Source-linked AI summary
SkillAdam: Stable and Efficient Skill Evolution for Agents
Gaoyuan Li, Meihao Fan, Yizhe Liu, Shaolei Zhang, Ju Fan, Siyi Wang, Jiaheng Hou, Xudong Weng, Honghan Tian, Zang Li
TL;DR
High-quality skills are costly to author and heuristic self-evolution can revise them unstably. SkillAdam adapts Adam’s stabilization and step-size principles to discrete skill documents, achieving stronger benchmark performance with fewer iterations and lower cost.
Problem
Automatically constructing high-quality skills remains difficult because manual authoring is labor-intensive and heuristic refinement can depend on task-specific strategies.
Method
SkillAdam combines optimization memory for preserving prior corrections with a volatility-driven edit budget for adapting revision magnitude.
Results
Across seven benchmarks, SkillAdam achieves state-of-the-art performance at substantially lower optimization cost, including gains on both short- and long-horizon tasks.
Takeaways & Limitations
SkillAdam provides a more stable and efficient framework for iterative skill self-evolution across short- and long-horizon agentic tasks.
Takeaways & Limitations
The judge experiment evaluates the complete framework and aggregates six dimensions, so it isolates neither individual components nor single quality dimensions.
Abstract
from arXiv · showhide
Agent skills provide a lightweight way to equip frozen language-model agents with domain knowledge and procedural guidance, yet obtaining high-quality skills remains costly and difficult to scale. Expert-written skills require substantial human effort. Recent skill self-evolution methods automate an iterative loop that uses execution feedback to revise skills, but their heuristic update strategies often yield unstable optimization and low iteration efficiency. We identify two challenges in realizing stable and efficient skill self-evolution. Direction Stability requires effective corrections to accumulate rather than be overwritten by iteration-local feedback. Update Adaptivity requires the scope of each revision to reflect the consistency of recent case-level improvements. We introduce SkillAdam, an Adam-inspired framework for optimizing discrete and non-differentiable skill documents. As a functional analogue of Adam's first moment, an optimization memory records identified problems and the outcomes of prior solution attempts to stabilize the update direction. As a functional analogue of Adam's second moment, a volatility-driven edit budget tracks the history-weighted variation of recent case-level improvements and adaptively controls the update magnitude. Across seven benchmarks that span short- and long-horizon tasks, SkillAdam achieves state-of-the-art performance with more stable optimization dynamics. It also obtains stronger skills with substantially fewer optimization iterations and lower cost than prior methods. Code repository: https://github.com/ruc-datalab/SkillAdam
1 INTRODUCTION
High-quality skills are costly to create, while heuristic self-evolution can produce unstable revisions. SKILLADAM addresses this by accumulating effective corrections and adapting revision scope to recent evidence.
- Motivation: High-quality skills improve agent performance but remain difficult to obtain automatically and costly to author manually.LLM-generated skills often require refinement, while expert authoring is time-consuming and labor-intensive.
- Motivation: Prompt- and heuristic-based skill refinement reduces manual effort but depends on human-designed, task-specific strategies.
- Related approaches: Recent methods automate skill construction from interaction trajectories, execution traces, and iterative self-evolution.
- Challenges: Direction Stability requires preventing later feedback from overwriting effective earlier corrections, while Update Adaptivity requires matching revision scope to evidence consistency.The paper motivates these challenges with revisions that can trade off itinerary cost against plan richness.
- SKILLADAM: SKILLADAM uses an issue tracker as optimization memory and a volatility-driven edit budget as functional analogues of Adam’s first- and second-moment mechanisms.The tracker preserves prior issues and solution outcomes, while volatility controls allowable modification scope.
- Contributions: Across seven benchmarks, SKILLADAM achieves state-of-the-art performance with substantially fewer optimization iterations and lower overall cost.
2 RELATED WORK
Prior work externalizes reusable capabilities, optimizes discrete language artifacts, and increasingly automates skill construction and revision. These efforts motivate SKILLADAM’s focus on natural-language skill documents and iterative self-evolution.
- Agent Skills: Agent Skills are modular instruction packages and supporting resources that equip agents with specialized capabilities without retraining compatible models.
- Agent Skills: Earlier systems externalize reusable capabilities through executable program libraries, structured rules, verified web skills, and retained experience or workflows.
- Prompt Optimization: Prompt optimization improves discrete language artifacts using evaluated candidates, textual feedback, reflective rollouts, or retained earlier feedback.
- Skill Self-Evolution: Recent skill self-evolution methods diagnose, consolidate, revise, or jointly evolve skills using execution traces, failure analysis, or surrogate verification.
3 PRELIMINARIES
The paper models skills as discrete natural-language documents evaluated on task batches through numerical metrics and diagnostic feedback. It then uses Adam as a conceptual basis for stabilizing directions and adapting effective step sizes.
- Skill optimization problem: A skill is a Markdown-formatted natural-language instruction document guiding an agent’s behavior in a target domain.
- Skill optimization problem: For each skill-task pair, the evaluator returns a metric vector and optional diagnostic information such as error descriptions or judge rationales.
- Skill optimization problem: Numerical metrics support aggregation and acceptance decisions, while diagnostic content remains language-space evidence for patch generation and issue tracking.
- Skill optimization problem: Evaluation feedback is collected over task batches, and a task-dependent aggregation function maps task-level metrics to a scalar objective.
- Adam optimization: Adam maintains exponential moving averages of stochastic-gradient moments to stabilize update direction and adapt the effective step size.The first moment aggregates gradient information, while the second moment rescales the base learning rate using recent squared-gradient magnitude.
4 METHOD
SkillAdam performs discrete skill self-evolution with two Adam-inspired mechanisms: an issue tracker stabilizes update direction, while a volatility-driven budget adapts modification scope. Each iteration compares current and candidate skills on the same cases, updates persistent states, and conditionally accepts the candidate.
- Adam-inspired design: SkillAdam’s Adam analogy is functional: language-space issue histories and patches serve roles analogous to first-moment memory and gradient-like updates without computing gradients.Evaluation feedback plays a loss-like role, while volatility determines the next edit budget.
- Skill-evolution cycle: At each iteration, the agent rolls out the current skill on a sampled mini-batch, generates a candidate patch, and evaluates both skills on the same cases.Paired trajectories and feedback provide directly comparable case-level evidence for the update and volatility calculations.
- Optimization memory: The Evolving Issue Tracker records error patterns, statuses, prior solution attempts, and their outcomes to stabilize updates across iterations.It integrates current evidence with accumulated issue histories rather than using only iteration-local feedback.
- Adaptive edit budget: Volatility measures how unevenly a candidate changes case-level performance, with high volatility indicating inconsistent effects across cases.When fewer than two valid case-level comparisons are available, volatility is set to zero.
- Adaptive edit budget: High volatility yields a smaller next-iteration edit budget, while low volatility permits a broader modification.The budget is determined from the state accumulated through the previous iteration and is bounded by base, minimum, and saturation parameters.
- Skill-evolution cycle: A benchmark-specific acceptance gate decides whether the candidate replaces the current skill, after which the tracker and volatility state are carried into the next iteration.The resulting accepted skill and optimizer states close the optimization loop.
5 EXPERIMENTS
The experiments evaluate SkillAdam across seven benchmarks spanning knowledge work and interactive planning, using native environments, interfaces, and evaluators. The benchmark suite contains ten evaluation slices, with DeepPlanning contributing four slices.
- Benchmark suite: Seven benchmarks cover knowledge work and interactive planning, producing ten evaluation slices in total.Six benchmarks contribute one slice each, while DeepPlanning contributes Shopping Levels 1–3 and Travel EN.
- Optimization procedure: Algorithm 1 initializes a skill from trajectory and feedback data, then iterates over sampled batches while maintaining tracker, volatility, and edit-budget states.The procedure returns the skill produced at the maximum iteration.
- Benchmark suite: Five benchmarks are classified as short-horizon because they require fewer than ten tool calls per task on average.They include SearchQA, SpreadsheetBench, OfficeQA, DocVQA, and LiveMathematicianBench.
- Evaluation metrics: Primary metrics vary by benchmark: Exact Match, hard task success, ANLS-hard, episode goal-completion rate, or domain-level case accuracy.DP-Avg averages the unrounded DP-Shopping and DP-Travel accuracies rather than the four individual DeepPlanning slices.
5.2 BASELINES
SkillAdam is compared with seven baselines spanning no-skill, human-authored, one-shot LLM-authored, and automated skill-optimization approaches. Comparisons use matched task settings, with protocol differences retained where baselines require native procedures.
- Compared methods: The seven baselines include NoSkill, HumanSkill, LLMSkill, Trace2Skill, TextGrad, GEPA, and SkillOpt.They cover both authored skills and automated optimization methods.
- Comparison protocol: Within each benchmark, methods use the same target-agent configuration, test cases, and evaluator.SkillAdam and SkillOpt also share the same initial skill constructed from fixed baseline trajectories.
- Comparison protocol: SkillOpt retains its original train and selection split, native selection and slow-update rules, and optimizer-memory procedures, and is run for four epochs.This preserves SkillOpt’s reported protocol for comparison.
5.3 SETUP
Experiments use frozen target models in a no-harness direct-chat setting, with skills injected as natural-language instructions and benchmark-specific optimization protocols. Acceptance depends on improvement thresholds for designated metrics while protecting others from regression.
- Common protocol: All experiments use a no-harness direct-chat setting with native task interfaces and evaluators, natural-language skill instructions, and frozen target models.GPT-5.5 is used outside DeepPlanning, while DeepPlanning uses a different target model configuration.
- Initialization and optimization: SkillAdam shares an initial skill with SkillOpt, uses unchanged test partitions for final evaluation, and applies benchmark-specific optimization splits.Outside DeepPlanning, SkillAdam merges train and selection partitions; DeepPlanning uses odd-numbered cases for optimization and even-numbered cases for evaluation.
- Initialization and optimization: At each iteration, the current skill and proposed revision are evaluated on the same sampled cases before optimizer states are updated.Their trajectories and evaluation feedback are passed to the optimizer states.
- Acceptance protocol: A candidate is accepted when at least one designated metric reaches its improvement threshold and every protected metric stays within its regression boundary.The gate does not use an additional validation set, and an auxiliary metric can support acceptance when the primary metric is unchanged.
- Main results: SkillAdam wins four of five short-horizon benchmarks and ties on one, averaging 14.45% and 31.20% improvements over HumanSkill and LLMSkill.Against SkillOpt, it improves DocVQA by 1.21% and LiveMath by 1.20%.
- Main results: SkillAdam raises DP-Avg from 21.7% to 28.3% versus SkillOpt, a gain of 6.7 percentage points.On DP-Travel, where NoSkill and SkillOpt nearly fail, SkillAdam reaches 11.7% accuracy.
- Main results: Overall, SkillAdam achieves the best performance among the compared baselines.The paper presents this as evidence for the effectiveness of its skill-evolution strategy.
5.5 ABLATION STUDIES
The ablations show that both Optimization Memory and the Volatility-driven Edit Budget improve long-horizon planning, while transfer tests show that SKILLADAM’s skills retain stronger performance across agent backbones.
- Ablation Results: 28.3% DP-Avg is achieved by full SKILLADAM, up from 19.2% without both mechanisms.Adding the Volatility-driven Edit Budget with Optimization Memory fixed improves DP-Avg by 6.7 percentage points.
- Ablation Results: The edit budget supports update adaptivity by shrinking revisions after inconsistent case-level effects and permitting broader updates after consistent improvements.This can avoid damaging previously correct constraints during long-horizon planning.
- Ablation Results: Optimization Memory raises DP-Avg from 19.2% to 21.7% without the edit budget, with gains on Shopping Levels 2 and 3 but a decrease on Level 1.The cumulative ablation does not independently isolate memory when the edit budget is enabled or their interaction.
- Cross-Model Transfer: SKILLADAM improves average target-model score from 63.1% to 67.8% and retention ratio from 76.2% to 81.3% over SkillOpt.It scores higher on five of six benchmarks and has higher retention on four of six.
- Cross-Model Transfer: The transfer experiment evaluates the complete framework, so it does not isolate which component produces the cross-backbone gain.Skills generated with GPT-5.5 were deployed verbatim on GPT-5.4-mini without further optimization or adaptation.
5.7 SKILL QUALITY ANALYSIS
SKILLADAM produces higher judged skill quality than SkillOpt across all ten benchmark slices and reduces optimization cost while improving DeepPlanning performance. The quality evaluation complements task-performance results but does not isolate individual components or dimensions.
- Skill Quality: Every paired quality comparison favors SKILLADAM, indicating that the improvement is consistent across tasks.The evaluation uses a blinded input that does not reveal the method name.
- Skill Quality: SKILLADAM’s mean skill-quality score rises from 3.30 to 3.88 and exceeds SkillOpt on all ten benchmark slices.Scores average six dimension medians from three independent calls to the same frozen GPT-5.5 judge.
- Interpretation: The judge experiment aggregates six quality dimensions and evaluates the complete framework, so it isolates neither an individual component nor a single dimension.The benchmark metrics measure task completion, whereas the judge evaluates relevant, operational, and usable guidance.
- Optimization Cost: SKILLADAM reduces total token consumption by 67.3% and API requests by 68.8% versus SkillOpt while increasing DP-Avg from 21.7% to 28.3%.The lower optimization cost is not achieved by sacrificing final-skill quality.
- Optimization Cost: The cost reduction comes from fewer optimization requests and modification attempts rather than shorter individual calls.The two methods consume a similar number of tokens per request, with SKILLADAM slightly higher on this measure.
6 ANALYSES
On DeepPlanning Shopping Level 1, SKILLADAM produced a more stable accepted optimization path and reached a stronger retained skill with substantially lower token cost than SkillOpt in the reported run.
- 6.1 Stability: SKILLADAM found a strong revision in the first iteration, and every subsequent accepted skill remained above the 40% initial test accuracy.SkillOpt’s accepted regular updates fluctuated around or below its starting performance.
- 6.1 Stability: Optimization Memory helps preserve useful corrections by retaining identified issues, their status, and prior solution outcomes across iterations.The acceptance gate also prevents insufficiently supported revisions from replacing the current skill.
- 6.1 Stability: The stability result characterizes the combined behavior of memory and the acceptance gate in one run, not the isolated causal effect of memory.A rejected revision had a high post-hoc test score because acceptance used sampled optimization cases and protected metrics instead.
- 6.2 Efficiency: SKILLADAM reached and retained a 56% test-accuracy skill after approximately 4M tokens, while SkillOpt produced a comparably strong candidate only after approximately 20M tokens and did not accept it.In this run, SKILLADAM therefore used about one fifth of the token cost.
- 6.2 Efficiency: The efficiency gain came from fewer unproductive modification attempts rather than cheaper individual requests.The comparison controlled for similar token cost per request in Table 6.
7 CONCLUSION
SKILLADAM adapts Adam’s principles to discrete skill self-evolution through optimization memory and a volatility-driven edit budget. Across seven short- and long-horizon benchmarks, it achieved state-of-the-art performance at substantially lower training cost, with better cross-model transfer.
- 7 CONCLUSION: SKILLADAM adapts Adam’s two principles to the skill space through accumulated optimization memory and a volatility-driven edit budget.Memory provides stability, while the edit budget adapts update scope to improvement-evidence reliability.
- 7 CONCLUSION: Across seven benchmarks spanning short-horizon and long-horizon agentic tasks, SKILLADAM achieved state-of-the-art performance at substantially lower training cost.
- 7 CONCLUSION: SKILLADAM’s skills transferred better across models.