Hyra: An AI Research Agent That Improves Its Own Solutions

Hyra-1.0 is Tencent's Hunyuan Research Agent: a general framework for performance-driven research and engineering. Instead of answering once, it turns a task into a continuing search process in which solutions are written, executed, measured, compared, and improved.

The same harness is designed to work wherever progress can be tested: model training, systems optimization, mathematical construction, scientific data analysis, molecular design, games, music, and programmatic 3D creation.

A Context Agent maintains the Experience Bank and assembles diverse inspiration packages from earlier source code, files, run logs, artifacts, scores, and evaluator comments. Those packages enter a task queue instead of being compressed into a single generic memory.

Multiple Proposal Agents consume different contexts in parallel and write new solution folders with solve.sh as the entry point. Each candidate runs in a clean sandbox, receives a score, and returns to the Experience Bank, creating an asynchronous producer-consumer pipeline that can keep compute, sandboxes, and model calls busy.

Strong search quickly discovers weaknesses in a metric. The official article describes candidates that lowered language-model loss by leaking future tokens and GPU kernels that cached answers during correctness checks before doing no work during timing. Both scored well without solving the intended problem.

When a task lacks a reliable evaluator, Hyra uses a nested loop: the inner loop improves solutions, then the outer loop revises the evaluator for speed, precision, resistance to reward hacking, and room for further progress. The upgraded evaluator starts the next search cycle.

In the launch results, Hyra improved three established AI research tasks: NanoChat validation BPB reached 0.9015, NanoGPT reached 3.28 validation loss in 76.4 seconds, and SOL-ExecBench mean SOL reached 0.771 across the published setup.

The official results repository also publishes advances in mathematical optimization, packing, quantum routing, symbolic forecasting, drug docking, and compact neural architectures. A 15-parameter Transformer performed 10-digit addition at the target accuracy, while the reported sunspot formula achieved approximately 0.78 forecast R² on a long held-out interval.

Tencent's August 19 update added four notable results: the 3D Blaschke–Lebesgue universal lower bound rose from about 0.380799w³ to above 0.411040w³; the Beurling–Ahlfors uniform coefficient fell from 1.575 to 1.523958; the known power-law threshold for partial Hadamard matrix counting moved from exponent 3 to 2; and the commutator construction improved from O(log⁴(1/ε)) to O(log³(1/ε)).

The repository includes manuscripts and, for three of the newest results, Lean 4 formalizations of their finite or numerical cores. That makes the public release more useful than a leaderboard claim alone: readers can inspect the argument, artifacts, and verification code.

For 3D creation, Hyra receives a single 2D reference, writes Blender Python, renders the result, and uses a rubric-based vision-language judge to score silhouette, proportion, structural completeness, materials, and visual similarity. The loop keeps editing modeling code and render parameters until the asset improves.

The open artifacts include a procedural QQ penguin and a Hunyuan logo orb. This is code-driven Blender modeling rather than a general image-to-mesh service, but it demonstrates how a research agent can turn visual feedback into repeated, measurable improvements to a 3D asset.