The View from Nowhere: Population-Centered Agent Evolution

Shuo Yan; Ningning Hou; Haoyu He; Haochen Pei; Shang Gao

cs.AIOct 27, 2026

The View from Nowhere: Population-Centered Agent Evolution

Authors: Shuo Yan, Ningning Hou, Haoyu He, Haochen Pei, Shang Gao

Organizations: Macquarie University · Monash University

Abstract

Self-improving AI systems can revise the designs that govern their own behavior. Yet existing approaches center improvement on selected parent agents and their evolutionary experience, leaving the broader population largely underused. We identify a substantial gap between population coverage (i.e., the tasks solved by at least one evolved variant) and the performance of the best individual agent. This gap shows that useful capabilities and diagnostic evidence emerge across the population without being consolidated into its leading agents. These observations motivate the view from nowhere, a principle for agent evolution: an agent should be improved by interpreting its behavior through population-wide patterns of success, failure, and prior self-modification. We instantiate this principle as Population-Centered Evolution (PCE). PCE distils task records and modification histories from the population into a population block, which captures recurring failures, complementary successes, and the effects of previously attempted modifications to guide improvements to a selected parent's harness. Under matched models, initial harnesses, and evolution schedules, PCE achieves 72.1% success on SWE-bench Verified and 96.7% on Polyglot, improving over a strong group-evolution baseline by 3.2 and 5.9 percentage points. In our evaluation, PCE reduces the gap between population coverage and best-agent performance by one-third on SWE-bench and closes it on Polyglot. These results suggest that treating the population as evidence, not merely as candidates for selection, provides a complementary principle for agent evolution.

Figures & tables

Figure 1: Population potential and task-level evidence. (a) The union of tasks solved by the GEA population exceeds the performance of its best individual agent by 15.0 percentage points (9 tasks in Held-in 60). (b) Parent-local diagnosis uses one parent trajectory to identify missed API cases and add verification instructions, while the other 15 population records are not inspected. (c) Population-centered diagnosis uses patterns across 17 records to identify a missing patch check and add bounded repair logic.

Figure 2: Parent-local and population-centered evolution. (a) Parent-local evolution diagnoses and modifies an agent using execution records associated with the selected parent. (b) Population-centered evolution distils task records and modification histories across the population into a population block. The block summarizes outcome evidence, identifies cross-agent observations, and suggests candidate files for inspection. Diagnosis combines this block with the selected parent’s records and source code to propose a harness modification.

Benchmark	Method	Model	Gen.	Best $\uparrow$	Overall $\uparrow$
SWE-bench	OpenHands †	GPT-5	N/A	N/A	71.8
	DGM	Claude 3.5 Sonnet	80	—	50.0
	DGM ‡	Claude 3.7 Sonnet	80	—	59.5
	DGM ‡	o3-mini	80	—	33.0
	DGM (GEA) †	Haiku/Sonnet 4.5	60	—	56.7
	GEA †	Haiku/Sonnet 4.5	30	—	71.0

Table 1: Performance (%). Best: highest Held-in 60 search score.

\dagger

: reported by GEA ( Weng et al., 2026 ) .

\ddagger

: model transfer. Comparison settings appear in Appendix A .

Figure 3: SWE-bench search

Benchmark	Method	Top-1 $\uparrow$	Top-3 min. $\uparrow$	Top-5 min. $\uparrow$	Union	Gap $\downarrow$
SWE-bench	GEA	70.0 $\pm$ 2.4	66.7 $\pm$ 2.4	66.7 $\pm$ 2.4	84.2 $\pm$ 3.5	14.2 $\pm$ 1.2
SWE-bench	PCE	75.0 $\pm$ 0.0	72.5 $\pm$ 1.2	70.8 $\pm$ 1.2	85.0 $\pm$ 0.0	10.0 $\pm$ 0.0
Polyglot	GEA	90.0	85.0	83.3	100.0	10.0
Polyglot	PCE	100.0	98.3	96.7	100.0	0.0

Table 2: Leading-variant performance and population coverage on Held-in 60 (%). Top-1 is the highest search score, and Top-

k

min. is the lowest score among the top-

k

evolved variants. Union is the fraction of tasks solved by at least one evaluated child, and Gap is Union minus Top-1 in percentage points. SWE-bench reports mean

\pm

sample standard deviation over two runs; Polyglot reports one run.

Method	Best	Mean $\pm$ sd	Intersection	Union
GEA	90.0	92.0 $\pm$ 2.5	85.0	100.0
Parent distillation	91.7	92.0 $\pm$ 4.3	78.3	98.3
Population distillation (PCE)	100.0	96.0 $\pm$ 2.8	88.3	100.0

Table 3: Parent-only distillation ablation on Polyglot Held-in 60 (%). Best is the highest search score during evolution. The selected agent is then evaluated five times. Mean

\pm

sd summarizes these evaluations, Intersection is the fraction of tasks solved in every evaluation, and Union is the fraction solved at least once. Parent-only distillation matches PCE’s distillation model, tools, budget, and output format but uses only evidence from the selected parents.

Metric	SWE-bench		Polyglot
Metric	GEA	PCE	GEA	PCE
Overall success (%) $\uparrow$	68.9	72.1	90.8	96.7
Total tokens (billions) $\downarrow$	2.98	1.92	0.99	0.41
Total requests $\downarrow$	114,629	69,539	52,034	28,823
Distillation calls	0	758	0	775
Evaluation episodes	1,100	1,200	1,147	1,177

Table 4: Performance and resource use in the main evolution runs. Token and request counts are provider totals for the complete run, covering task execution, diagnosis, self-modification, and population distillation. Distillation calls report the number made by the distillation agent. An evaluation episode is one task attempt, and trajectory length is measured in characters. We report tokens and requests rather than monetary cost because provider prices vary over time.

Appendix figures & tables3 assets

Supplementary material from the paper’s appendix.

Appendix

	SWE-bench				Polyglot
Gen	PCE 1	PCE 2	GEA 1	GEA 2	PCE	GEA	Parent dist.
0	40, 34	40, 29	41, 31	25, 15	38, 47	42, 47	45, 48
1	38, 40	40, 36	0/10, 33	27, 19	38, 45	50, 41	45, 46
2	40, 40	42, 45	42, 39	28, 25	42, 58	51, 50	47, 47
3	42, 37	42, 40	40, 37	39, 13	58, 59	47, 54	45, 51
4	44, 38	41, 40	39, 19	37, 27	60, 57	49, 44	0/10, 44

Appendix

Table 5: Search scores of the children in all reported runs. Each cell lists the two children of a generation, scored out of 60 tasks. An entry of 0/10 denotes an evaluation stopped at the ten-task gate, and a dash denotes self-improvement that produced no patch. Min, Max, and Median are over fully evaluated children.

Prompt condition	Sphinx-7748	Sphinx-8638	Sphinx-9229
Prompt condition	I / E / T	I / E / T	I / E / T
GEA	3 / 2 / 0	4 / 0 / 1	5 / 0 / 0
+Success	—	4 / 0 / 1	5 / 0 / 0
+Own failures	—	4 / 0 / 1	5 / 0 / 0
+Population failures	—	3 / 0 / 2	5 / 0 / 0
+Block	0 / 5 / 0	5 / 0 / 0	3 / 2 / 0

Appendix

Table 6: Proposal counts from five diagnoses per condition. I: instructions. E: episode logic. T: new tool. Dashes indicate untested trajectory controls due to insufficient eligible trajectories at generation 1.

Benchmark	Set	Agent	Rep 1–5	Mean $\pm$ sd (%)
SWE-bench	Held-in 60	Initial	33, 33, 36, 33, 35	56.7 $\pm$ 2.4
		GEA	39, 39, 39, 42, 40	66.3 $\pm$ 2.2
		PCE	39, 43, 45, 44, 40	70.3 $\pm$ 4.3
	Held-out 60	Initial	38, 30, 31, 29, 32	53.3 $\pm$ 5.9
		GEA	41, 39, 39, 40, 37	65.3 $\pm$ 2.5
		PCE	40, 42, 43, 42, 44	70.3 $\pm$ 2.5

Appendix

Table 7: Repeated and held-out evaluations of the selected agents. Rep 1–5 are solved-task counts, and the last column is mean success percentage

\pm

sample standard deviation. Held-out 140 and Held-out 165 are evaluated once, and their rows list the solved counts and success percentages of the initial agent, GEA, and PCE. All tasks of a set remain in the denominator, including never-graded attempts.

LLM Agent Self-Improvement

Large language model agent self-improvement studies how agents use experience and feedback to improve their behavior, tools, and workflows.

See all 277 papers

SAGE: A Quantitative Evaluation of Socialized Evolution in Agent Ecosystems
Jun 2, 2026Linyue Pan, Yaoming Zhu, Lin Qiu +2Multi-Agent Collaboration AI Agent Benchmarks
Tsinghua University, China · Meituan, China
PACE: Anytime-Valid Acceptance Tests for Self-Evolving Agents
Jun 6, 2026Zayx ShawnSequential Hypothesis Testing LLM Agent Reliability
Independent Researcher
AgentEvolver: System-Wide Self-Evolution Through Task Execution
Oct 8, 2026Wentao Zhang, Fuchao Yang, Yilei Zhao +2LLM Agent Self-Improvement Self-Evolving Agents

Jun 2, 2026J/K move · Enter open · S save

SAGE: A Quantitative Evaluation of Socialized Evolution in Agent Ecosystems

Linyue Pan, Yaoming Zhu, Lin Qiu, Xuezhi Cao+1

Tsinghua University, China · Meituan, China

Multi-Agent Collaboration AI Agent Benchmarks Multi-Agent Systems Self-Improving Agents

Open paper AskPDF arXiv

Self-improving language agents are typically evaluated in isolation: an agent attempts a task, receives feedback, and iteratively refines its own behavior. Yet agents increasingly operate alongside peers whose strategies and outcomes are publicly visible. This raises an under-studied question: when does shared experience produce improvements that self-improvement alone cannot achieve? We introduce SAGE (Social Agent Group Evolution),an evaluation framework that compares two compute-matched conditions: SocialEvo, where agents from five distinct model families co-evolve with access to all peers' histories; and SelfEvo, where each agent receives the same number of task attempts but sees only its own past, which is conventional in self-improving agent studies. We instantiate SAGE in three arenas: open-ended ML research, long-horizon economic planning, and strategic multiplayer play, evaluated across multiple evolutionary rounds. We find that group history is not a universal amplifier: the strongest agent does not exceed its self-evolution ceiling. However, agents that plateau under self-improvement can achieve significant breakthroughs when peer experience is available. In competitive settings, counterfactual controls reveal that agents improve generally rather than developing opponent-specific strategies. Across different forms of shared history, filtered peer traces and reflective summaries often outperform raw logs, indicating that social gains depend on abstraction rather than exposure volume. These findings reveal that peer-history gains are agent-specific, arena-dependent, and contingent on the capacity to abstract transferable knowledge from public traces.

Abstract

LLM Agent Self-Improvement

Explore similar work

SAGE: A Quantitative Evaluation of Socialized Evolution in Agent Ecosystems

PACE: Anytime-Valid Acceptance Tests for Self-Evolving Agents

AgentEvolver: System-Wide Self-Evolution Through Task Execution