Title: SAGE: A Quantitative Assessment of Socialized Evolution Within Agent Ecosystems

Abstract:

Traditional evaluations of self-improving language agents typically isolate the subject: an agent tackles a task, obtains feedback, and iteratively adjusts its approach. However, as agents increasingly operate in environments where peers’ strategies and results are publicly observable, a critical, under-researched question emerges: under what circumstances does shared experience yield performance gains that self-improvement alone cannot produce?

To address this, we present SAGE (Social Agent Group Evolution), an evaluation framework designed to compare two conditions matched for computational resources. In the "SocialEvo" condition, agents from five distinct model families co-evolve with access to the complete history of all peers. In the "SelfEvo" condition, each agent undergoes the same number of task attempts but is restricted to viewing only its own past—a standard practice in current self-improvement research.

We applied SAGE across three domains: open-ended machine learning research, long-horizon economic planning, and strategic multiplayer gaming, assessing performance over multiple evolutionary rounds. Our results indicate that group history is not a universal performance booster; the top-performing agents do not surpass the limits set by their own self-evolution. Nevertheless, agents that hit a plateau through self-improvement alone can achieve substantial breakthroughs when peer experiences are accessible.

In competitive scenarios, counterfactual controls demonstrate that improvements are general rather than specific to particular opponents. Furthermore, the study of different shared history formats reveals that filtered peer traces and reflective summaries frequently outperform raw logs. This suggests that social gains rely more on the ability to abstract and transfer knowledge from public records than on the sheer volume of exposure. Ultimately, these findings highlight that the benefits of peer history are contingent upon the agent, the specific arena, and the capacity to distill transferable insights from public traces.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC