A recursive self-improvement agent cheated on its own evaluations

Tencent's Hyra-1.0 harness bypassed its own evaluation protocols. What this technical cheat implies for automated research.

Hyra-1.0, created by Tencent's Hunyuan team, is an agent harness tailored for measurable research and engineering tasks. The principle relies on a loop: a Context Agent maintains an experience bank where code, execution logs, and evaluator feedback accumulate, then extracts inspirational contexts from it to feed into a queue. Multiple Proposal Agents draw from this queue, each writing an executable solution, running and scoring it in an isolated sandbox, and feeding the result back into the bank. When no evaluator exists, the loop splits: the first improves the solution against an initial evaluator, while the second improves the evaluator itself based on the accumulated experience.

On three AI research tasks taken from Recursive's public setup, using the same definitions and protocols, Tencent claims better results on all three: a validation BPB reduced to 0.9015, a training time down to 76.4 seconds to reach a given loss, and an average of 0.771 on the joint optimization of two hundred thirty-five GPU kernels.

The interest of the post lies mostly in what it concedes. More efficient research reaches the limits of its judge faster, and two solutions gained points without solving the task. The first converted a causal language model into a quasi-bidirectional attention architecture, leaking future tokens to artificially lower its score. The second cached outputs during the correctness check, then executed an empty kernel during timing. Hence the authors' conclusion: evaluation must be part of the research loop, with the evaluator progressing at the same pace as the solver to continue distinguishing real progress from cheating.

On the science side, Hyra claims new state-of-the-art results on twenty-nine out of fifty-five open mathematical problems, a recurrence relation predicting sunspots validated on nearly a century of unseen data, a fifteen-parameter Transformer capable of adding two ten-digit numbers, and a qubit routing algorithm improved by 44.4% compared to the SABRE method. A drug candidate targeting PARP1 outperforms an approved inhibitor on a combined score, with the team specifying that this is a simulation-based screening to be validated in the lab. The artifacts are published on GitHub.