From Trajectories to Differentiated Skills: Cognitive Differentiation for Self-Improving Agents

Yewei Zhao; Derong Xu; Wei Chen; Yin Xu; Peilun Zhou; Xiaohan Wang; Guojun Yin; Wei Lin

cs.AIOct 16, 2026

From Trajectories to Differentiated Skills: Cognitive Differentiation for Self-Improving Agents

Authors: Yewei Zhao, Derong Xu, Wei Chen, Yin Xu, Peilun Zhou, Xiaohan Wang, Guojun Yin, Wei Lin,

Organizations: University of Science and Technology of China · City University of Hong Kong · Meituan

Abstract

Large language model (LLM) agents tackle complex tasks through reasoning and interaction with external environments. Agent skills provide persistent, transferable capabilities that guide task planning, inform contextual decisions, support tool use, and enable executable procedures. Yet existing skill-evolution methods largely organize knowledge around skill artifacts, overlooking the functional heterogeneity within individual skills and its implications for knowledge maintenance and use. Inspired by cognitive science, we propose Cognitive Differentiation in Skill Evolution (CDSE), which evolves a unified skill through four functionally distinct yet coordinated components: Meta, Experience, Tool, and Script. CDSE extracts evidence-grounded semantic gradients from successful and failed trajectories, consolidates them by functional role, and applies component-specific updates and validation with cross-component coordination and adaptive invocation. An analysis of 500 top-ranked skill packages shows that 96.8% contain knowledge spanning at least two functional roles. Across six benchmarks, three agent models, and two execution harnesses, CDSE achieves the highest scores in all 36 settings, improving over the Base agent by 7.6 percentage points and the strongest competing method in each setting by 2.6 points on average. Ablation studies further support the importance of functional differentiation and adaptive invocation.

Figures & tables

Panel of Figure 1 — Figure 1: Performance and framework of CDSE.

Figure 2: Functional composition of 500 widely used skill packages. Left: prevalence of each component; right: the number of components per skill. Component labels are non-exclusive.

Figure 3: CDSE discovers semantic gradients from successful and failed executions, consolidates them by functional role, and updates the four components through specialized maintenance and coordination. Revised invocation descriptions connect the evolving skill to subsequent agent behavior.

Agent model	Method	OfficeQA	Spreadsheet2	SealQA	LiveMath	SWE-Pro	Automation
GLM-5.2 (Claude Code)	Base	71.5	24.0	45.9	83.0	60.2	20.8
	Trace2Skill	74.1 +2.6	28.3 +4.3	48.2 +2.3	82.3 -0.7	62.6 +2.4	21.4 +0.6
	SkillOpt-lite	76.4 +4.9	26.8 +2.8	49.7 +3.8	85.0 +2.0	63.1 +2.9	24.7 +3.9
	EvoSkill	76.0 +4.5	25.9 +1.9	52.4 +6.5	84.4 +1.4	62.6 +2.4	25.5 +4.7
	CDSE	77.2 +5.7	30.7 +6.7	54.1 +8.2	87.7 +4.7	68.4 +8.2	32.9 +12.1
DeepSeek-V4-Flash (Claude Code)	Base	68.0	16.8	51.8	59.9	57.8	14.3

Table 1: Task scores (%) across models, harnesses, and skill-evolution methods.

Figure 4: Component removal effects and invocation behavior on GLM-5.2.

Figure 5: Score gains across three CDSE evolution rounds.

Figure 6: Cross-model and cross-harness transfer gains.

Source	Component	OfficeQA	Spreadsheet2	SealQA	LiveMath	SWE-Pro	Automation
Spreadsheet2	Meta	+4.1 (18%)	–	+4.3 (20%)	+0.0 (0%)	+4.1 (10%)	+7.1 (28%)
	Experience	+0.8 (6%)	–	-2.8 (14%)	-0.4 (0%)	+0.0 (7%)	-1.6 (2%)
	Tool	-1.3 (1%)	–	+0.0 (0%)	+1.0 (0%)	-0.6 (1%)	+0.0 (0%)
	Script	-0.6 (2%)	–	-0.9 (0%)	+0.7 (0%)	-3.9 (3%)	-1.6 (0%)
Automation	Meta	+5.7 (10%)	+1.4 (19%)	+6.7 (35%)	+0.2 (0%)	+2.6 (6%)	–
	Experience	+2.2 (7%)	-1.2 (0%)	+3.8 (38%)	-0.8 (0%)	+0.0 (4%)	–

Table 2: Cross-benchmark component transfer on GLM-5.2.

Variant	Spreadsheet Bench2	SWE-Pro	Automation Bench
Base	24.0	60.2	20.8
Full CDSE	30.7	68.4	32.9
w/o Cognitive Differentiation	25.8	62.6	24.3
w/o Specialized Update	27.1	65.4	28.0
w/o Cross-Component Coordination	29.6	67.5	30.7
w/o Adaptive Invocation Interface	26.4	64.3	27.1

Table 3: Mechanism ablations on GLM-5.2 with Claude Code. All entries are task scores (%).

Method	Spreadsheet2	SWE-Pro	Automation
Trace2Skill	15.41	41.30	23.90
SkillOpt-lite	29.33	59.82	6.56
EvoSkill	42.67	75.47	5.06
CDSE	7.60	15.04	2.19

Table 4: Input tokens per percentage point of score gain over Base after one evolution round (million tokens/pp,

\downarrow

), using GLM-5.2 with Claude Code.

Appendix figures & tables14 assets

Supplementary material from the paper’s appendix.

Appendix

	OfficeQA		Spreadsheet Bench2		SealQA		LiveMathBench		SWE-Pro		Automation Bench
Method	Input	Output	Input	Output	Input	Output	Input	Output	Input	Output	Input	Output
Trace2Skill	35.80	1.350	49.30	2.700	10.60	0.480	9.51	1.050	123.90	5.600	23.90	1.050
SkillOpt-lite	38.50	1.450	52.80	2.900	11.40	0.520	10.10	1.150	131.60	5.950	25.60	1.150
EvoSkill	37.20	1.410	51.20	2.820	11.10	0.510	9.89	1.110	128.30	5.830	24.80	1.110
CDSE	37.00	1.400	50.90	2.800	11.00	0.500	9.84	1.100	127.80	5.800	24.70	1.100

Appendix

Table 5: Input and output token consumption (millions of tokens) in a single evolution round with GLM-5.2 and Claude Code.

	OfficeQA		Spreadsheet Bench2		SealQA		LiveMathBench		SWE-Pro		Automation Bench
Method	Input	Output	Input	Output	Input	Output	Input	Output	Input	Output	Input	Output
Trace2Skill	12.34	0.466	15.41	0.844	5.89	0.267	N/A	N/A	41.30	1.867	23.90	1.050
SkillOpt-lite	6.64	0.250	29.33	1.611	3.17	0.144	11.22	1.278	59.82	2.705	6.56	0.295
EvoSkill	10.94	0.415	42.67	2.350	1.61	0.074	6.59	0.740	75.47	3.429	5.06	0.227
CDSE	7.87	0.298	7.60	0.418	1.55	0.070	3.08	0.344	15.04	0.682	2.19	0.097

Appendix

Table 6: Single-round input and output tokens per percentage point of score gain over Base on all six benchmarks (million tokens/pp,

\downarrow

), using GLM-5.2 with Claude Code. Bold marks column minima before rounding. N/A indicates no positive gain: Trace2Skill decreases the LiveMathBench score by 1.4 points.

Benchmark	Meta	Exp.	Tool	Script
OfficeQA	1.60	1.10	1.80	3.60
SpreadsheetBench2	1.40	2.70	4.90	12.30
SealQA	1.00	1.13	2.24	2.67
LiveMathBench	6.25	2.76	–	–
SWE-Pro	1.30	1.70	3.80	10.20
AutomationBench	1.10	1.10	1.10	1.80

Appendix

Table 7: Mean first invocation turn of the four skill components in Figure 4 , evaluated with GLM-5.2 under Claude Code. A dash indicates no invocation.

Figure 7: Evolution over three rounds on all six benchmarks with GLM-5.2, DeepSeek-V4-Flash, and MiniMax-M3 under Claude Code. Scores are shown as percentage-point gains over each model’s no-skill baseline (R0). Each curve shows one reported score at each round; the same trajectory corpus is reused across rounds. The first round provides most of the gain, with smaller, benchmark-dependent changes thereafter.

Benchmark	R0	R1	R2	R3
OfficeQA	71.5	76.2	77.8	77.2
SpreadsheetBench2	24.0	30.7	29.9	30.7
SealQA	45.9	53.0	53.5	54.1
LiveMathBench	83.0	86.2	87.3	87.7
SWE-Pro	60.2	68.7	69.0	68.4
AutomationBench	20.8	32.1	33.3	32.9

Appendix

Table 8: Multi-round evolution scores (%) for GLM-5.2 under Claude Code. R0 is the no-skill baseline.

Benchmark	R0	R1	R2	R3
OfficeQA	68.0	76.5	77.3	77.3
SpreadsheetBench2	16.8	26.9	26.9	26.9
SealQA	51.8	59.4	59.8	59.4
LiveMathBench	59.9	68.7	68.3	69.1
SWE-Pro	57.8	64.7	63.5	64.1
AutomationBench	14.3	27.6	28.2	28.9

Appendix

Table 9: Multi-round evolution scores (%) for DeepSeek-V4-Flash under Claude Code. R0 is the no-skill baseline.

Benchmark	R0	R1	R2	R3
OfficeQA	71.6	80.8	81.4	81.8
SpreadsheetBench2	25.3	34.1	34.7	34.4
SealQA	49.4	57.9	57.5	57.9
LiveMathBench	73.5	82.2	82.5	82.7
SWE-Pro	57.2	65.0	64.7	65.3
AutomationBench	9.1	18.8	19.5	20.0

Appendix

Table 10: Multi-round evolution scores (%) for MiniMax-M3 under Claude Code. R0 is the no-skill baseline.

	GLM-5.2			DeepSeek-V4-Flash
Benchmark	Base	N	T	Base	N	T
OfficeQA	71.5	77.2	72.5	68.0	77.3	72.5
SpreadsheetBench2	24.0	30.7	30.0	16.8	26.9	20.7
SealQA	45.9	54.1	49.4	51.8	59.4	48.5
LiveMathBench	83.0	87.7	85.8	59.9	69.1	71.6
SWE-Pro	60.2	68.4	63.5	57.8	64.1	60.4

Appendix

Table 11: Cross-model transfer scores (%) for Figure 6 (Claude Code fixed). Column groups identify the receiving configuration. N: Native; T: Transferred from the other configuration. Gains are relative to Base.

	Claude Code			Codex
Benchmark	Base	N	T	Base	N	T
OfficeQA	71.5	77.2	75.3	68.8	74.9	74.3
SpreadsheetBench2	24.0	30.7	28.6	22.7	31.3	26.1
SealQA	45.9	54.1	47.5	22.6	27.6	25.5
LiveMathBench	83.0	87.7	84.2	79.0	83.5	80.2
SWE-Pro	60.2	68.4	64.2	62.3	68.9	65.6

Appendix

Table 12: Cross-harness transfer scores (%) for Figure 6 (GLM-5.2 fixed). Column groups identify the receiving configuration. N: Native; T: Transferred from the other configuration. Gains are relative to Base.

Component	Precision	Recall	F1
Meta	0.860	0.840	0.850
Experience	0.880	0.840	0.860
Tool	0.900	0.860	0.880
Script	0.920	0.900	0.910
Macro average	0.890	0.860	0.875

Appendix

Table 13: LLM annotation quality against expert labels on a random 100-package subset. F1 is computed from the reported precision and recall.

Component	ID	Label	Count	Percentage
Meta	M1	Task Planning & Decomposition	356	71.2%
Meta	M2	Progress & Resource Control	267	53.4%
Meta	M3	Global Verification & Evidence Control	214	42.8%
Meta	M4	Stopping & Finalization	345	69.0%
Experience	E1	Context-Conditioned Heuristic	486	97.2%
Experience	E2	Failure Pattern & Anti-Pattern	421	84.2%

Appendix

Table 14: Fine-grained label prevalence among 500 skills. Labels are non-exclusive; each count is the number of skills containing that label.

Component	Count / 500	Percentage
Meta	443	88.6%
Experience	493	98.6%
Tool	398	79.6%
Script	119	23.8%

Appendix

Table 15: Component prevalence. A skill is counted once for a component if it contains any of that component’s labels.

Components	Skills	Percentage
0	0	0.0%
1	16	3.2%
2	104	20.8%
3	291	58.2%
4	89	17.8%
Total	500	100.0%

Appendix

Table 16: Distribution of the number of functional components per skill. Unlike component marginals, these categories are mutually exclusive.

Benchmark	Task family	Dataset selection
OfficeQA	Document-grounded question answering	Easy and Hard subsets
SpreadsheetBench2	Spreadsheet manipulation	Financial_Model and Template subsets
SealQA	Long-context evidence synthesis	Complete dataset
LiveMathBench	Mathematical reasoning	202605 and 202606 subsets
SWE-Pro	Repository-level software engineering	Three Python subsets: pandas , scikit-learn , and xarray
AutomationBench	Tool-based business workflows	All instances from the Sales and Marketing domains

Appendix

Table 17: Task families and dataset selections. Each selected collection is split into training and test sets at a 2:7 ratio.

LLM Agent Skill Learning

Large language model agent skill learning studies how agents acquire, represent, reuse, and refine procedures for multi-step tasks.

See all 398 papers

SkillFocus: Evolving Agent Skills via Capability Decomposition
Sep 28, 2026Ning Wang, Zhiren Gong, Bingdong Li +2LLM Agent Skill Learning Task Decomposition
East China Normal University · Nanyang Technological University · Southern University of Science and Technology +1
SkillCommit: Evolving Agent Skills through Behaviorally Validated Scope Expansion
Aug 15, 2026Yu He, Weikai YangLLM Agent Self-Improvement LLM Agent Skill Learning
Nanjing University · The Hong Kong University of Science and Technology (Guangzhou)
SkillGrad: Optimizing Agent Skills Like Gradient Descent
May 26, 2026Hanyu Wang, Yifan Lan, Bochuan Cao +2Large Language Model-Guided Optimization LLM Agent Self-Improvement
College of Information Sciences and Technology The Pennsylvania State University University Park, PA, USA

Sep 28, 2026J/K move · Enter open · S save

SkillFocus: Evolving Agent Skills via Capability Decomposition

Ning Wang, Zhiren Gong, Bingdong Li, Peng Yang+1

East China Normal University · Nanyang Technological University · Southern University of Science and Technology+1

LLM Agent Skill Learning Task Decomposition

Open paper AskPDF arXiv

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.

Abstract

LLM Agent Skill Learning

Explore similar work

SkillFocus: Evolving Agent Skills via Capability Decomposition

SkillCommit: Evolving Agent Skills through Behaviorally Validated Scope Expansion

SkillGrad: Optimizing Agent Skills Like Gradient Descent