Roundup of EvoSkill's Research: Paving the Self-Evolving Agents Frontier for Industry Leaders

In just four months, EvoSkill has been cited by more than 60 papers from researchers across 100+ institutions—and frontier labs are now positioning their own contributions in relation to it.

Recently accepted at COLM, EvoSkill has drawn citations from researchers at MIT, Carnegie Mellon, Tsinghua, Microsoft, Google, ETH Zürich, Alibaba, and Amazon.

We’re highlighting 14 of these papers to explore how researchers are building on, challenging, and extending EvoSkill’s approach to self-evolving AI.

SkillOS: Learning Skill Curation for Self-Evolving Agents

Frozen executor, trainable curator RL over grouped task streams

An RL-trained curator manages a frozen agent's skill repository, learning from feedback that arrives late and indirectly.

Skill curation is a learning problem, not a heuristic.

Here's the problem Google’s SkillOS paper goes after: self-evolving agents pile up experience faster than anyone can decide what's worth keeping. Their answer is to pair a frozen agent executor with a trainable skill curator, and to train the curator with reinforcement learning on grouped task streams with composite rewards — so later related tasks end up grading the curation decisions made earlier. That closes the loop on the hardest part of skill management. The payoff for a good curation decision only shows up on tasks later, and now something is learning from that.

SkillOS cites EvoSkill when it frames skill-based agent memory. That's the territory both systems now share.

Authors: Siru Ouyang, Jun Yan, Yanfei Chen, Rujun Han, Zifeng Wang, Bhavana Dalvi Mishra, Rui Meng, Chun-Liang Li, Yizhu Jiao, Kaiwen Zha, Maohao Shen, Vishy Tirumalashetty, George Lee, Jiawei Han, Tomas Pfister, Chen-Yu Lee. Google Cloud AI Research, with collaborators at UIUC and MIT.

Paper: https://arxiv.org/abs/2605.06614

SkillFoundry: Building Self-Evolving Agent Skill Libraries from Heterogeneous Scientific Resources

71.1% novel skills found in no existing library

A skill library mined from repositories, APIs, notebooks, and papers — most of it material no other library has.

The scientific literature is already a skill library. Nobody had compiled it.

SkillFoundry, from Jian Ma's lab at Carnegie Mellon University, mines heterogeneous scientific resources into validated, executable skill packages, organizes them with a domain knowledge tree, and refines them through closed-loop validation. It improves coding-agent performance on five of six MoSciBench datasets and on two real genomics workflows: cell type annotation and the scDRS pipeline.

When CMU lays out the field, EvoSkill is their example of the other path to a skill library — "iterative skill discovery through failure analysis." SkillFoundry compiles knowledge that already exists. EvoSkill evolves knowledge from experience. The field needs both.

Authors: Shuaike Shen, Wenduo Cheng, Mingqian Ma, Alistair Turcan, Martin Jinye Zhang, Jian Ma. Carnegie Mellon University.

Paper: https://arxiv.org/abs/2604.03964

From Context to Skills: Can Language Models Learn from Context Skillfully?

Zero supervision skills from pure self-play

Context-specific skills discovered and refined with no human supervision and no external feedback at all.

Skill discovery works even when there's no feedback signal to learn from.

Ctx2Skill, from Tsinghua University's NLP group with Maosong Sun among the authors, runs a multi-agent self-play loop — a Challenger, a Reasoner, and a Judge, plus Proposer and Generator skill updaters and a cross-time replay mechanism — that discovers context-specific skills entirely on its own. Across four context-learning tasks, the discovered skills substantially improve context learning, and they transfer across backbone models.

Their related work anchors on EvoSkill as the reference for automated skill construction by performing “failure-driven refinement into structured skill folders." That's the design their feedback-free setting positions itself against.

Authors: Shuzheng Si, Haozhe Zhao, Yu Lei, Qingyi Wang, Dingwei Chen, Zhitong Wang, Zhenhailong Wang, Kangyang Luo, Zheng Wang, Gang Chen, Fanchao Qi, Minjia Zhang, Maosong Sun. Tsinghua University, with collaborators at DeepLang AI, UIUC, Fudan, and CUHK.

Paper: https://arxiv.org/abs/2604.27660

SkillOpt: Executive Strategy for Self-Evolving Agent Skills

52 / 52 best or tied, every evaluated cell

Best or tied across six benchmarks, seven target models, and three harnesses — with EvoSkill as the named competitor.

A skill is the external state of a frozen model, and it can be trained like one.

SkillOpt, from Microsoft, Fudan University, Shanghai Jiao Tong University, and Tongji University, treats skill evolution the way an optimizer treats weights. An optimizer model turns scored rollouts into bounded add, delete, and replace edits, accepting them only when they strictly improve held-out validation. A textual learning rate, a rejected-edit buffer, and slow-update and meta-update cycles keep the process in check. The payoff: the skills it learns transfer across models, harnesses, and nearby benchmarks.

Their experiments name EvoSkill "the harness-side competitor" — skill-folder evolution under failure analysis is the reference point they measure against, and SkillOpt beats it by +14.0% and +3.2% in the Codex and Claude Code loops. We'll take that trade. Frontier labs now define their contributions relative to EvoSkill.

Authors: Yifan Yang, Ziyang Gong, Weiquan Huang, Qihao Yang, Ziwei Zhou, Zisu Huang, Yan Li, Xuemei Gao, Qi Dai, Bei Liu, Kai Qiu, Yuqing Yang, Dongdong Chen, Xue Yang, Chong Luo. Microsoft, with collaborators at Shanghai Jiao Tong, Tongji, and Fudan Universities.

Paper: https://arxiv.org/abs/2605.23904

Trace2Skill: Distill Trajectory-Local Lessons into Transferable Agent Skills

+57.65% skills from a 35B lifting a 122B

Skills evolved from Qwen3.5-35B trajectories improve a Qwen3.5-122B agent by up to +57.65% on WikiTableQuestions.

Skills distilled from a smaller model transfer up the scale ladder.

Trace2Skill, from Alibaba's Qwen Large Model Application Team with ETH Zürich, the University of Zurich, Peking University, and Zhejiang University, consolidates many execution trajectories in parallel into a unified skill directory via inductive reasoning — deepening existing human-written skills and growing new ones from weak LLM drafts. The evolved skills transfer across model scales, model families, and out-of-distribution settings.

Their Appendix G faithfully reproduces EvoSkill as one of "three full concurrent skill-evolution systems" in a head-to-head on SpreadsheetBench-Verified, under a shared open base model in both Claude Code and ReAct harnesses, where Trace2Skill outperforms it (65.8% to 69.8% against 33.5% to 59.5%). Independent reproduction under controlled conditions — that's exactly how this field should work, whichever way the numbers fall.

Authors: Jingwei Ni, Yihao Liu, Xinpeng Liu, Yutao Sun, Mengyu Zhou, Pengyu Cheng, Dexin Wang, Erchao Zhao, Xiaoxi Jiang, Guanjun Jiang. Qwen Large Model Application Team, Alibaba, with collaborators at ETH Zürich, University of Zurich, Peking University, and Zhejiang University.

Paper: https://arxiv.org/abs/2603.25158

Workflow Closure Is Not Scientific Closure in Auto-Research Systems

100+ papers · 21 systems audited for epistemic standing

A survey of over 100 papers and a structured audit of 21 auto-research systems, from Yale University.

Closing the loop is not the same as doing science.

Auto-research systems can now run idea-to-experiment-to-writing-to-self-evaluation loops end to end. This Yale position paper argues that closing that loop doesn't confer scientific standing. It diagnoses three recurring failure patterns — objective collapse, validation collapse, acceptance collapse — and argues for autonomous execution under non-autonomous epistemic control.

EvoSkill shows up in their literature review of skill acquisition and evolution, as part of the evidence for how far autonomous self-improvement has come — and why its epistemics now deserve real scrutiny. Fair. When your work becomes the reason a critique gets written, the field is taking it seriously.

Authors: Shuai Wang, Xinyuan Tian, Pangpang Liu, Yize Zhao. Yale University.

Paper: https://arxiv.org/abs/2605.26200

Pomona: Continuous Code Quality Improvement via Small, Automated Changes at Bloomberg

15 / 17 PRs merged median time-to-close under 2 hours

A skill-based agent shipping real pull requests at Bloomberg, with 8 of 10 senior engineers asking to adopt it.

Skill-based agents are already merging code in production.

Pomona is an experience paper from a live deployment, and that makes it our favorite kind of citation. It's a lightweight agentic tool: a Scanning skill builds a prioritized backlog of code-quality tasks, a Repair skill generates tiny, roughly ten-line pull requests, and humans stay in the loop. In a one-month team deployment at Bloomberg, 15 of 17 generated PRs were merged, with median time-to-close under two hours.

Pomona cites EvoSkill on the question it faces next for how to "evolve skills at scale." Once skills run in production, evolution stops being a research topic and becomes a maintenance strategy.

Authors: David Williams, Angelos Evripiotis, Serkan Kirbas, Harry Morgan, Sergey Magidovich, Peter Wainwright, Federica Sarro. University College London and Bloomberg.

Paper: https://arxiv.org/abs/2606.06752

SkillComposer: Learning to Evolve Agent Skills for Specification and Generalization

4B lifts a 27B +4.5% agent tasks · +3.4% code tasks

A 4B composer model improves a 27B executor by up to +4.5% on agent tasks and +3.4% on code tasks.

Skill evolution itself becomes a trainable operation.

SkillComposer, from Alibaba's Tongyi Lab with Zhejiang University and National University of Singapore, breaks skill construction into three learnable operations — create, improve, merge — trained via rejection sampling, so language models can self-evolve skills at inference time in offline, online, or hybrid deployment modes. It generalizes to unseen domains across tau2-Bench, LiveCodeBench v6, and AppWorld.

The authors cite EvoSkill as the skill-evolution method "closest to our work," pointing to how it discovers and edits skills through iterative failure analysis and validation.

Authors: Qi Zhang, Zhaopeng Feng, Xiaonan Shi, Xiaomeng Hu, Chu Liu, Pengjun Xie, Xiaobin Wang, Jieping Ye, Bryan Hooi, Haobo Wang, Junbo Zhao. Zhejiang University, Tongyi Lab, and National University of Singapore.

Paper: https://arxiv.org/abs/2606.06079

SkillGenBench: Benchmarking Skill Generation Pipelines for LLM Agents

1 of 5 pipelines the benchmark evaluates

Nine institutions — Peking University, Tsinghua University, Shanghai Jiao Tong University, National University of Singapore, Nanyang Technological University among them — pick five canonical skill-generation pipelines to measure. EvoSkill is one of them.

Skill generation finally gets measured in isolation.

Skill-generation results have always been tangled up with harness effects and evaluation quirks. SkillGenBench pulls the generation step out on its own: generators produce standardized skill artifacts from repository-grounded and document-grounded corpora, which run under fixed harnesses with deterministic, execution-based checks. Across methods and backbones it finds substantial performance variation, and distinct failure modes between repository-sourced and document-sourced skills.

EvoSkill is evaluated as the representative of "methods that derive skills from execution experience" — one of the five pipelines the benchmark treats as the field's canon. Nine institutions had to agree on that list.

Authors: Yifan Zhou, Zhentao Zhang, Ziming Cheng, Shuo Zhang, Qizhen Lan, Zhangquan Chen, Zhi Yang, Qianyu Xu, Ronghao Chen, Huacan Wang, Sen Hu. Shanghai Jiao Tong University, Xi'an Jiaotong University, NUS, QuantaAlpha, Tsinghua, SUFE, NTU, Peking University, and UCAS.

Paper: https://arxiv.org/abs/2605.18693

SkillGraph: Self-Evolving Multi-Agent Collaboration with Multimodal Graph Topology

4 benchmarks · 5 MAS structures consistent gains, 4 base models

Jointly evolving what agents know and how they communicate, for visual multi-agent systems.

Who talks to whom should evolve along with what they know.

SkillGraph, from National University of Singapore with Technical University Munich and Zhejiang University, evolves two things most systems treat separately. A Multimodal Graph Transformer predicts query-conditioned collaboration topologies, while a Skill Designer distills reasoning heuristics from failure cases into a self-evolving multimodal Skill Bank. Together they improve results consistently across four benchmarks, five multi-agent structures, and four base models.

They cite EvoSkill in the lineage of "dynamic injection and deep co-evolution of agentic skills through reinforcement learning and closed-loop analysis." Failure-driven skill design, carried into the multimodal world.

Authors: Zheng Nie, Ruolin Shen, Xinlei Yu, Bo Yin, Jiangning Zhang, Xiaobin Hu. National University of Singapore, with collaborators at Technical University of Munich and Zhejiang University.

Paper: https://arxiv.org/abs/2604.17503

SkillEvolBench: Benchmarking the Evolution from Episodic Experience to Procedural Skills

180 tasks · 6 environments adversarial shortcuts, context shift, composition

A diagnostic benchmark from seven universities and Amazon, testing whether agents can turn experience into reusable skills.

The stress test for every claim this field makes.

SkillEvolBench compares skill-evolution conditions against no-skill and raw-trajectory controls, under context shift, adversarial shortcuts, and composition. The verdict, across ten model configurations and three harnesses, is a challenge to the whole field: current agents adapt locally but rarely form robust reusable skills, and raw-trajectory reuse frequently outperforms distilled skills.

EvoSkill sits in the benchmark's cited foundations of skill-discovery work. The claims this benchmark exists to test are the ones our line of work put on the table — and we'd rather have them tested hard than taken on faith.

Authors: Yingtie Lei, Zhongwei Wan, Jiankun Zhang, Samiul Alam, Zixuan Zhong, Peizhou Huang, Xin Wang, Jingxuan Zhang, Donghao Zhou, Yunta Hsieh, Zhihao Dou, Hui Shen, Yan Xu, Dimitrios Dimitriadis, Tuo Zhang, Mi Zhang. Ohio State, with collaborators at UChicago, UCL, Michigan, CUHK, Case Western, and Amazon.

Paper: https://arxiv.org/abs/2605.24117

CoEvoSkills: Self-Evolving Agent Skills via Co-Evolutionary Verification

Highest pass rate on SkillsBench, Claude Code and Codex

The top score among five baselines on SkillsBench, generalizing across eight LLMs in total.

Skill evolution without ground truth.

CoEvoSkills, from a six-institution team including Columbia University, McGill University, and University of British Columbia, pairs a Skill Generator that iteratively refines multi-file skill packages with a co-evolving Surrogate Verifier that provides feedback without ever seeing ground-truth test content.

They're explicit about their starting point. EvoSkill "relies heavily on ground-truth supervision for failure diagnosis, limiting applicability in real-world settings" — that's the limitation their co-evolutionary verifier exists to remove. Our failure-analysis loop, their label-free verifier. This is how research compounds.

Authors: Hanrong Zhang, Shicheng Fan, Henry Peng Zou, Yankai Chen, Zhenting Wang, Jiayu Zhou, Chengze Li, Wei-Chieh Huang, Yifei Yao, Kening Zheng, Xue (Steve) Liu, Xiaoxiao Li, Philip S. Yu. University of Illinois Chicago, with collaborators at MBZUAI, McGill, Columbia, Zhejiang University, and UBC.

Paper: https://arxiv.org/abs/2604.01687

Experience Compression Spectrum: Unifying Memory, Skills, and Rules in LLM Agents

50–500x the compression band where skills live

Amazon's unifying framework places episodic memory at 5–20x compression, procedural skills at 50–500x, declarative rules at 1000x+.

Memory, skills, and rules are one axis, not three separate fields.

From the AWS Generative AI Innovation Center, this paper maps more than 20 agent-learning systems onto a single spectrum of experience compression. What motivated it is telling: a citation analysis showing the three communities cite each other less than 1% of the time. The mapping exposes what the authors call the missing diagonal — every system operates at one fixed compression level, and none learns across the full spectrum.

EvoSkill plays three roles in the paper: It's mapped onto the spectrum as a Level-2 procedural-skill system, cited as evidence via our SealQA-to-BrowseComp +5.3% zero-shot transfer, and it's proposed as a component of the paper's future architecture — "EvoSkill's Pareto selection for promotion." Baseline, evidence, and building block, all in one paper.

Authors: Xing Zhang, Guanghui Wang, Yanwei Cui, Wei Qiu, Ziyuan Li, Bing Zhu, Peiyang He. AWS Generative AI Innovation Center and HSBC Technology Center.

Paper: https://arxiv.org/abs/2604.15877

From Raw Experience to Skill Consumption: A Systematic Study of Model-Generated Agent Skills

Extractor ≠ consumer decoupled abilities, independent of scale

A model can be a strong skill extractor and a weak skill consumer — and scale does not fix it.

The field's first systematic reality check on model-generated skills.

This study from Fudan University, Microsoft Research, and Shanghai Jiao Tong University builds a utility-grounded evaluation across the full skill lifecycle — experience generation, skill extraction, skill consumption — over five agentic task domains. Model-generated skills help on average but show non-trivial negative transfer, and extractor and consumer turn out to be decoupled capabilities.

They cite EvoSkill among the methods that pushed skill production toward automation at scale. In other words: the shift this paper exists to measure.

Authors: Zisu Huang, Jingwen Xu, Yifan Yang, Ziyang Gong, Qihao Yang, Muzhao Tian, Xiaohua Wang, Changze Lv, Xuemei Gao, Qi Dai, Bei Liu, Kai Qiu, Xue Yang, Dongdong Chen, Xiaoqing Zheng, Chong Luo. Fudan University, with collaborators at Microsoft Research and Shanghai Jiao Tong University.

Paper: https://arxiv.org/abs/2605.23899

And then it left the lab

The fourteen above are research. But a citation trail this size picks up something better than research: production systems.

Alibaba Group's Taobao team built SkillChain for the image-based AI assistants running at production scale on their e-commerce platforms, and places EvoSkill in the lineage of systems that evolve skills from interaction traces via failure-driven refinement. ByteDance built MUSE-Autoskill for self-evolving agents and works through EvoSkill's design in detail — the failure analysis, the Pareto-frontier selection — in its lifecycle comparison of the field. An Elastic security researcher cites EvoSkill in a position paper on self directed skill acquisition for cybersecurity agents, noting that evolved skills transfer zero-shot to unseen tasks. Last but not least, Shanghai AI Laboratory's Safactory, an infrastructure for training trustworthy autonomous agents, builds on the idea it credits to EvoSkill and its neighbors: discovering high-level policies from historical trajectories and continuously evolving them.

E-commerce, security operations, code quality, agent infrastructure. Four months from arXiv to other people's production stacks.

The surveys have arrived

There's a moment in every research area when the surveys show up — when the work stops being a scatter of papers and becomes a field with a shape. That moment came fast here, and EvoSkill is on every map.

Carnegie Mellon's "They Are Not Static: A Survey of Dynamic Agent Skills" references EvoSkill twenty-one times, placing it among the execution-grounded systems that gate skill admission at write time — the proposal, verify, admit, execute family. "A Survey on AI Agent Harness" classifies it as automated multi-agent skill discovery in its taxonomy of the skill layer. A dedicated survey of agent skills credits it with a specific mechanism: "EvoSkill proposes multi-agent revision loops that retain only validated improvements." Another places it at the refinement stage of the skill lifecycle it defines. A fifth, on self-evolving agents as dynamic graph transformation, folds it into the broader self-evolution story.

Five surveys, five taxonomies, one framework showing up in all of them. When the people mapping the field need your work to draw the map, you've become part of the infrastructure.

The numbers

Four months after release:

• 60+ citing papers, from over 100 institutions across four continents
• 15 of the biggest names in AI among the affiliations: MIT, Carnegie Mellon University, Tsinghua University, Microsoft Research, Google Cloud AI Research, ETH Zürich, Yale University, University College London, Alibaba's Qwen and Tongyi teams, Peking University, National University of Singapore, University of Chicago, Columbia University, Amazon, and Shanghai Jiao Tong University, Fudan University — plus ByteDance, Shanghai AI Laboratory, Bloomberg, Elastic, Meituan, OPPO, and Xiaohongshu beyond them
• A named baseline or evaluated pipeline in work from Microsoft, Alibaba's Qwen team, and the nine-institution SkillGenBench coalition; a proposed building block in Amazon's Experience Compression Spectrum
• Five survey papers already placing EvoSkill in their taxonomies of the field
• Production citations from Bloomberg, Alibaba, ByteDance, Elastic, and Shanghai AI Laboratory
• 1,000 stars on https://github.com/sentient-agi/EvoSkill

And the results that started it all. Installed into Claude Code, EvoSkill pushed SOTA on OfficeQA from 60.6% to 68.1% and on SealQA from 26.6% to 38.7%, and transferred a SealQA-evolved skill zero-shot to BrowseComp for a +5.3% gain (43.5% to 48.8%) — no human involved at any step. We saw similarly large improvements with OpenCode, OpenHands, Goose, and Codex CLI.

Learn more about EvoSkill

• Read the original EvoSkill Research Paper
• Explore our EvoSkill Technical Deep Dive
• See how we put the upcoming EvoSkill V2 to the test: by building an AI coach tasked with helping another AI improve its test scores

Install EvoSkill and get started here: https://github.com/sentient-agi/EvoSkill