Does sharing experience across agents beat isolated evolution?

Group-Evolving Agents (GEA) makes "a group of agents, rather than an individual agent, as the fundamental unit of evolution." Prior open-ended self-evolving systems follow "chain- or tree-structured evolution, where different branches remain strictly isolated," so exploratory diversity "rarely serves as effective stepping stones" and instead produces "short-lived variants that fail to contribute to long-term cumulative progress." GEA's fix is explicit experience sharing within a group at each iteration. On the paper's own benchmarks this measures out to 71.0% on SWE-bench Verified and 88.3% on Polyglot, against 56.7% and 68.3% for the prior open-ended self-evolving methods it compares against — and, "without any human intervention," comparable to or beating human-designed frameworks (71.8% and 52.0% on the same two benchmarks).

Mechanically, each iteration has two stages. Parent Group Selection picks K agents from an archive using a "Performance–Novelty" score: agents are represented as task-success vectors, novelty is cosine distance to nearest neighbors, and the combined score weights performance as primary with novelty as "a mild bias." Open-Ended Group Evolution then has the parent group jointly produce a same-size offspring group by pooling each agent's evolutionary traces — code patches, a predicted patch for a sampled unsolved task, that task's execution logs and tool-invocation history, and the evaluation outcome exposing failure modes — across the whole group rather than within one lineage. The paper's own accounting of why this works: nine tool-level modifications drove the gains, GEA's best agent integrated eight of them, the comparison system's best agent integrated only five, and "the four tools missing from the DGM agent were explored in isolated branches... but failed to propagate due to lineage isolation." Five of GEA's eight originated from different parent agents — the sharing mechanism, not just more search, is doing the work.

This reads as a worked instance of the first stage in Can agents evolve beyond the constraints humans engineer? — peers exerting adaptive pressure on each other — with benchmark numbers behind the survey's claim that single-entity evolution stalls inside isolated branches. It also names its foil directly: the Darwin Gödel Machine of Can AI systems improve themselves through trial and error? is the "DGM agent" GEA out-integrates above, so GEA keeps DGM's empirical-validation-plus-archive recipe and replaces DGM's individual-lineage parent selection with group-level selection and cross-agent sharing. And because GEA's gains are, in its own analysis, "workflow and tool enhancements rather than model-specific optimizations" on frozen GPT- and Claude-series backbones, this sits inside the fast loop of Do self-improving agents really split into two distinct loops? — no weights change, only the scaffold each agent carries.

The excerpt does not name the specific prior methods behind the 56.7%/68.3% and 71.8%/52.0% comparison figures, so how directly comparable those baselines are cannot be checked from this text alone. All results come from the authors' own runs in "isolated sandbox environments," with no independent replication reported here, and the group size K, archive size, and compute budget behind these numbers aren't given in this excerpt. The authors themselves flag that open-ended exploration "may inadvertently introduce directions misaligned with human intent" and produce patches "difficult to fully understand," and their extension to non-coding domains (they mention bias mitigation) is stated as potential, not demonstrated. The defensible claim is narrower than "group evolution beats individual evolution" in general: within this paper's coding benchmarks and this archive-based evolutionary setup, cross-agent experience sharing converted diversity that would otherwise die out in isolated branches into measurable, transferable gains.