SkillFocus: Evolving Agent Skills via Capability Decomposition

Ning Wang; Zhiren Gong; Bingdong Li; Peng Yang; Aimin Zhou

cs.AISep 28, 2026

SkillFocus: Evolving Agent Skills via Capability Decomposition

Authors: Ning Wang, Zhiren Gong, Bingdong Li, Peng Yang, Aimin Zhou

Organizations: East China Normal University · Nanyang Technological University · Southern University of Science and Technology · Shanghai Innovation Institute

Abstract

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.

Figures & tables

Figure 1: Existing and capability-space skill evolution. Both follow the same execution–revision loop. Existing methods base revisions on current execution evidence, whereas ours also derives a capability space from tasks before execution and uses it during revision.

Figure 2: Overview of SkillFocus. Before evolution, capability induction builds a fixed capability space

(\mathcal{C},Q)

from task specifications, from which the selection set is built. In each round, capability scores select

c^{*}

, which sets the requirement to improve and the matched execution evidence; the resulting candidate replaces the current skill only after selection and training validation.

Method	SearchQA	SpreadsheetBench	LiveMath	IFBench	Avg.
No Skill	64.5	21.8	14.4	67.8	42.1
One-shot Skill	70.7	34.0	30.6	80.6	54.0
Trace2Skill	73.7	30.7	39.5	76.3	55.1
GEPA	74.6	58.9	37.9	73.4	61.2
SkillOpt	76.1	62.1	41.1	73.9	63.3
SkillFocus	80.1	77.1	43.6	82.0	70.7

Table 1: Test accuracy (%) of SkillFocus and baselines on four benchmarks. Bold and underlined values denote the best and second-best results, respectively; Avg.: unweighted mean over the four benchmarks.

Figure 3: Evolution trajectories of SkillFocus and SkillOpt. Each curve plots selection set accuracy gain over the initial skill in percentage points; markers show accepted updates, and horizontal segments are rejected candidates that consume tokens without changing the retained skill.

Figure 4: Task–capability assignments and their effect on final skill accuracy. (a) capabilities assigned to each task and the number of tasks per capability; (b) overlap between the task sets of different capabilities; (c) final skill accuracy when capabilities are derived from different sources. Panel annotations report labels per task, coverage, and median support overlap.

Condition	Accept (%)	$\Delta_{\rm sel}$	$\Delta_{\rm tr}$
TM: Top + Matched	40.0	+4.10	+9.37
T $\neg$ M: Top + Mismatched	26.7	+3.62	+4.97
RM: Random + Matched	26.7	+2.28	+5.33
R $\neg$ M: Random + Mismatched	20.0	+2.09	+3.39

Table 2: Focus–evidence crossover over 15 historical revision states, macro-averaged across SearchQA, SpreadsheetBench, and LiveMath. Accept: candidate acceptance rate;

\Delta_{\rm sel}

\Delta_{\rm tr}

: paired candidate gains (points) on the selection and training sets.

Appendix figures & tables17 assets

Supplementary material from the paper’s appendix.

Appendix

Benchmark	Tasks	$K$	Labels per task	Support	Mean $J$	Median $J$	Coverage
SearchQA	600	19	4.18	3–600	.055	.022	100%
SpreadsheetBench	120	34	2.46	3–89	.018	.000	99.2%
LiveMath	53	12	5.11	5–53	.202	.167	100%
IFBench	89	21	1.29	3–18	.018	.000	86.5%

Appendix

Table 3: Benchmark-level statistics of the frozen task–capability matrices. Tasks is the number of task records the matrix covers;

K

is the number of induced capabilities; Labels per task is the mean number of active capabilities per task; Support is the minimum–maximum number of supporting tasks;

J

is pairwise support-set Jaccard overlap; and Coverage is the fraction of covered tasks carrying at least one capability.

Benchmark	Representative behavioral requirement	# Tasks
SearchQA	Return exactly one non-empty answer element containing only the final answer text, with no reasoning, citations, or extra prose.	600
SearchQA	Use only the supplied question, documents, and metadata as evidence, avoiding outside knowledge, tools, and unsupported inference.	600
SearchQA	Select an answer only when the same candidate satisfies all substantive clue constraints, rather than combining separate partial matches.	191
SearchQA	Follow the relation expressed in the clue to extract the correct participant, such as an author, actor, founder, owner, or creator.	127
SpreadsheetBench	Write the exact formula or formula-based expression required by the task directly in the specified answer cells.	89
SpreadsheetBench	Generate each output with correct row-specific references, conditions, and source matches rather than positional assumptions.	31

Appendix

Table 4: Representative behavioral requirements and their task support.

capability source	SearchQA	SSB	LiveMath
SkillFocus	80.1	77.1	43.6
task-semantic	61.3	68.9	25.8
execution-derived	73.5	66.1	30.7
randomized assignment	69.0	59.6	23.4

Appendix

Table 5: Test accuracy (%) when capabilities are derived from different sources, under the same downstream SkillFocus pipeline.

Benchmark	Views	$K$	Common refs.	ARI	NMI
SearchQA	3	18/13/17	47.7	.883	.953
SSB	3	17/16/19	35.7	.913	.976
LiveMath	3	11/11/10	33.3	.973	.989
IFBench	3	30/24/42	112.3	.912	.974

Appendix

Table 6: Assignment agreement across three independent proposal calls on the same task specifications.

K

: proposal-specific capability counts; Common refs.: aligned behavioral-requirement references; ARI and NMI: adjusted Rand index and normalized mutual information between proposal assignments.

SearchQA
Condition	Accept (%)	$\Delta_{\rm sel}$	$\Delta_{\rm tr}$
TM	40.0	$+5.30$	$+10.50$
T $\neg$ M	20.0	$+5.30$	$+3.42$
RM	40.0	$+7.50$	$+3.92$
R $\neg$ M	40.0	$+3.13$	$+1.92$
SpreadsheetBench

Appendix

Table 7: Per-benchmark focus–evidence crossover (five historical revision states per benchmark). Accept: candidate acceptance rate;

\Delta_{\rm sel}

\Delta_{\rm tr}

: paired candidate gains (points) on the selection and training sets.

Variant	SearchQA	SSB	LiveMath
SkillFocus	80.1	77.1	43.6
w/o matched evidence	71.4	58.6	20.8
w/o capability prioritization	68.7	38.3	20.8

Appendix

Table 8: Ablations of capability-guided revision on SearchQA, SpreadsheetBench, and LiveMath (test accuracy, %).

Benchmark	capability	Rounds	Support	Share
SearchQA	$c_{1}$	3	600	100.0%
SearchQA	$c_{2}$	1	600	100.0%
SearchQA	$c_{3}$	1	289	48.2%
SpreadsheetBench	$c_{1}$	5	89	74.2%
LiveMath	$c_{1}$	2	53	100.0%
LiveMath	$c_{3}$	2	31	58.5%

Appendix

Table 9: Capabilities selected by the priority rule over the five rounds of each formal run. Support is the number of tasks requiring the capability in the frozen

Q

; Share is its fraction of the task population. A capability is selected in consecutive rounds when it is the only eligible one.

Benchmark	Record	T/N $n$	$\Delta_{T}/\Delta_{N}$	$R_{d}(e)$
SQA	Screen.	93/107	$+1.1/+0.0$	$+1.1$
SSB	Acc.-1	89/31	$+46.1/+16.1$	$+29.9$
LM	Acc.	31/22	$+6.5/+4.5$	$+1.9$
IFB	Acc.	3/86	$+0.0/+8.1$	$-8.1$

Appendix

Table 10: Target and non-target edit responses. T/N: group sizes;

\Delta_{T}/\Delta_{N}

: net changes (points);

R_{d}(e)

: their difference; Screen./Acc.: screening record/accepted candidate (Acc.-1: first accepted).

	SkillFocus		SkillOpt
Benchmark	Sel.	Test	Sel.	Test
SearchQA	49.5	80.1	76.0	76.1
SpreadsheetBench	80.0	77.1	85.0	62.1
LiveMath	38.9	43.6	44.4	41.1
IFBench	77.8	82.0	89.7	73.9

Appendix

Table 11: Final selection set and held-out test accuracy (%). The two columns cover different task sets, so their difference is not a calibrated generalization gap across methods.

Figure 5: Representative capability-guided revision on SpreadsheetBench. The case connects diverse task requests to a recurring behavioral requirement, an evidence-grounded revision, and paired target/non-target responses.

Group	$n$	Repair/reg.	Net $\Delta$
Target	89	42/1	$+46.1\%$
Non-target	31	7/2	$+16.1\%$

Appendix

Table 12: Paired response of the representative SpreadsheetBench revision. Repair/reg. counts failure-to-success and success-to-failure changes; the target-minus-non-target contrast is

+29.9

points.

Benchmark	Task format	Evaluation
SearchQA	Jeopardy!-style clues answered in a closed-context, single-turn setting	Deterministic answer matching
SpreadsheetBench	Spreadsheet manipulation on the Verified subset; the agent reads and writes workbook files	Target cells compared with the reference workbook
LiveMath	Multiple-choice mathematical reasoning with deterministically shuffled options; code execution is disabled	Exact match of the selected option
IFBench	Single-turn free-text responses under verifiable output constraints	Programmatic constraint verification

Appendix

Table 13: Benchmarks used in the experiments. Task format states what the agent reads and produces; Evaluation states the deterministic checker applied to each response.

Condition	No Skill	One-shot Skill	Trace2Skill	GEPA	SkillOpt	SkillFocus
Evaluator and test split	Shared	Shared	Shared	Shared	Shared	Shared
Target model	DeepSeek-V4-Flash	DeepSeek-V4-Flash	DeepSeek-V4-Flash	DeepSeek-V4-Flash	DeepSeek-V4-Flash	DeepSeek-V4-Flash
Initial skill	–	–	Shared	Shared	Shared	Shared
Optimizer	–	–	GPT-5.4	GPT-5.4	GPT-5.4	GPT-5.4
Evolution budget	–	–	One analysis and consolidation pass	160 metric calls	Official setting	At most five rounds, one screened candidate per round
Candidate retention	–	–	Single consolidated skill	Pareto-based selection	Retained-skill update	Positive paired change on selection and training sets

Appendix

Table 14: Shared and method-specific conditions across compared methods. “–” denotes a condition that does not apply to the method.

Benchmark	C/A	Init.	Iter.	$C_{\mathrm{evo}}$	Test
SQA	5/2	5.10M	15.55M	20.66M	8.51M
SSB	5/2	6.45M	22.61M	29.06M	4.73M
LM	5/2	0.38M	5.98M	6.36M	3.40M
IFB	5/1	0.97M	2.48M	3.45M	1.58M

Appendix

Table 15: SkillFocus inference-token ledger (millions). SQA, SSB, LM, and IFB abbreviate the four benchmarks; C/A: proposed/accepted candidates; Init., Iter.: initialization and iterative evolution.

Benchmark	SkillFocus	SkillOpt	Relative change
SearchQA	20.66	45.21	$-54.3\%$
SpreadsheetBench	29.06	40.55	$-28.3\%$
LiveMath	6.36	8.14	$-21.9\%$
IFBench	3.45	3.21	$+7.5\%$
Mean	–	–	$-24.0\%$

Appendix

Table 16: Pre-final-test evolution cost (millions of inference tokens) of SkillFocus and SkillOpt. The final row averages the per-benchmark relative changes.

Interface	Role	Calls
capability induction
ExtractRequirements	task $\rightarrow$ requirements	per task
ProposeCapabilities	cross-task proposals	3
Consensus	frozen registry	1
Assign	requirement $\rightarrow$ capability	batched
capability revision

Appendix

Table 17: Overview of the prompt interfaces and their call multiplicity in the formal runs.

Stage / perspective	Frozen operational condition
Diagnosis: missing guidance	No current-skill rule addresses the operation; the explorer leaves cited_rules empty.
Diagnosis: ineffective guidance	An existing rule addresses the operation but fails; the explorer quotes at least one such rule.
Diagnosis: behavior organization	Interacting rules create an ordering, conditioning, or precedence problem; the explorer quotes at least two rules.
Generation: minimal addition	Propose one non-duplicative standalone rule with op=add .
Generation: revise existing guidance	Replace the core content of one existing rule, quoting the rule being revised.
Generation: structural change	Change a rule’s trigger, ordering, precedence, interaction, or removal target, quoting the affected rule(s).

Appendix

Table 18: Frozen Explorer perspectives in the formal runs.

Explore similar work

SkillSpec: Consensus-Gated Agent Skill Evolution via Representation Specialization
Sep 30, 2026Huancheng Chen, Xiaodi Sun, Zhaoqiong Huang +2Self-Evolving Skill Libraries Skills
Microsoft AI
SkillAxe: Sharpening LLM-Authored Agent Skills Through Evaluation-Guided Self-Refinement
Jun 9, 2026Srishti Gautam, Arjun Radhakrishna, Sumit GulwaniSkills Instruction
Microsoft
SkillGrad: Optimizing Agent Skills Like Gradient Descent
May 26, 2026Hanyu Wang, Yifan Lan, Bochuan Cao +2Skills Large Language Model Agents
College of Information Sciences and Technology The Pennsylvania State University University Park, PA, USA

Sep 30, 2026, cs.LGJ/K move · Enter open · S save

SkillSpec: Consensus-Gated Agent Skill Evolution via Representation Specialization

Huancheng Chen, Xiaodi Sun, Zhaoqiong Huang, Shenyang Huang Shreya Singhal+1

Microsoft AI

Self-Evolving Skill Libraries Skills Large Language Model Agents Reusable Knowledge Semantic Memory

Open paper AskPDF arXiv

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.

Abstract

Explore similar work

SkillSpec: Consensus-Gated Agent Skill Evolution via Representation Specialization

SkillAxe: Sharpening LLM-Authored Agent Skills Through Evaluation-Guided Self-Refinement

SkillGrad: Optimizing Agent Skills Like Gradient Descent