Hyra Autonomous System Achieves New Results In Mathematics, Quantum Computing, And Drug Design Via Continuous Feedback Loop
In Brief
Tencent introduces Hyra, an autonomous research harness that iteratively generates and refines solutions across AI R&D, scientific discovery, and creative design.
Chinese technology company Tencent has introduced Hyra, a general-purpose autonomous research system designed to convert computational intelligence into practical value across diverse, high-stakes environments. The platform operates through a continuous feedback loop in which agents generate, execute, and refine solutions using accumulated experience, with applications spanning AI research, scientific discovery, and creative content generation.
Hyra’s architecture centers on a lightweight, asynchronous harness that follows the principle of keeping the framework simple while granting agents broad autonomy. At its core, a Context Agent maintains an Experience Bank—a repository of execution logs, evaluator feedback, source code, and artifacts from previous solution attempts. The agent synthesizes “inspirations” from this bank and places them in a task queue. Multiple Proposal Agents then consume these contexts, each producing a new solution with a standardized `solve.sh` entry point.
Solutions execute in isolated sandboxes, receive scores, and return their artifacts to the Experience Bank. Semaphores manage resource allocation across sandboxes, model inference, and the task queue, enabling quality to scale efficiently with available compute time. For tasks lacking predefined evaluators, Hyra upgrades to a bilevel loop: an inner cycle improves solutions against a current evaluator, while an outer cycle refines the evaluator itself using accumulated experience, thereby reducing reward-hacking risks and creating headroom for genuine progress.
Demonstrated Performance Across AI, Science, and Creative Fields
Tencent tested Hyra against three established benchmarks from Recursive’s automated AI research system: NanoChat Autoresearch, NanoGPT Speedrun, and SOL-ExecBench, covering model training, training acceleration, and GPU kernel optimization respectively. Using identical task definitions and evaluation protocols, Hyra-1.0 outperformed Recursive’s reported results across all three tasks, reducing NanoChat’s validation BPB to 0.9015, cutting NanoGPT’s time-to-target-loss to 76.4 seconds, and achieving a mean SOL of 0.771 on SOL-ExecBench. Tencent noted that stronger search capabilities also exposed evaluator vulnerabilities—such as solutions exploiting bidirectional attention leakage or empty kernel caching—which reinforced the necessity of evolving evaluation mechanisms alongside solution search.
Beyond AI infrastructure, Hyra demonstrated scientific and creative capabilities. The system set new best-known results on 29 of 55 open mathematical problems, discovered a recurrence relation for sunspot prediction with strong out-of-sample accuracy across nearly a century of data, and designed a 15-parameter Transformer for 10-digit addition—58.3% fewer parameters than the prior record. In quantum computing, Hyra produced a qubit-routing algorithm improving efficiency on IBM Q20 by 44.4% over the established SABRE method. In drug design, it generated a PARP1 inhibitor candidate outperforming the approved drug olaparib on combined binding and drug-likeness metrics, though Tencent emphasized this remains an early simulation requiring further validation.
Creative applications included an Othello bot that reached third place among 730 entries on Botzone through self-play evolution from minimax search to a hybrid AlphaZero-style approach; 3D modeling from 2D reference images judged by vision-language model rubrics; and multi-instrument musical arrangements refined through seven rounds of evaluator evolution based on harmonic, rhythmic, and aesthetic criteria.
Tencent framed Hyra as a foundation for a broader evolutionary loop in which improved scaffolds generate better data and experience, yielding stronger models that in turn enhance future scaffolds. The company indicated that future generations of its AI models and select products will co-evolve with the Hyra framework, and invited collaboration through an open form for academic partnerships, internships, and feedback.
Disclaimer
In line with the Trust Project guidelines, please note that the information provided on this page is not intended to be and should not be interpreted as legal, tax, investment, financial, or any other form of advice. It is important to only invest what you can afford to lose and to seek independent financial advice if you have any doubts. For further information, we suggest referring to the terms and conditions as well as the help and support pages provided by the issuer or advertiser. MetaversePost is committed to accurate, unbiased reporting, but market conditions are subject to change without notice.
About The Author
Alisa, a dedicated journalist at the MPost, specializes in crypto, AI, investments, and the expansive realm of Web3. With a keen eye for emerging trends and technologies, she delivers comprehensive coverage to inform and engage readers in the ever-evolving landscape of digital finance.
More articles
Alisa, a dedicated journalist at the MPost, specializes in crypto, AI, investments, and the expansive realm of Web3. With a keen eye for emerging trends and technologies, she delivers comprehensive coverage to inform and engage readers in the ever-evolving landscape of digital finance.
