SkillNEO: Non-Elitist Evolutionary Optimization of Agent Skills
Authors: Enze Ma, Chunyu Miao, Yankai Chen, Jie Yang, Liancheng Fang, Kening Zheng, Hanrong Zhang, Muyao Xu, Xinle Yu, Chengze Li, Wei-Chieh Huang, Ziwei Ren, Yunhao Du, Zhixiang Shen, Xue Liu, Philip S. Yu, Zhen Wang
Organizations: University of Illinois at Chicago · Mohamed bin Zayed University of Artificial Intelligence · Independent Researcher · University of California, San Diego · University of California, Los Angeles · University of Texas at Arlington
Agent skills turn procedural knowledge into reusable artifacts, allowing LLM agents to accumulate and refine expertise through experience. Building on this, skill evolution iteratively improves skills using execution feedback. However, improvement may require several interdependent edits, so low-reward versions can still be worth developing. Requiring immediate reward gains can discard useful partial repairs before their benefits emerge. Non-elitist evolution allows new versions to receive further edits despite scoring below earlier ones. For agent skills, this calls for evidence of what to preserve and repair next. We introduce SkillNEO (Non-Elitist Evolutionary Optimization for Agent Skills), which separately selects a version to use and versions to keep improving. Comparing how related versions perform the same task reveals working steps to preserve and remaining failures to fix. It can also revisit and revise earlier versions using later evidence. Across 86 SkillsBench tasks in eight domains, including software engineering and office work, SkillNEO raises GPT-5.4 Mini's average reward from 49.1% to 66.4%, within GPT-5.5 and GPT-5.6 Luna's original-skill range. Mini executes tasks, proposes revisions, and edits skills; final evaluation uses the same task instances. Gains span all eight domains. The same skills improve stronger models without further editing. On the 41 tasks with modified skills, reuse improves Luna by 10.5 percentage points and places GPT-5.5 10.6 points above its published original-skill reference. Execution records show these skills can benefit stronger models even when Mini still fails some task requirements. Search histories also show lower-scoring versions leading to skills ultimately selected for use. Together, these findings show that skill evolution can turn a weaker model's partial progress into reusable procedural knowledge for improving more capable agents.
Figures & tables
Figure 1: Non-elitist evolution allows weaker revisions to lead to better skills. (a) Single-direction refinement discards alternative revisions, potentially losing useful partial repairs. (b) SkillNEO develops and revisits such revisions, using execution-trace comparisons to guide further repairs that can yield a new best skill. States si pair skill bundles with their execution records. The dark-blue node marks the new deployment incumbent; crossed-out nodes in (a) denote discarded revisions. Solid tree edges denote edits; in (b), the dashed orange arrow denotes reselection.
Figure 2: SkillNEO separates choosing a skill to use from choosing versions to improve. A version can receive further edits (green) even when it scores below the version selected for use (blue). To guide these edits, the Proposer compares saved task executions to identify steps that work and failures that need repair. The Skill Editor applies the proposed changes to instructions, scripts, and resources. Edit Validation checks the revised files before the Executor runs the task with the new version. The resulting scores inform selection, and execution records guide the next round. Mini fills every model-based role in our experiments.
Model
Original / reference
Final
Δ (pp)
GPT-5.4 Mini
49.1
66.4
+17.3
GPT-5.5
66.1
69.6
+3.5
GPT-5.6 Luna
67.3
71.9
+4.6
GPT-5.4 †
51.7
65.5
+13.8 †
Table 1: Overall rewards and cross-model reuse. Average task reward (%) for the 86-task assignment (a) and 41 modified tasks (b); Δ is in percentage points. † GPT-5.4 compares Vals (87 tasks, OpenHands) with final evaluation on 86 tasks using Codex. GPT-5.5 uses published, task-aligned SkillsBench references; Luna uses post-freeze evaluations of both conditions. Subset gains / ties / losses are 29/12/0 (Mini), 15/18/8 (GPT-5.5), and 11/23/7 (Luna), with ties within ±0.05 pp.
Domain
N
No skill †
Original skills
SkillNEO
Δ (pp)
Software engineering
16
26.9
48.3
75.7
+27.4
Office / white collar
14
40.5
47.6
61.9
+14.3
Natural science
14
38.3
53.0
71.4
+18.4
Industrial / physical
14
28.6
48.3
53.9
+5.6
Finance / economics
9
18.5
22.2
48.1
+25.9
Mathematics / formal
7
38.1
61.9
79.5
+17.6
Table 2: Mini’s rewards across eight domains. Average task reward (%) on the same 86 tasks. No skill † is a published GPT-5.4 Mini/OpenHands reference , reaggregated by domain; Original skills and SkillNEO use our Mini/Codex high evaluation. Δ=\mboxSkillNEO−Original skills , in percentage points. The Overall row averages all 86 tasks.
Figure 3: Mini improves across all eight domains. No skill † uses the OpenHands reference in Table 2 ; Original skills and SkillNEO use our Mini/Codex high evaluation. Parenthesized gains are SkillNEO minus Original skills (pp). Domain counts sum to 86.
Figure 4: Case studies of skill evolution in SkillNEO. (a) A zero-reward partial repair leads to the next selected skill in Gravitational Wave Detection. (b) Sales Pivot Analysis reuses a sibling’s PDF extraction procedure in another branch. (c) Revisiting an earlier parent produces the final skill in Paratransit. (d) Restarting from the original bundle uses execution evidence collected later in Wyckoff. Panel (a) reports three-run mean rewards at creation. In (c) and (d), solid edges show primary parents, broken arrows show reuse or reselection, and rewards come from independent final Mini evaluations (Appendices H and I ).
Model
Reward
Intended scan scope
Correct output fields
Mini
66.7
3/3
2/3
GPT-5.4
100.0
3/3
3/3
GPT-5.5
100.0
3/3
3/3
Luna
100.0
3/3
3/3
Table 3: Execution behavior with the same final skill bundle. Models differ in which requirements they satisfy despite receiving identical guidance. Counts give passes out of three executions; skipped checks are not passes. For Sales Pivot, Valid pivot objects requires the row-field, aggregation, and cache-field checks to pass. Reward is mean task reward (%).
Appendix figures & tables14 assets
Supplementary material from the paper’s appendix.
Appendix
Model / agent
No skill
Original skill
Δ
Mini / OpenHands (official)
29.9
40.7
+10.8
Mini / Codex high (this study)
39.1
49.1
+10.0
GPT-5.5 / OpenHands (official)
50.9
66.9
+16.0
GPT-5.5 / Codex (official)
46.2
66.1
+19.9
Appendix
Table 4: No-skill and original-skill references. Original skills increase average reward in each listed configuration on the aligned 86 tasks. Rewards are percentages and Δ is in percentage points. Official rows are reaggregated from the published per-task registry. Differences are computed from the displayed means. No Luna no-skill result has been measured.
Study
Executor
Tasks
Human
Developed
SkillAudit ( Gao et al., 2026 )
Opus 4.8
89
56.7
73.9
MUSE-Autoskill ( Lin et al., 2026a )
GPT-5.5
75
59.67
53.42
SPARK ( Zhou et al., 2026b )
GPT-5.4 Mini
86
≈47
≈52
SkillNEO (this study)
GPT-5.4 Mini
86
49.1
66.4
Appendix
Table 5: Reported skill-development results on SkillsBench. Average rewards (%) use each study’s task cohort and evaluation protocol. Human denotes each study’s human-skill reference; Developed uses its resulting skills. SkillEvolver and CoEvoSkills instead report complete-pass frequencies. SkillAudit, MUSE-Autoskill, and SkillNEO use the listed model for development and execution; SPARK uses stronger-model teacher trajectories; SPARK’s scores follow Section 4.1 of the source. SkillAudit and SkillNEO refine supplied skills; MUSE-Autoskill and SPARK generate skills, with Human as a separate reference. MUSE counts its 28 uncovered tasks as zero. SPARK describes an 86-task suite; exact Mini task alignment is unavailable. For SkillNEO , 41 modified bundles receive new final executions; 45 retained bundles share their original Mini measurements.
GPT-5.4
GPT-5.5
GPT-5.6 Luna
Mini outcome
n
Evolved
Original
Evolved
Gap
Original
Evolved
Gap
All modified
41
66.8
60.5
71.1
+10.6
63.9
74.4
+10.5
Mini improved
29
69.1
64.3
78.1
+13.8
62.8
81.0
+18.3
Mini tied
12
61.1
51.4
54.2
+2.8
66.7
58.3
-8.3
Appendix
Table 6: Receiver rewards grouped by Mini’s outcome. GPT-5.5 and Luna have larger mean gains on tasks where Mini improves than on tasks where Mini ties. Scores report average reward (%); Gap is evolved minus original reward in percentage points. The 41 modified tasks are grouped post hoc by Mini’s final evaluation outcomes, using its ±0.05 pp display band. Evolved denotes the same frozen Mini-developed bundle for each task. GPT-5.4 reports frozen-skill endpoints; original-skill subgroup scores are not reported. Evaluation settings follow Section 4.1 .
GPT-5.5 direction
Luna gain
Luna tie
Luna loss
Total
Gain
6
6
3
15
Tie
4
13
1
18
Loss
1
4
3
8
Total
11
23
7
41
Appendix
Table 7: Joint directions of receiver reward changes. Gains and losses need not coincide across GPT-5.5 and Luna on the 41 modified tasks. Rows use GPT-5.5 reference differences; columns use Luna differences from post-freeze evaluations. Both axes use the existing ±0.05 pp display band. Original-skill reference sources appear in Appendix B .
Stratum
n
Original
Evolved
Δ
Full-set contribution
Modified skills
41
63.9
74.4
+10.5
+5.0
Mini positive subset
29
62.8
81.0
+18.3
+6.2
Mini tied subset
12
66.7
58.3
-8.3
-1.2
Original retained
45
70.4
69.6
-0.7
-0.4
Appendix
Table 8: Contributions to Luna’s overall reward difference. Gains on modified tasks outweigh the decrease on retained tasks. Rewards are percentages; Δ and Full-set contribution are in percentage points. The Mini-positive and Mini-tied subsets partition the 41 modified tasks. The 41 modified and 45 retained tasks partition the benchmark. Their contributions are nΔ/86 and sum using unrounded values. Mini-positive/tied strata are defined post hoc from Mini’s final evaluation outcomes.
Original-performance stratum
n
Original
Evolved
Δ
G/T/L
Zero observed reward
8
0.0
50.0
+50.0
4/4/0
Partial observed reward
16
57.5
61.4
+3.9
7/3/6
Perfect observed reward
17
100.0
98.0
-2.0
0/16/1
Appendix
Table 9: Luna transfer by observed original performance. Among the 41 modified tasks, the group with zero original reward has the largest mean gain. Rewards are percentages, Δ is in percentage points, and G/T/L counts gains, ties, and losses with the ±0.05 pp display tolerance of Table 1 . Groups are defined post hoc from Luna’s three original-skill repeats.
Domain
All
Modified
Mini mod.
Luna mod.
Luna all
Cybersecurity
7
4
+22.2
+44.4
+25.4
Finance / economics
9
4
+58.3
+25.0
+3.7
Industrial / physical
14
7
+11.3
-11.9
-10.7
Mathematics / formal
7
3
+41.1
-10.6
+0.2
Media / content
5
2
+33.3
+0.0
+13.3
Natural science
14
6
+42.9
+11.1
+9.5
Appendix
Table 10: Domain-level improvement and transfer. Mini improves in every domain, while Luna differences vary by domain. Differences are percentage points. All and Modified give task counts; the modified-task and full-domain differences use their respective denominators.
Task
Mini O
Mini E
5.5 O
5.5 E
Luna O
Luna E
3d-scan-calc
R
100.0
100.0
100.0
100.0
100.0
100.0
ada-bathroom-plan-repair
R
33.3
33.3
100.0
100.0
100.0
100.0
adaptive-cruise-control
M
33.3
66.7
33.3
100.0
66.7
100.0
azure-bgp-oscillation-route-leak
R
0.0
0.0
0.0
0.0
0.0
0.0
citation-check
R
100.0
100.0
100.0
100.0
100.0
100.0
civ6-adjacency-optimizer
M
0.0
23.3
95.0
63.3
63.3
31.7
Appendix
Table 11: Complete task-level rewards. Each row reports mean reward (%) on one of the 86 tasks. M denotes modified bundles; R denotes byte-identical retained bundles. O/E denote original/final assignments. Mini’s R rows reuse the same original three-run score in both columns. Evaluation follows Section 4.1 ; Appendix B identifies the sources of the original-skill references.
Task
Original
Evolved
Change (pp)
Energy Market Pricing
66.7
0.0
−66.7
Grid Dispatch Operator
66.7
0.0
−66.7
Organize Messy Files
66.7
0.0
−66.7
Civ6 Adjacency Optimizer
63.3
31.7
−31.7
Drone Planning Control
84.4
67.8
−16.7
JPG OCR Statistics
33.3
0.0
−33.3
Appendix
Table 12: Luna tasks with observed transfer losses. Seven modified tasks lose reward under evolved bundles. Values are mean rewards (%); Change (pp) uses unrounded measurements.
Property
Value
Tasks with changed skill bundles
41/86 (47.7%)
Skill packages changed (counted per task)
62/228 (27.2%)
Total lines: original → deployed
106,632 → 108,233 (+1.5%)
Inserted + deleted lines: total / mean / median
8,147 / 198.7 / 157
Mean net line change per task: SKILL.md
-35.0
Mean net line change per task: support scripts
+74.0
Appendix
Table 13: Changes from original to deployed skill bundles. Core instructions become shorter on average while support scripts grow. The first two rows cover all 86 tasks; remaining statistics use the 41 modified tasks. Means and medians are per task. Packages and files are counted separately for each task; 228 counts the union of original and deployed packages. Line edits = insertions + deletions; net change = insertions − deletions. ρ is Spearman correlation.
Task
Generation
Retained parent
Higher-scoring saved versions
Gravitational Wave
4
e04 : 0.0
e01 : 66.7; d01 , b07 : 33.3
Data-to-D3
3
d03 : 33.3
c01 , c02 , b02 : 66.7
Appendix
Table 14: Parent eligibility beyond the historical top three. Both searches retain a parent despite three higher-scoring saved versions. Scores give each version’s latest complete three-execution mean (%) at that generation. Independent final Mini evaluation gives both tasks 100.0, up from 0.0.
Adopted lineage depth
Tasks
1
6
2
13
3
5
4
6
5
11
Appendix
Table 15: Adopted lineage depth. The 41 adopted lineages have mean depth 3.07 and median depth three. Depth counts primary-parent edits from the recorded root to deployment. It can be smaller than the generation number when a search revisits an earlier node.
Operator
Candidates
Edits on final paths
Final edits
consolidate
60
2
2
create
12
3
2
distill a
129
13
7
drop
5
0
0
extend
244
39
7
graft
64
8
3
Appendix
Table 16: Recorded edit operators. Counts cover the 41 experiment trees that produced adopted non-root skills. Candidates counts retained candidate versions, Edits on final paths counts adopted ancestry edges, and Final edits counts the 41 edges producing the final deployed bundles. These frequencies do not measure operator success rates. The 971 candidate records and 126 edits on final paths all record Mini as both diagnosis and writing model.
Task
Revision illustrated
Mini
Luna
Software dependency audit
Compare executions; refine report checks, then scan scope
0.0→66.7
0.0→100.0
PDDL airport planning
Complete a missing helper; strengthen saved-plan checks
33.3→100.0
100.0→100.0
Sales pivot analysis
Graft PDF extraction from a non-deployed sibling
0.0→33.3
0.0→100.0
Gravitational wave detection
Consolidate a non-deployed branch’s conditioning and search workflow
0.0→100.0
100.0→100.0
Appendix
Table 17: Skill-revision cases and final bundle performance. Scores show original → evolved reward (%) for complete bundles. These cases illustrate recorded revisions; the scores do not isolate individual edit effects. Evaluation settings follow Section 4.1 .
Natural-language skills are textual procedural memories through which large language model (LLM) agents retain reusable task knowledge without updating model weights. Existing methods typically treat skills as either static artifacts or monolithic documents optimized using aggregate validation scores as feedback. However, representing a skill as a monolithic document restricts optimization to its textual content, without explicitly modeling the structure through which procedural knowledge is retrieved and executed. We identify a key distinction between learning what knowledge to retain and determining how to organize it: textual updates should first be validated through execution evidence, after which the retained knowledge should be structured according to its procedural dependencies and retrieval requirements. To this end, we introduce SkillSpec, a two-phase framework comprising consensus-gated evolution and representation specialization. In the consensus-gated phase, complementary editing intents generate complete candidate skills. An update is committed only when paired evaluations reach consensus, requiring sufficient overall improvement and non-negative aggregate paired gain in every repeated evaluation. In the specialization phase, signals of process and redundancy sensitivity derived from the full optimization trajectory, including accepted and rejected candidates, guide the selection of a flat, graph, or hybrid representation.Across six benchmarks and three target language models, SkillSpec improves average success rate over SkillOpt by 6.89%, averaged across the three models. These results demonstrate that reliable skill evolution and representation specialization address complementary objectives: deciding what knowledge to retain and how to structure it for inference.
Agent skill evolution seeks to improve reusable procedural guidance for large language model (LLM) agents through iterative revision. Existing methods base each revision mainly on execution trajectories or feedback, leaving recurring behavioral requirements across tasks implicit and tying revision to the behavior of the current skill. We introduce SkillFocus, which decomposes recurring task requirements into a capability space that remains fixed as the skill evolves, separating what tasks require from how the current skill behaves. SkillFocus maps current task outcomes to this space to identify the capability that leaves the most tasks unresolved, then uses that capability to determine what to revise and which evidence to use. Across four benchmarks spanning heterogeneous tasks, SkillFocus achieves the best held-out accuracy on all four, outperforming the strongest competing result by 5.7 points on average while using 24% fewer evolution tokens on average than the closest iterative baseline. Controlled studies further show that capabilities derived from recurring task requirements outperform task-semantic and execution-derived alternatives, while randomizing task--capability assignments reduces final accuracy by up to 20.2 points. Matching evidence to the selected capability increases candidate gain by 4.4 points under prioritized revision.
Ning Wang, Zhiren Gong, Bingdong Li +2
East China Normal University · Nanyang Technological University · Southern University of Science and Technology +1
Agent skills are procedural artifacts that enable LLM agents to execute workflows, verify constraints, and recover from failures. Existing self-evolving methods refine skills using accumulated trajectories. However, they struggle in cold-start settings, where only an initial, imperfect skill is available. Consequently, skill construction defaults to expert authoring or one-shot LLM generation. Expert-authored skills are costly and may not align with how LLM agents actually execute tasks, while one-shot generated skills can be syntactically well formed yet behaviorally weak. To bridge this gap, we propose SkillRevise, an execution-grounded framework designed to iteratively refine these initial skills. SkillRevise diagnoses skill defects from execution evidence, retrieves relevant repair principles from a general memory, and applies execution-anchored edits. By re-executing candidates, it retains the first verifier-passing skill within the revision budget and falls back to empirical utility only when no candidate succeeds. Evaluated across three benchmarks and five LLMs, SkillRevise substantially outperforms one-shot baselines, improving the base agent's success rate on SkillsBench from 36.05% to 61.63%. Furthermore, the revised skills transfer across both executors and task environments, suggesting that SkillRevise captures reusable procedural knowledge beyond any single executor.
Yuxuan Liu, Zhaochen Su, Lingyun Xie +11
The Hong Kong University of Science and Technology · Harbin Institute of Technology · Harbin Institute of Technology, Shenzhen +2