Hyperagents — part 7/7

Cleaned-text code points [93356, 98124). Bibliography omitted here; source.md includes it.

Over time, the temperature parameter itself becomes adaptive, responding to archive size, recent progress, or stagnation. Later iterations further enhance parent selection with multi-component scoring systems that integrate additional signals such as diversity, recency, and elite performance. A representative evolved formulation is:

Several runs also introduce explicit stagnation detection, increasing exploration when progress slows:

Despite the increasing sophistication, the automatically modified parent selection does not outperform a carefully handcrafted score-child-prop mechanism ( Section A.2 ). Qualitatively, this appears to result from the added complexity and sensitivity of the learned mechanisms. While the DGM-H reliably creates effective components (e.g., UCB, softmax sampling, and adaptive parameters), it does not consistently identify a strong parent selection strategy across different runs. Nonetheless, these results demonstrate that the DGM-H can autonomously recreate classic selection algorithms, extend them with adaptive heuristics, and explicitly reason about failure modes such as stagnation, highlighting its ability to improve core elements of the open-ended optimization process itself.

This is the pseudocode of DGM-H with modifiable parent selection:

## Appendix F Additional Safety Discussion

Reflection and amplification of human biases. In this work, objectives are specified through fixed benchmarks and evaluation criteria. The DGM-H does not alter the underlying task definitions; instead, it optimizes performance with respect to the provided objectives. For example, in paper review, the DGM-H learns to predict acceptance decisions that reflect existing human review data, rather than modifying the review process itself. As a result, the DGM-H reflects the norms and biases present in the data and benchmarks on which it is trained. In this sense, the system acts both as a clarifier and an amplifier of existing human behavior. By making implicit preferences and biases explicit, measurable, and reproducible, the DGM-H can surface latent assumptions in human decision-making processes. This creates the possibility of co-evolution between humans and AI systems, where human institutions adapt their norms and objectives in response to insights revealed by automated optimization. However, if the benchmarks encode undesirable biases or misaligned incentives, the DGM-H will faithfully optimize for them and may exacerbate their effects. This underscores the importance of careful benchmark design, dataset curation, and periodic re-evaluation of evaluation criteria. Within this framing, safety concerns include critically examining and improving the human-defined objectives against which agents are optimized.

Evaluation gaming. Another safety concern arises from the risk of evaluation gaming, a manifestation of Goodhart’s law [ Strathern, 1997 ] , where optimizing for a metric leads to improvements on the metric without progress on the intended underlying objective. Because the DGM-H optimizes empirical evaluation signals, self-improving agents may discover strategies that exploit weaknesses or blind spots in the evaluation procedure. Such strategies can yield higher measured performance while deviating from the true goal the benchmark was designed to capture. Mitigating evaluation gaming requires robust, diverse, and periodically refreshed evaluation protocols, as well as complementary metrics, held-out tests, and human oversight. More broadly, these considerations highlight that as self-improving systems become more powerful, safety increasingly depends on the fidelity and robustness of the evaluation signals that guide optimization, rather than solely on the transparency or constraints of the learning algorithm itself.

## Instructions for reporting errors

We are continuing to improve HTML versions of papers, and your feedback helps enhance accessibility and mobile support. To report errors in the HTML that will help us improve conversion and rendering, choose any of the methods listed below:

Click the "Report Issue" ( ) button, located in the page header.

Tip: You can select the relevant text first, to include it in your report.

Our team has already identified the following issues . We appreciate your time reviewing and reporting rendering errors we may not have found yet. Your efforts will help us improve the HTML versions for all readers, because disability should not be a barrier to accessing research. Thank you for your continued support in championing open access for all.

Have a free development cycle? Help support accessibility at arXiv! Our collaborators at LaTeXML maintain a list of packages that need conversion , and welcome developer contributions .