cache consistency issues were identified.
Continuous evolution: Turning today’s research into tomorrow’s tools
Currently, Hyra2.0 is also deeply involved in the pre-training, post-training, and inference optimization of Tencent Hunyuan’s next-generation Hy model, reducing costs by shortening the experimental cycle and automatically generating high-quality training data.
With the official release of Hyra2.0, automated scientific research is undergoing a profound paradigm shift: The contribution of AI is no longer limited to providing a specific answer, but rather effectively enhancing the research capabilities for a category of problems. By turning valuable experience from individual attempts into reusable tools, workflows, and skills, the value of each development effort continues to accumulate through repeated use. Ultimately, through the positive cycle of “the better the tools, the stronger the system that creates them”, the boundaries of human science are continuously expanded.