PaperlayerView exact revision on arXiv

Source-linked AI summary

Skill is Not One-Size-Fits-All: Model-Aware Skill Alignment for LLM Agents

Jianxiang Yu, Jiapeng Zhu, Bochen Lin, Qier Cui, Zichen Ding, Xiang Li

arXiv:2605.30723v1cs.CL

TL;DR

Existing skill libraries reuse the same formulations across different backbones despite evidence that skills can help one model and harm another. MASA aligns skills to each target backbone through hierarchical evolution and a lightweight rewriter, achieving the best success rate across settings with gains up to +25.8 points.

  • Problem

    Existing skill libraries reuse shared skill formulations across backbones, despite model-dependent effectiveness and differing preferences for skill granularity and expression.

  • Method

    MASA uses teacher-guided hierarchical skill evolution with hill climbing and UCB-driven search, then trains a lightweight model-conditioned rewriter for single-pass adaptation.

  • Results

    +25.8 points was the maximum gain over the strongest baseline, with MASA achieving the best success rate across three environments and four Qwen3 backbones.

  • Takeaways & Limitations

    The results support treating skills as model-aware artifacts adapted to target backbones rather than shared uniformly across models.

  • Takeaways & Limitations

    Empirical evidence is restricted to the Qwen3 family, so broader generality across model families and environments remains untested.

Abstract

from arXiv · show

LLM agents increasingly retrieve externally curated skills-procedural instructions retrieved at decision time-to improve performance on long-horizon interactive tasks. Existing skill libraries are typically treated as model-agnostic, reusing the same skill formulations across backbones with substantially different capacities and behaviors. However, our controlled experiments across multiple model scales show that skill effectiveness is strongly model-dependent: a skill that benefits one backbone can harm another. Motivated by this observation, we propose MASA Model-Aware Skill Alignment, a framework that adapts skills to each target backbone without modifying agent weights. MASA operates in two stages: (1) a hierarchical skill evolution pipeline that iteratively rewrites general and task-specific skills using hill climbing and UCB-driven tree search, guided by environment feedback and model capability profiles; and (2) a lightweight model-conditioned skill rewriter trained on evolution trajectories to reproduce the adaptation in a single forward pass. Experiments across three interactive environments and four backbones show that MASA consistently achieves the best overall performance, with gains of up to 25.8 points over the strongest baseline. The learned rewriter further generalizes to unseen tasks and environments without additional search, consistently outperforming a much larger teacher LLM at a fraction of the inference cost.

1 Introduction

Existing skill libraries reuse shared procedural formulations across LLM backbones, but experiments show that skill effectiveness is model-dependent and can vary across model families and sizes. MASA addresses this mismatch by aligning skills with each target backbone through hierarchical evolution and lightweight model-conditioned rewriting without modifying agent weights.

  • Motivation: LLM agents use externally retrieved procedural knowledge, called skills, to steer long-horizon interactive tasks without modifying model weights.The tasks include web navigation, embodied control, and tool use.
  • Problem: Existing skill-library systems typically build one shared library and reuse it across different LLM backbones.These libraries may be hand-crafted or distilled from agent trajectories.
  • Empirical motivation: The optimal skill granularity varies across four Qwen3 backbones spanning 4B–32B, and a skill can boost one backbone while degrading another.The experiment fixes the skill-library principles and varies only granularity on ALFWorld.
  • Empirical motivation: The same model-dependent preference pattern holds across the Gemma3 family and across models of the same size from different families.These findings indicate that skill effectiveness depends on formulation as well as encoded knowledge.
  • MASA: MASA aligns each skill library with its target backbone without modifying agent weights by treating skill alignment as a hierarchical search driven by environment feedback.It combines capability profiles with iterative skill reformulation.
  • MASA: MASA evolves general skills through hill climbing and task-specific skills through UCB-driven tree search, then trains a lightweight rewriter on evolution trajectories.The rewriter is designed to transform unaligned skills into model-appropriate formulations in a single forward pass.

2 Preliminary Study: One Skill Library Does Not Fit All

The preliminary study tests whether one skill library serves all model scales equally by holding behavioral principles fixed while varying only textual granularity. Results show that optimal skill form is model-dependent, and mismatches can reduce performance.

  • Experimental question: The study isolates skill-form effects by keeping underlying behavioral principles fixed and varying only their textual granularity.The evaluation compares skill formulations rather than different skill content.
  • Experimental setup: The three skill variants encode identical behavioral principles while differing in representational depth: Concise, Moderate, and Detailed.Moderate follows Xia et al. (2026), while Concise and Detailed are controlled rewrites preserving the same principles.
  • Finding 1: 32.1% is Qwen3-8B’s best result under No Skill, while all three skill variants reduce its performance.Qwen3-4B performs best with Moderate skills, whereas Qwen3-14B and Qwen3-32B perform best with Detailed skills.
  • Finding 1: No single granularity level is uniformly optimal across models, demonstrating that skill-form mismatches can hurt performance.The finding does not imply that skills are inherently incompatible with Qwen3-8B.

Hierarchical Model-Conditioned Skill Evolution

MASA evolves general and task-specific skills through model-conditioned, search-based adaptation, then distills this process into a model-conditioned skill rewriter. Its design addresses non-monotonic model-scale effects, sharp task-type variation, and the need for a model-aware skill library.

  • Framework: MASA separates adaptation into Stage 1 general skill evolution and Stage 2 task-specific skill evolution.The framework also includes a model-conditioned skill rewriter and a model-aware skill library.
  • Design requirements: Skill granularity has a non-monotonic, model-specific relationship with performance, ruling out simple alignment heuristics.The findings note that neither concise nor more explicit procedural guidance consistently benefits smaller or larger models.
  • Design requirements: Performance can vary by over 60 points across task types within a single model–granularity pairing, requiring task-type-specific adaptation.This spread exceeds the differences between granularity levels for any single task, making global optimization insufficient.
  • Design requirements: The same scale-dependent trend appears across the Gemma3 4B/12B/27B family, suggesting that the phenomenon generalizes across model families.The parallel experiment is reported in Appendix C.3.
  • Design requirements: MASA performs open-ended, model-aware rewriting to address misalignment in decision strategy, framing, and format beyond textual granularity.The controlled study varies only textual granularity while holding content fixed, but the paper identifies broader sources of misalignment in practice.

3 Method: MASA

MASA aligns skills to each target backbone through model-conditioned evolution and a deployment-time rewriter. Its search pipeline optimizes general and task-specific skills separately, while the rewriter amortizes this adaptation into one forward pass.

  • Framework: MASA conditions skill libraries on a structured model card rather than using model-agnostic formulations.The model card includes architecture metadata, training provenance, and backbone strengths and weaknesses.
  • Framework: The frozen backbone retrieves separate top-kG general and top-kT task-specific skills, with only the skill library optimized.General skills encode cross-task strategy principles, while task-specific skills provide procedures tailored to task types.
  • Objective: MASA maximizes expected adjusted reward by balancing task success against nothing-happens-rate penalties that proxy skill-induced stalling.The penalty strength is controlled by λ ∈[0, 1].
  • Skill evolution: The evolution pipeline uses teacher-driven failure attribution and rewrites, optimizing general skills by hill climbing and task-specific skills by per-type UCB1 tree search.General-skill candidates require full-suite evaluation, whereas task-specific search explores diverse strategies independently for each task type.
  • Deployment-time rewriter: MASA-Rewriter learns from evolved skill sets to adapt new skills in a single forward pass without environment interaction or iterative search.The evolution pipeline supplies training targets, while the rewriter is intended for compute-constrained deployment.

4 Experiments

Across three interactive environments and four Qwen3 backbones, MASA consistently outperforms model-agnostic and one-shot rewriting baselines. Its evolved skills and lightweight rewriter also generalize across datasets, task types, and environments without additional search or target-environment data.

  • Experimental setup: Experiments span ALFWorld, WebShop, and search-augmented QA across Qwen3-4B, 8B, 14B, and 32B backbones.The QA evaluation covers seven single-hop and multi-hop benchmarks; all Qwen3 models use non-thinking mode.
  • Search-augmented QA: MASA achieves the highest average search-augmented QA success rate for every backbone, improving 4B Bamboogle from 12.9 for the best baseline to 61.3 and ranking first on 5 of 7 datasets for 32B.Skill evolution uses only NQ and HotpotQA, yet gains generalize to out-of-domain benchmarks.
  • ALFWorld: MASA achieves ALFWorld average success rates of 31.4, 57.9, 64.3, and 65.7 for 4B, 8B, 14B, and 32B, gaining +4.3, +25.8, +20.0, and +20.7 over the strongest baselines.For Qwen3-14B and 32B, MASA ranks first on all six ALFWorld task types.
  • WebShop: MASA achieves the highest WebShop success rate and score for every backbone, while larger models underperform 4B without adaptation, including 14B No Skill at 2.8% versus 4B No Skill at 23.0%.The passage attributes the larger models’ weakness to verbose chain-of-thought preambles before actions.
  • Generalization: The 4B-parameter MASA-Rewriter consistently surpasses DeepSeek-V4-powered DS-Adapter at a fraction of the inference cost.Held-out ALFWorld task types are CLEAN, HEAT, and COOL, with general skills unchanged.
  • Generalization: Cross-env MASA-Rewriter, trained only on Search and WebShop traces, outperforms DS-Adapter on held-out ALFWorld by +1.5, +3.0, +2.9, and +3.0 for 4B, 8B, 14B, and 32B.Cross-task training increases gains over DS-Adapter to +8.8, +13.2, and +7.4 for 8B, 14B, and 32B.

5 Conclusion

MASA addresses the breakdown of one-size-fits-all skill design by aligning skills with target backbones through hierarchical evolution and a lightweight model-conditioned rewriter. Across three environments and four Qwen3 backbones, it achieved the best success rate in all settings, while the rewriter generalized to unseen tasks and environments at negligible deployment cost.

  • Conclusion: MASA achieved the best success rate in all settings across three environments and four Qwen3 backbones.Its gains reached +25.8 points.
  • Conclusion: +25.8 points was the maximum reported gain for MASA across the evaluated settings.The framework combined hierarchical skill evolution with a lightweight model-conditioned rewriter that amortizes search into one forward pass.
  • Conclusion: MASA supports treating skills as model-aware artifacts adapted to target backbones rather than shared uniformly across models of different capacities.The authors envision MASA-Rewriter as lightweight plug-and-play middleware for automatically rewriting existing skill libraries.

Limitations

The empirical evidence and skill rewriter training are limited to the Qwen3 family and three environments. The evolution pipeline also depends on automatic success/failure signals, restricting use in domains that require external evaluators or human annotations.

  • Model and environment coverage: Empirical evidence is restricted to Qwen3 models (4B/8B/14B/32B), while broader model-family and environment coverage would require substantially more compute.The paper identifies open-weight and proprietary model families as potential extensions.
  • Training coverage: The skill rewriter is trained on skill-evolution trajectories collected from ALFWorld, WebShop, and Search-QA.
  • Evaluation requirements: The evolution pipeline relies on environments providing automatic success/failure signals, such as task-completion flags, to evaluate rewritten skills.
  • Evaluation requirements: Applying the framework to open-ended web tasks or real-world applications would require external evaluators or human annotations.Such evaluators could broaden the framework’s range of agent applications.

Ethical Considerations · A Related Work

MASA uses public benchmarks and licensed models without collecting new human-subject data, but its increased agent capability and evolved skills require oversight and validation. Related work spans reusable skill libraries and model-aware prompt optimization, while MASA targets dynamically retrieved skills rather than monolithic prompts.

  • Ethical Considerations: MASA introduces no new human-subject data collection and evaluates on standard public benchmarks, including ALFWorld, WebShop, and open-domain QA datasets.
  • Ethical Considerations: The experiments use publicly released models accessed according to their respective licenses.
  • Ethical Considerations: Improving agent effectiveness through skill adaptation may increase capabilities in interactive environments, requiring monitoring, policy constraints, and human oversight in high-risk settings.
  • Ethical Considerations: The skill evolution pipeline may inherit biases or unsafe heuristics from optimization trajectories and feedback, so evolved skill libraries require inspection and validation before deployment.
  • A Related Work: Reusable procedural knowledge improves LLM-agent performance without modifying model weights; prior systems use textual feedback, growing libraries, or cached successful behaviors.
  • A Related Work: Because LLM behavior is sensitive to instruction phrasing, OPRO and EvoPrompt iteratively refine task instructions, but their outputs remain independent of backbone identity.
  • A Related Work: MAPO and PromptBridge account for model identity when optimizing or transferring task instructions, whereas MASA addresses dynamically retrieved skill libraries rather than single monolithic prompts in non-agent settings.

B Ablations · C Preliminary Study: Supplementary Details · C.1 Skill Variant Comparison

The ablations show that MASA’s two search stages contribute differently across backbones and environments, while model-card conditioning consistently improves the rewriter. The preliminary study compares skill variants that preserve task coverage but expose different amounts of procedural guidance.

  • B Ablations: Removing task-specific search causes the largest ALFWorld drops for Qwen3-8B (−25.0) and Qwen3-32B (−15.7).These results indicate that per-task-type procedural guidance is critical for these backbones.
  • B Ablations: Removing general search most severely affects Qwen3-14B (−16.4), indicating that high-level behavioral priors matter for this backbone.The passage attributes this to sufficient capacity to follow such priors while still benefiting from stronger guidance.
  • B Ablations: WebShop depends more on consistent high-level decision strategies encoded by general skills, whereas ALFWorld depends on fine-grained procedural sequences encoded by task-specific skills.This asymmetry explains why the two search-stage ablations affect environments and backbones differently.
  • B Ablations: Removing the target model’s capability card consistently degrades MASA-Rewriter performance on held-out ALFWorld tasks.The comparison uses Cross-env training on Search + WebShop and Cross-task training on Search + WebShop + ALFWorld Pick/Look/Pick2; results are average SR on Clean/Heat/Cool.
  • B Ablations: The model-card effect is especially pronounced for the Cross-task training variant.The ablation evaluates the rewriter with and without the target model’s capability card.
  • C.1 Skill Variant Comparison: Table 4 presents three nonempty ALFWorld skill variants for the preliminary study.All variants retain the same skill IDs and task coverage.
  • C.1 Skill Variant Comparison: The variants differ in how much procedural text is exposed to the agent, with more detailed versions adding trigger conditions, executable steps, and failure-prevention cues.Bold text highlights these additions in the comparison.

C.2 Per-Task Breakdown: Qwen3 … G WebShop Supplementary Results

The supplementary results show that skill effectiveness varies substantially by task, model family, scale, and architecture. The appendices detail model-card construction, two-stage skill evolution, tree-search settings, and rewriter-training data.

  • C.2 Per-Task Breakdown: Qwen3: Qwen3 per-task success rates show large within-condition swings, motivating task-specific skill search.Qwen3-14B Concise scores 74.2 on PICK versus 13.7 on COOL, while Qwen3-4B Detailed scores 1.6 on PICK versus 46.7 on LOOK.
  • C.3 Supplementary Validation: Gemma3: Gemma3-4B and Gemma3-12B perform best with Concise skills, whereas Gemma3-27B performs best with Detailed skills.The fixed-granularity sweep supports the conclusion that the best skill form is model-dependent rather than universally transferable.
  • C.3 Supplementary Validation: Gemma3: Equal-sized Gemma3-4B and Qwen3-4B backbones prefer different skill granularities, isolating architecture and training effects from scale.Gemma3-4B peaks with Concise skills, whereas Qwen3-4B peaks with Moderate skills despite identical parameter budgets.
  • D Model Card Construction: Model cards combine architecture metadata, training provenance, and capability profiles from documentation and automated rollout analysis.The rubric includes parameters such as model family, context window, alignment pipeline, training-data scale, multilingual support, strengths, and observed weaknesses.
  • E.1 Stage 1: Hill Climbing: Stage 1 uses hill climbing with up to 10 iterations, patience 3, and the top 5 historical skill sets supplied to the teacher.A candidate general skill set is accepted only when its average adjusted reward strictly exceeds the current best.
  • E.2 Stage 2: UCB-Driven Tree Search: Stage 2 performs per-task UCB1 tree search, selecting nodes by mean adjusted reward, visit counts, and an exploration constant of 1.4.The procedure runs 10 iterations per task type with 100 episodes per node evaluation.
  • F Skill Rewriter Training Details: The Qwen3-4B rewriter is trained with augmented in-domain and cross-model pairs, including combined and environment-specific variants.The combined rewriter uses 769 samples across three environments, while the environment-specific rewriter uses 499 WebShop and Search samples.

G.1 Trajectory Analysis: Why Larger Models Fail · G.2 Per-Category Breakdown

Trajectory analysis attributes larger Qwen3 models’ baseline failures to verbose reasoning that consumes the fixed step budget instead of enabling enough environment interaction. Across categories, MASA broadly improves success rates, especially for OTHER and ELECTRONICS, rather than relying on one easy category.

  • G.1 Trajectory Analysis: Why Larger Models Fail: The analysis examines failed WebShop trajectories to explain why larger Qwen3 models underperform 4B under baseline conditions.The comparison covers Qwen3-8B, 14B, and 32B against 4B.
  • G.1 Trajectory Analysis: Why Larger Models Fail: 0% of Qwen3-4B steps contain chain-of-thought, compared with extensive reasoning preambles in Qwen3-8B/14B/32B actions.Qwen3-4B averages ∼73 chars per action, while larger models prepend verbose reasoning before commands.
  • G.1 Trajectory Analysis: Why Larger Models Fail: 97% of Qwen3-14B steps contain verbose reasoning, making it the most severe case among the larger models.The statistic comes from Table 7’s trajectory analysis, where CoT denotes the fraction of steps containing reasoning preambles.
  • G.1 Trajectory Analysis: Why Larger Models Fail: The verbose-reasoning pattern exhausts the fixed step budget on deliberation rather than environment interaction, preventing enough purchase actions for success.The agents consider multiple options but do not complete sufficient purchase actions.
  • G.2 Per-Category Breakdown: Most categories have near-zero SR for 8B/14B/32B baselines, consistent with the verbose-reasoning bottleneck.This pattern appears in the full per-category success-rate breakdown.
  • G.2 Per-Category Breakdown: MASA achieves the best SR in the vast majority of categories across all backbones, with particularly large gains on OTHER and ELECTRONICS.The improvement is broad rather than category-specific and does not rely on a single easy category inflating the average.

H Skill Rewriter OOD: Per-Task Breakdown · I Qualitative Analysis: Evolved Skill Examples · J The Use of Large Language Models (LLMs)

The OOD breakdown shows that transfer benefits depend on task and backbone, while qualitative examples reveal fundamentally different evolved strategies for each model’s dominant failure mode. Supplementary materials document skill variants, per-task results, and the limited use of LLMs for presentation refinement.

  • H Skill Rewriter OOD: Per-Task Breakdown: Figure 5 reports per-task success-rate breakdowns for the OOD generalization experiment across held-out task types and target backbones.The figure covers Clean, Heat, and Cool for 4B–32B backbones.
  • H Skill Rewriter OOD: Per-Task Breakdown: Cross-task transfer improves over Cross-env on all three ALFWorld tasks, with especially large gains on COOL and CLEAN.Examples include 14B COOL rising from 24.0 to 44.0 and 8B CLEAN rising from 29.6 to 51.9.
  • H Skill Rewriter OOD: Per-Task Breakdown: Cross-environment transfer is strongest on HEAT, reaching 56.2% for both 14B and 32B, but is mixed on CLEAN and COOL.On some larger-backbone cells, Cross-env falls short of DS-Adapter; 14B CLEAN is tied at 40.7.
  • I Qualitative Analysis: Evolved Skill Examples: On WebShop’s color-matching subtask, MASA evolves qualitatively distinct skills tailored to each backbone’s dominant failure mode.The case study identifies color matching as the highest-failure-rate subtask during skill evolution.
  • I Qualitative Analysis: Evolved Skill Examples: The evolved strategies range from exact-match rejection for Qwen3-4B to approximate purchasing for Qwen3-8B and algorithmic substring matching with tiebreaking for Qwen3-32B.Qwen3-14B instead receives a one-step fast-fail exit strategy, reflecting its distinct failure pattern.
  • J The Use of Large Language Models (LLMs): The paper states that LLMs were used exclusively for grammatical polishing and stylistic refinement to improve presentation clarity and readability.The supplementary pages include additional tables and figures, including skill-library variants, per-task results, OOD results, and evolved WebShop skills.
  • J The Use of Large Language Models (LLMs): Supplementary examples include systematic-exploration skills and multiple granularity conditions—CONCISE, MODERATE, and DETAILED—with NO SKILL omitted for brevity.The supplementary tables report ALFWorld and WebShop performance across the listed conditions and backbones.
Loading 2605.30723v1…