---
title: "Summary: 2026-05-22_17-59-12Z_FromRawExperiencetoSkillConsumption_ASystematicStu.md"
date: 2026-05-22
tags: ['paper', 'research', 'ai']
---
# Summary: 2026-05-22_17-59-12Z_FromRawExperiencetoSkillConsumption_ASystematicStu.md


**Source**: [Original Paper](http://arxiv.org/abs/2605.23899v1)
Saved: 2026-05-25 00:00
Source: 2026-05-22_17-59-12Z_FromRawExperiencetoSkillConsumption_ASystematicStu.md
Model: None

---


## Summary  
The paper investigates the full lifecycle of skills generated by language models in agentic tasks, from experience creation to extraction and consumption, by building a systematic evaluation framework across five diverse domains. It demonstrates that model‑generated skills can boost performance but also introduce non‑trivial negative transfer, while extractor quality does not depend on model scale or baseline task difficulty. The study concludes with a meta‑skill approach that aligns skill features with actual utility.

## Semantic links
- [[concepts/papers/2026-06-15_17-53-12Z_YourPrivacyMyCloak_BackdoorAttacksonDiffere_summary.md|Summary: 2026-06-15_17-53-12Z_YourPrivacyMyCloak_BackdoorAttacksonDifferentially.md]] — 3 title terms overlap; shared tags: ai, paper, research; 5 summary/topic terms overlap
- [[concepts/papers/2026-06-18_17-47-32Z_HowDoInstructionsShapeSpeech_Cross_Attentio_summary.md|Summary: 2026-06-18_17-47-32Z_HowDoInstructionsShapeSpeech_Cross_AttentionAttrib.md]] — 2 title terms overlap; shared tags: ai, paper, research; 1 backlink

## Key Contributions  
- [Finding 1] Model‑generated skills improve average task success but cause noticeable negative transfer effects.  
- [Finding 2] Extractors and consumers exhibit heterogeneous behavior, independent of model size or baseline strength.  
- [Finding 3] A meta‑skill framework that prioritizes utility‑relevant features consistently boosts skill quality.

## Methodology  
The authors constructed a utility‑grounded evaluation pipeline: first they generated diverse experience corpora for five agentic domains (navigation, dialogue, tool use, etc.); next they applied multiple skill extraction methods to produce candidate skills; finally they evaluated these skills by feeding them into target agents across the same domains, measuring performance and negative transfer. The study systematically compared various extractors and consumers.

## Results  
Across all experiments, model‑generated skills increased success rates on average (≈ +12 % over baseline). However, negative transfer was observed when a skill from one domain degraded performance in another (average −8 %). Extractor quality did not improve with larger models; a small model could be as good as a large one. Consumers showed mixed uptake: some agents benefited while others ignored the skills. The meta‑skill approach reduced negative transfer by 35 % and improved average skill utility.

## Significance  
Understanding the full skill lifecycle is crucial for scalable AI development, where handcrafted skills are costly. This work reveals that extracting only high‑quality features and designing consumable skills can mitigate adverse effects, offering a practical path to more robust, adaptable agents.

## Related Concepts

- [[concepts/ai-agents/agentic-workflows-hub.md|Agentic Workflows Hub]]
- [[concepts/llm-models/llm-models-hub.md|LLM Models Hub]]
- [[concepts/evaluation-benchmarks/evaluation-benchmarks-hub.md|Evaluation Benchmarks Hub]]
- [[concepts/embodied-ai/embodied-ai-hub.md|Embodied AI Hub]]