MOSCOPT: Mixture-of-Skills Collective Optimization for LLM Agents

Zhenyu Zhang; Jiudong Yang

cs.AISep 13, 2026

MOSCOPT: Mixture-of-Skills Collective Optimization for LLM Agents

Authors: Zhenyu Zhang, Jiudong Yang

Organizations: ML Center, Coupang, Shanghai, China · AI Center, Futu AI, Shenzhen, China

Abstract

Natural language prompts and skills serve as the strategic backbone of LLM-based agents. Recent advances in prompt and skill optimization have achieved notable gains, yet all existing methods optimize a single text template---missing the synergy among multiple complementary strategies. We propose MOSCOPT, a text-native, parameter-free algorithm that jointly optimizes a pool of $N$ skills and a gating skill $G$ that dynamically selects $K$ skills per step. To effectively optimize the skills, we build the EditAdam with internally maintained dual states. Through the three-phase interleaved updates with EditAdam, the system monotonically improves without gradient or parameter tuning. Extensive experiments and detailed ablations across 5 benchmarks and 3 target LLMs demonstrate that MOSCOPT consistently outperforms all baselines, and confirm that both the mixture-of-skills architecture with selective activation and the collective evolution with three-phase interleaving are essential to its superior performance. Code is released https://github.com/zhangzhenyu13/SummerClaw/tree/master/summerclaw/agent_trainer/algorithms/moscopt.

Figures & tables

Figure 1: MOSCOPT maintains three core components: a skill pool $\mathcal{S}=\{s_{1},\ldots,s_{N}\}$ , a gating skill $G$ , and a summary table $\Sigma(\mathcal{S})$ . Given a task input, $G$ selects $K$ skills via $\Sigma_{t}(\mathcal{S})$ , outputting ACTIVATE IDs (with top- $K$ -by-Q-score fallback). The $K$ activated skills are loaded into agent context while non-activated skills remain invisible.

Method	SearchQA	GPQA-D	HLE	AIME	LiveMath	avg.
No Skill	48.8	90.7	46.3	94.5	29.9	62.0
5-shot LLM Skill	41.0	88.3	46.5	93.1	29.4	59.7
Reflexion	76.1	90.4	48.7	98.3	37.1	70.1
DSPy	62.2	90.7	46.1	93.8	31.5	64.9
SE-Agent	79.9	89.4	47.0	94.1	46.3	71.3
OPRO	53.2	90.9	46.5	95.7	41.6	65.6

Table 1: Main results with Qwen3.6-Plus and cross models evaluation for MOSCOPT.

Ablation Configuration	Avg. Score
MOSCOPT (full)	74.1
w/o EditAdam (simple reflect edit)	68.5
w/o Gating Optimization ( $G$ fixed)	70.9
Task level Gating	73.1
w/o Collective co-breeding	73.7
w/o Progressive Disclosure	73.9

Table 2: Ablation study on Qwen3.6-Plus. Removing any component degrades performance, confirming the necessity of each design choice.

Deployment Mode	Avg. Score
Full Mixture ( $\{\mathcal{S},G,K\}$ )	74.1
Static Routing Table (top Q scores)	70.6
Distilled Single Skill (merged top skills)	71.2

Table 3: Comparison of three deployment modes on Qwen3.6-Plus. Full mixture with dynamic gating outperforms both static routing and skill distillation.

Appendix figures & tables4 assets

Supplementary material from the paper’s appendix.

Appendix

Mode	Behavior	Use Case
constant	Fixed budget	Uniform task difficulty
linear	Decay from $L_{\max}$ to $L_{\min}$	Skills stabilizing
cosine	Smooth annealing	Avoid abrupt changes
autonomous	LLM decides budget	Fluctuating difficulty

Appendix

Table 5

Rollout	Activated	Hard $R(\tau)$	Co-occur?	Notes
1	$\{2,3\}$	1 (success)	$C(2,3){+}{+}$	Both credited $+1$
2	$\{2,5\}$	0 (failure)	—	Both credited $+0$
3	$\{1,3\}$	1 (success)	$C(1,3){+}{+}$	Both credited $+1$
4	$\{2,3\}$	1 (success)	$C(2,3){+}{+}$	Both credited $+1$

Appendix

Table 6

Hyperparameter	Symbol	Default	Description
Pool size	$N$	10	Total skills in pool
Activation count	$K$	3	Skills activated per step
Elim/reproduce	$M$	1	Skills replaced per cycle
Evolution interval	$E$	1	Epochs between evolution
Gating granularity	—	task	task-level or step-level
Max edit budget	$L_{\max}$	8	Max edits per opt_step

Appendix

Table 7

Property	SkillOpt	MOSCOPT
Optimization target	Single skill	Skill pool + gating
Exploration	Single path	Multi-skill parallel
Long-horizon tasks	Fixed strategy	Phase-aware switching
Synergy exploitation	None	Explicit modeling
Safe editing	Bounded + gate	Inherited + 3-phase
LLM calls/opt_step	$O(B+R)$	$O(BT+N\|V\|+\|V\|)$

Appendix

Table 8

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 402 papers

MOCHA: Multi-Objective Chebyshev Annealing for Agent Skill Optimization
May 19, 2026Md Mehrab Tanjim, Jayakumar Subramanian, Branislav Kveton +6LLM Agent Skill Learning Multi-Objective Optimization
Adobe Research
SkillOpt: Executive Strategy for Self-Evolving Agent Skills
May 22, 2026Yifan Yang, Ziyang Gong, Weiquan Huang +12Large Language Model-Guided Optimization LLM Agent Self-Improvement
Microsoft · Shanghai Jiao Tong University · Tongji University +1
COBRA-Skills: Contextual Bandit-Guided Evolution for Agent Skill Optimization
Sep 10, 2026Pingchen Lu, Xiangyi Wang, Xiang Li +6Evolutionary Optimization LLM Agent Skill Learning
The Chinese University of Hong Kong, Shenzhen · Tianjin University · The Hong Kong University of Science and Technology (Guangzhou) +1

May 19, 2026J/K move · Enter open · S save

MOCHA: Multi-Objective Chebyshev Annealing for Agent Skill Optimization

Md Mehrab Tanjim, Jayakumar Subramanian, Branislav Kveton, Xiang Chen+5

Adobe Research

LLM Agent Skill Learning Multi-Objective Optimization Pareto Optimization

Open paper AskPDF arXiv

LLM agents organize behavior through skills -- structured natural-language specifications that govern how an agent reasons, retrieves, and responds. Unlike monolithic prompts, skills are multi-field artifacts with platform limits: descriptions are indexed for routing, instruction bodies are loaded into a shared context window, and every extra line costs the agent context it could spend elsewhere. Skill optimization is therefore inherently multi-objective: a skill should be accurate and economical at once. Existing prompt optimizers collapse these goals into a single score, and we find that this makes them stall: in most of their runs, 1000 rollouts yield zero progress over the seed skill. We introduce MOCHA (Multi-Objective CHebyshev Annealing), which selects skills to improve by Chebyshev scalarization and accepts improvements that expand the Pareto front, annealing from exploration to exploitation. Chebyshev scalarization scores a skill by its largest weighted shortfall from a perfect skill; with random weights, it can select any skill on the Pareto front for improvement, including trade-offs that an average would never favor. Across six diverse skills and two models, MOCHA escapes the seed in nearly every run, achieving 7.9% higher mean accuracy than the strongest baseline (up to 23.7% on DebugBench with GPT 5.5 and 14.9% on FEVER with Claude Haiku 4.5) while discovering Pareto fronts with the highest hypervolume and nearly twice as many Pareto-optimal skills.

Abstract

LLM Agent Skill Learning

Explore similar work

MOCHA: Multi-Objective Chebyshev Annealing for Agent Skill Optimization

SkillOpt: Executive Strategy for Self-Evolving Agent Skills

COBRA-Skills: Contextual Bandit-Guided Evolution for Agent Skill Optimization