SkillCome: Group Contrast Skill Optimization with Dual Memory

Haolin Li; Feng Hong; Ang Li; Chilin Fu; Weichang Wu; Ya Zhang; Yanfeng Wang; Xiaolu Zhang

cs.AISep 29, 2026

SkillCome: Group Contrast Skill Optimization with Dual Memory

Authors: Haolin Li, Feng Hong, Ang Li, Chilin Fu, Weichang Wu, Ya Zhang, Yanfeng Wang, Xiaolu Zhang, +1 more

Organizations: Fudan University · Ant Group · Shanghai Jiao Tong University

Abstract

Skill evolution improves the capabilities of large language models by analyzing trajectories generated under a given skill and modifying the skill accordingly. Existing approaches typically generate a single trajectory per question. However, this provides insufficient optimization signals since it requires inferring effective skill edits from a solitary path. It is difficult to pinpoint which actions caused the failure in a failed trajectory, or to determine which actions in a successful one should be incorporated into the skill. Furthermore, they rely on a local batch of trajectories for analysis, making the optimization direction susceptible to noisy evidence. To address these, we propose SkillCome, a Skill-evolution method based on group Contrast optimization with dual memory. For each question, SkillCome generates trajectories and performs group contrast analysis to precisely identify key behavioral divergences between successful and failed trajectories, offering reliable optimization signals. The dual memory system further accumulates evidence from historical steps to track patterns shared across different groups, leading to more generalized optimization directions. Together, SkillCome builds a systematic optimization process that transforms experience from observed successful trajectories into reusable skills. Extensive experiments on six benchmarks spanning question answering, reasoning, and agentic tasks demonstrate the effectiveness of our method. SkillCome consistently outperforms baselines across five models of varying families and scales, with gains up to +5.69 points.

Figures & tables

Figure 1: Left : An example of a skill learned by SkillCome on SpreadsheetBench. Through group contrast analysis, SkillCome learns to delete rows in descending order to avoid index shifts invalidating subsequent deletions, a skill absent from the single-rollout methods. Right : Post-hoc experiment on Qwen3.6-35B-A3B. After SkillOpt training, we perform one round of group contrast analysis. This single post-hoc skill edit can already improve performance on three datasets.

Figure 2: Overview of SkillCome. We first formulate the training as grouped rollouts. The generated trajectories are routed to either “Success Reflection” or “Failure Reflection” based on their outcome correctness. Groups with mixed outcomes undergo “Contrast Reflection”. During reflection, the optimizer model reads and updates failure memory and contrast memory to find generalizable optimization directions. The updated skill is accepted if it improves the validation score.

Model	Skill	Spreadsheet	SearchQA	LiveMath	ALFWorld	OfficeQA	DocVQA	Avg
Qwen3.8 Flash-Next	No skill	48.58	67.50	33.73	81.53	23.48	90.37	57.53
	Init skill	$-3.03$ 45.55 $-3.03$	$+0.47$ 67.97 $+0.47$	$+1.41$ 35.14 $+1.41$	$+0.94$ 82.47 $+0.94$	$+7.27$ 30.75 $+7.27$	$+1.01$ 91.38 $+1.01$	$+1.35$ 58.88 $+1.35$
	Trace2Skill	$+21.08$ 69.66 $+21.08$	$+2.43$ 69.93 $+2.43$	$-2.13$ 31.60 $-2.13$	$+3.54$ 85.07 $+3.54$	$+26.35$ 49.83 $+26.35$	$+0.94$ 91.31 $+0.94$	$+8.70$ 66.23 $+8.70$
	SkillOpt-Lite	$+25.17$ 73.75 $+25.17$	$+4.93$ 72.43 $+4.93$	$+8.02$ 41.75 $+8.02$	$+4.66$ 86.19 $+4.66$	$+36.82$ 60.30 $+36.82$	$+1.34$ 91.71 $+1.34$	$+13.49$ 71.02 $+13.49$
	SkillOpt	$+23.93$ 72.51 $+23.93$	$+5.07$ 72.57 $+5.07$	$+15.80$ 49.53 $+15.80$	$+3.17$ 84.70 $+3.17$	$+36.32$ 59.80 $+36.32$	$+0.87$ 91.24 $+0.87$	$+14.20$ 71.73 $+14.20$
	SkillCome	$+27.13$ 75.71 $+27.13$	$+6.64$ 74.14 $+6.64$	$+29.95$ 63.68 $+29.95$	$+10.26$ 91.79 $+10.26$	$+40.20$ 63.68 $+40.20$	$+2.41$ 92.78 $+2.41$	$+19.43$ 76.96 $+19.43$

Table 1: Performance across six benchmarks and five models. “–” denotes unsupported VQA evaluation for text-only LLMs. Subscripts in green / red represent the change relative to “No skill”. The best and second-best scores within each model are in bold and underlined , respectively.

Model	Method	Spreadsheet	SearchQA	LiveMath	ALFWorld	OfficeQA	DocVQA	Avg
DeepSeek-V4 Flash	SkillOpt	59.16	71.57	68.63	74.44	26.01	–	59.96
	SkillOpt w/ Memory	59.43	71.14	68.86	77.80	47.80	–	65.01
	SkillCome w/o Memory	57.92	71.43	63.92	75.37	37.84	–	61.30
	SkillCome	62.54	73.36	74.29	84.14	50.51	–	68.97
Qwen3.6 35B-A3B	SkillOpt	57.65	67.36	69.81	83.21	34.63	89.77	67.07
	SkillOpt w/ Memory	58.63	69.07	67.92	81.16	35.14	89.91	66.97

Table 2: Ablation experiment of method components on two models.

Figure 3: A case study of the evolution process of SkillCome on ALFWorld, where the agent needs to manipulate objects within a simulated embodied environment. SkillCome discovers a valuable pattern. Although the initial candidate skill is rejected, the extracted pattern remains in contrast memory. Its support group accumulates from 2 to 14 distinct tasks across optimization steps. During the process, the skill is accepted and refined multiple times.

Appendix figures & tables9 assets

Supplementary material from the paper’s appendix.

Appendix

Dataset	Train	Validation	Test	Total
SearchQA	400	200	1400	2000
SpreadsheetBench	40	79	281	400
OfficeQA	49	49	148	246
DocVQA	107	53	374	534
LiveMath	36	35	106	177
ALFWorld	200	140	134	474

Appendix

Table 3: The statistics of the datasets used in our paper. Transfer-only target datasets have no training or validation split.

Model	Skill	All	Meta	Non-Meta
GLM-5.3	No skill	36.08	14.67	52.50
	Init skill	38.44	16.30	55.42
	SkillCome	67.92	71.74	65.00
DeepSeek-V4 Flash	No skill	31.60	4.35	52.50
	Init skill	35.38	10.33	54.58
	SkillCome	74.29	96.74	57.08

Appendix

Table 4: Livemath Performance with the Meta and Non-Meta subsets.

Model	Spreadsheet	LiveMath	ALFWorld	OfficeQA	DocVQA
Qwen3.8-Flash-Next	1.78	1.81	1.83	2.09	0.69
DeepSeek-V4-Pro-0813	1.71	2.50	0.43	0.65	–
GLM-5.3	0.92	2.83	0.43	1.50	–
DeepSeek-V4-Flash	1.77	0.90	1.27	3.46	–
Qwen3.6-35B-A3B	3.39	0.90	1.65	2.01	0.86

Appendix

Table 5: Standard deviations of SkillCome. Mean scores are provided in Table 1 . “–” denotes unsupported evaluation.

Target model	Mixed groups	Proportion
DeepSeek-V4-Flash	142 / 430	33.02%
Qwen3.6-35B-A3B	192 / 480	40.00%
GLM-5.3	107 / 430	24.88%
DeepSeek-V4-Pro-0813	152 / 430	35.35%
Qwen3.8-Flash-Next	138 / 480	28.75%
Total	731 / 2250	32.49%

Appendix

Table 6: Mixed-group counts and proportions by target model. A mixed group contains both successful and failed rollouts of the same task. Counts are reported as mixed groups / total groups.

Model	Skill	Spreadsheet	SearchQA	LiveMath
Qwen3.8 Flash-Next	No skill	48.58	67.50	33.73
	Init skill	$-3.03$ 45.55 $-3.03$	$+0.47$ 67.97 $+0.47$	$+1.41$ 35.14 $+1.41$
	SkillCome	$+27.13$ 75.71 $+27.13$	$+6.64$ 74.14 $+6.64$	$+29.95$ 63.68 $+29.95$
	Codex-SkillCome	$+36.56$ 85.14 $+36.56$	$+10.79$ 78.29 $+10.79$	$+34.19$ 67.92 $+34.19$
DeepSeek-V4 Pro-0813	No skill	47.78	66.57	28.07
	Init skill	$+0.62$ 48.40 $+0.62$	$+0.50$ 67.07 $+0.50$	$+3.30$ 31.37 $+3.30$

Appendix

Table 7: Performance comparison of SkillCome under direct chat and the Codex CLI harness.

Target Model	Skill	Spreadsheet	SearchQA	LiveMath	ALFWorld	OfficeQA	Avg
DeepSeek-V4 Flash	No skill	47.33	64.85	31.60	59.51	5.41	41.74
	Init skill	$+1.25$ 48.58 $+1.25$	$+0.58$ 65.43 $+0.58$	$+3.78$ 35.38 $+3.78$	$-2.60$ 56.91 $-2.60$	$+6.25$ 11.66 $+6.25$	$+1.85$ 43.59 $+1.85$
	Qwen3.6 $\rightarrow$ Flash	$+12.72$ 60.05 $+12.72$	$+7.36$ 72.21 $+7.36$	$+37.27$ 68.87 $+37.27$	$+14.00$ 73.51 $+14.00$	$+16.89$ 22.30 $+16.89$	$+17.65$ 59.39 $+17.65$
	Flash $\rightarrow$ Flash	$+15.21$ 62.54 $+15.21$	$+8.51$ 73.36 $+8.51$	$+42.69$ 74.29 $+42.69$	$+24.63$ 84.14 $+24.63$	$+45.10$ 50.51 $+45.10$	$+27.23$ 68.97 $+27.23$
Qwen3.6 35B-A3B	No skill	41.99	66.00	26.65	66.04	37.16	47.57
	Init skill	$-1.24$ 40.75 $-1.24$	$+0.23$ 66.23 $+0.23$	$+3.54$ 30.19 $+3.54$	$+13.44$ 79.48 $+13.44$	$+1.52$ 38.68 $+1.52$	$+3.50$ 51.07 $+3.50$

Appendix

Table 8: Cross-target skill transferability experiments of SkillCome. “Qwen3.6

\rightarrow

Flash” denotes transferring the skill learned on Qwen3.6-35B-A3B to DeepSeek-V4-Flash. “Flash

\rightarrow

Flash” is the original performance of SkillCome on the DeepSeek-V4-Flash.

Source dataset	Target dataset	Model	No skill	Transferred skill	$\bm{\Delta}$
SearchQA	HotpotQA	DeepSeek-V4-Pro-0813	63.20	64.80	+1.60
		Qwen3.8-Flash-Next	66.30	67.90	+1.60
		DeepSeek-V4-Flash	62.70	64.80	+2.10
		Qwen3.6-35B-A3B	64.80	65.70	+0.90
OlympiadBench	Omni-MATH	DeepSeek-V4-Pro-0813	72.30	74.10	+1.80
		Qwen3.8-Flash-Next	86.40	86.60	+0.20

Appendix

Table 9: Full results of cross-dataset skill transferability.

\Delta

denotes the score improvement of transferred skill over no skill.

Model	Skill	Spreadsheet	SearchQA	LiveMath	ALFWorld	OfficeQA	DocVQA	Avg
DeepSeek-V4 Flash	No skill	47.33	64.85	31.60	59.51	5.41	–	41.74
	SkillCome	$+11.12$ 58.45 $+11.12$	$+6.94$ 71.79 $+6.94$	$+23.35$ 54.95 $+23.35$	$+19.59$ 79.10 $+19.59$	$+1.68$ 7.09 $+1.68$	–	$+12.54$ 54.28 $+12.54$
	Init skill	48.58	65.43	35.38	56.91	11.66	–	43.59
	SkillCome	$+13.96$ 62.54 $+13.96$	$+7.93$ 73.36 $+7.93$	$+38.91$ 74.29 $+38.91$	$+27.23$ 84.14 $+27.23$	$+38.85$ 50.51 $+38.85$	–	$+25.38$ 68.97 $+25.38$
Qwen3.6 35B-A3B	No skill	41.99	66.00	26.65	66.04	37.16	89.77	54.60
	SkillCome	$+16.02$ 58.01 $+16.02$	$+4.07$ 70.07 $+4.07$	$+38.21$ 64.86 $+38.21$	$+21.09$ 87.13 $+21.09$	$-3.88$ 33.28 $-3.88$	$+0.27$ 90.04 $+0.27$	$+12.63$ 67.23 $+12.63$

Appendix

Table 10: Performance of SkillCome evolving from empty skills.

Target model	Dataset	SkillCome (M)	SkillOpt (M)	Ratio
DeepSeek-V4 Flash	SpreadsheetBench	0.69	1.50	$0.460\times$
	SearchQA	1.69	4.64	$0.364\times$
	OfficeQA	0.97	2.20	$0.441\times$
	LiveMath	0.56	0.90	$0.622\times$
	ALFWorld	1.09	1.57	$0.694\times$
Qwen3.6 35B-A3B	SpreadsheetBench	0.56	1.31	$0.427\times$

Appendix

Table 11: Optimizer token usage in millions (M). The ratio is computed as SkillCome/ SkillOpt.

Explore similar work

Skill-R1: Agent Skill Evolution via Reinforcement Learning
May 10, 2026Yash Vishe, Rohan Surana, Xunyi Jiang +8Self-Evolving Skill Libraries Unsupervised Skill Discovery
UC San Diego · Adobe Research
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
SkillFocus: Evolving Agent Skills via Capability Decomposition
Sep 28, 2026Ning Wang, Zhiren Gong, Bingdong Li +2Self-Evolving Skill Libraries Skills
East China Normal University · Nanyang Technological University · Southern University of Science and Technology +1

May 10, 2026, cs.LGJ/K move · Enter open · S save

Skill-R1: Agent Skill Evolution via Reinforcement Learning

Yash Vishe, Rohan Surana, Xunyi Jiang, Zihan Huang+7

UC San Diego · Adobe Research

Self-Evolving Skill Libraries Unsupervised Skill Discovery Skill Evolution Agentic Reinforcement Learning Skills

Open paper AskPDF arXiv

Agentic large language models often rely on skills, reusable natural language procedures that guide planning, action, and tool use. In practice, skills are typically improved through prompt engineering or by aligning the task LLM itself, which is costly, model-specific, and often infeasible for closed-source models. Skill optimization is not a one-step problem but a recurrent process with two coupled levels of credit assignment: a useful skill must improve rollout quality under current conditioning, while a useful revision must turn observed outcomes into a better skill for the next round. We propose Skill-R1, a reinforcement learning framework for instance-level recurrent skill optimization from verifiable rewards. Rather than updating the task LLM, Skill-R1 trains a lightweight skill generator that conditions on the task context, prior rollouts, and their verified outcomes to produce skills that steer a frozen task LLM. This preserves black-box compatibility with both open- and closed-source models while making adaptation substantially cheaper than model-level updates. Skill-R1 proceeds over multiple generations: at each step, the current skill induces rollouts whose verified outcomes are fed back to produce the next revision. To optimize this recurrent process, we introduce a bi-level group-relative policy optimization objective combining intra-generation and inter-generation advantages. The intra-generation term compares rollouts under shared skill conditioning, while the inter-generation term rewards revisions that improve behavior across successive generations. Together, these provide a principled objective for directional skill evolution rather than one-shot self-refinement. Empirically, Skill-R1 achieves consistent gains over no-skill baselines and standard GRPO across benchmarks with verifiable rewards, with particularly strong improvements on complex, multi-step tasks.

Abstract

Explore similar work

Skill-R1: Agent Skill Evolution via Reinforcement Learning

SkillGrad: Optimizing Agent Skills Like Gradient Descent

SkillFocus: Evolving Agent Skills via Capability Decomposition