ScienceCUA: Resource-to-Skill Construction and Execution-Guided Skill Evolution for Scientific Computer Use Agents

Daiqiang Li; Jingyi Peng; Rui Hu; Jia Li; Xuyang Liu; Guanren Qiao; Yujia Xiang; Qiuzhuang Sun

cs.AINov 2, 2026

ScienceCUA: Resource-to-Skill Construction and Execution-Guided Skill Evolution for Scientific Computer Use Agents

Authors: Daiqiang Li, Jingyi Peng, Rui Hu, Jia Li, Xuyang Liu, Guanren Qiao, Yujia Xiang, Qiuzhuang Sun,

Organizations: DeepAutonomy Technology Co., Ltd. · Peking University · College of AI, Tsinghua University · The Hong Kong Polytechnic University ·

Abstract

Computer-use agents for scientific software require application-specific expertise. Much of this expertise is documented in human-authored software manuals and web pages. However, turning these resources into executable, reusable skills remains challenging due to mismatches between documented and deployed software versions and differences in scope between documentation coverage and task requirements. We introduce ScienceCUA, a two-stage framework combining resource-to-skill construction with execution-guided skill evolution. First, ScienceCUA identifies tasks supported by the software from its documentation, infers operational trajectories, and generalizes them into reusable Skills. Command catalogs, syntax constraints, and shortcuts are organized separately as References for on-demand lookup. Then, ScienceCUA uses execution feedback to drive skill self-evolution by contrasting successful and failed trajectories for the same task, identifying deficiencies, revising existing skills or adding new ones, and updating the skill library only after repeated validation in real environments. To enable reliable evaluation without benchmark noise or task leakage, we introduce two improved benchmarks. ScienceBoard-Verified provides an audited and corrected basis for measuring task performance, while ScienceBoard-VE separates skill evolution from final testing through disjoint evolution and test tasks and further supports native-agent evaluation. Experiments show that the initial Skills and References improve average task success rate by up to 7.23 percentage points over the base model, while skill self-evolution increases the total gain over the baseline to as much as 17.12 percentage points. ScienceCUA turns software documentation into an evolving set of executable skills and queryable references, offering a practical way for general-purpose agents to use scientific software reliably and at scale.

Figures & tables

Figure 2: Motivation for resource-to-skill construction. (a) Scientific applications and their resources. (b) Human-oriented narrative content. (c) Sparse task-relevant information. (d) Effects of directly injecting raw resources on success rate and total execution time.

Figure 3: An example in KAlgebra: a procedural Skill for plotting 3D functions (left) and a Reference cataloging supported commands (right).

Figure 4: Task distribution in ScienceBoard-VE. The inner ring groups tasks by application. The outer ring separates verified evaluation tasks (V) from evolution tasks (E). R+X denotes revised or removed tasks in the original benchmark. The removed task is excluded from the sector sizes.

Figure 5: Task execution time, initial library construction cost, and resource use during execution. (a) Total task execution time summed across all six applications for four models under six methods. Each curve represents one model, and radial values indicate hours. (b) Total initial library construction time across the six applications for VisualSkill, SkillFoundry, and ScienceCUA, using GPT-5.5 (xhigh). (c) Completed calls to Skills and References in final task attempts, averaged over all tasks in each application, for native Grok Build with Grok-4.6 (xhigh). White labels give Skill and Reference calls per task in their respective colors; the narrow GRASS GIS Skill and TeXstudio Reference segments have connected labels. Bar-end labels give the combined calls per task. Values are rounded independently. The fractions on the right show the number of tasks with at least one resource call over the total number of tasks in each application.

Model	Method	KAlgebra	Celestia	GRASS GIS	ChimeraX	Lean 4	TeXstudio	Average
	Full ScienceCUA	80.65%	63.64%	47.06%	85.71%	19.23%	31.25%	54.59%
	$-$ Knowledge addition	80.65%	63.64%	32.35%	85.71%	15.38%	31.25%	51.50%	( $\downarrow$ 5.66%)
Grok-4.6 (xhigh)	$-$ Failure history	70.97%	54.55%	41.18%	82.14%	15.38%	25.00%	48.20%	( $\downarrow$ 11.71%)
	initial library	70.97%	48.48%	26.47%	85.71%	11.54%	25.00%	44.70%	( $\downarrow$ 18.12%)
	Trace2Skill	70.97%	51.52%	23.53%	82.14%	19.23%	31.25%	46.44%
	SkillCAT	74.19%	63.64%	44.12%	85.71%	19.23%	31.25%	53.02%

Table 3: Self-evolution comparison and ablation on ScienceBoard-Verified. Results are success rates in percent. All variants are evaluated with frozen libraries and no evolution history.

Appendix figures & tables5 assets

Supplementary material from the paper’s appendix.

Appendix

Software	Environment version	Resource version	Format	Documentation source
KAlgebra	24.11.70	20.12	PDF	https://docs.kde.org/stable_kf6/en/kalgebra/kalgebra/kalgebra.pdf
Celestia	1.7.0	1.6.1	PDF	https://drive.google.com/file/u/1/d/1dxMnahknnVjrckqWOyJhvuS8IA0BCT0l/view
ChimeraX	1.9	1.9.0	HTML	https://web.archive.org/web/20241215145343id_/https://www.cgl.ucsf.edu/chimerax/docs/user/index.html ; root prefix /chimerax/docs/user/
GRASS GIS	8.4.1	8.4.1	HTML	https://lira.no-ip.org:8443/doc/grass-doc/html/index.html
Lean 4	4.14.0	4.26.0	HTML	https://lean-lang.org/theorem_proving_in_lean4/ ; root prefix /theorem_proving_in_lean4/
TeXstudio	4.8.6	4.9.8	HTML	https://texstudio-org.github.io/

Appendix

Table 8: Inputs used for the VisualSkill Stage 1 reproduction.

Stage	Executable contract
Evidence Architect (HTML)	Input. One normalized HTML root scope (root page plus all eligible in-scope child pages); root-authored structure; candidate Evidence skeleton Tools. inspect_source , get_root_structure , submit_evidence_revision Output / budget. EvidenceRevision ; exactly 3 calls in the default workflow
Evidence Architect (PDF)	Input. One normalized PDF document; selected navigation tree; content assignments; residual content; candidate skeleton Tools. inspect_source , get_pdf_document_structure , bounded read_pdf_content and read_evidence_node , submit_evidence_revision Output / budget. EvidenceRevision ; at most 6 calls
Reference Build	Input. Frozen CorpusSnapshot and Python-generated review manifest Tools. Candidate/catalog inspection, bounded Evidence reading, submit_reference_review Output / budget. Reference review decisions; at most 8 calls
Task Discovery	Input. Frozen CorpusSnapshot Tools. Evidence inventory, search/read/image inspection, Task batch submission Output / budget. TaskDraft ; exploration budget $\max(24,\min(40,3\,\texttt{max\_steps}))$
Trajectory Inference	Input. One validated Task plus Corpus/Knowledge snapshots Tools. Evidence, Task, Reference and Reference-entry reads; trajectory submission Output / budget. TrajectoryDraft ; $\max(24,2\,\texttt{max\_steps})$ plus 2 protocol calls
Skill Induction	Input. A batch of validated Trajectories plus KnowledgeSnapshot Tools. Upstream-object reads, Skill submission, coverage inspection/review Output / budget. SkillDraft ; $\max(40,3\,\texttt{max\_steps})$ plus 6 protocol calls

Appendix

Table 9: Executable contracts for the Resource-to-Skill child sessions. Budgets refer to the default production configuration; protocol-reserved calls are listed separately where applicable.

Task	Proof goal
A-01	Derive reflexivity from symmetry, transitivity, and an existence condition for a relation.
A-02	Prove a recurrence inequality for the Ackermann function.
A-03	Relate transitivity of a ZF set to powerset and union conditions.
A-04	Prove a cofinality property of a limit ordinal.
A-05	Prove a statement involving cardinal operations.
B-01	Prove a modular congruence under primality and coprimality assumptions.

Appendix

Table 10: Lean 4 Verified tasks and their proof goals. The entries summarize the full formal task statements supplied with ScienceBoard-VE.

Task	Task goal	Additional information or tool
A-01	Enter a molecular angle in the article.	Measurement in ChimeraX.
A-02	Enter a residue name in the article’s table.	Residue information from ChimeraX.
A-03	Correct a DNA sequence in the article.	Sequence information from ChimeraX.
A-04	Correct a Fibonacci value in the article.	Computation in KAlgebra.
B-01	Complete a \cite command for an existing bibliography entry.	None.
B-02	Enter the COSPAR ID of Gemini 9A and add a citation.	Object information from Celestia.

Appendix

Table 11: TeXstudio Verified tasks and information required beyond editor operation. The entries summarize the full task instructions supplied with ScienceBoard-VE.

Figure 11: Success rates across the six applications in ScienceBoard-Verified. All models use the xhigh reasoning setting. ScienceCUA provides Skills and References that have undergone self-evolution. Labels above the bars give success rates in percent.

Computer-Use Agents

Computer-use agents perceive and operate graphical interfaces and software tools to complete tasks across websites, apps, and desktops.

See all 171 papers

SKILLFOUNDRY: Building Self-Evolving Agent Skill Libraries from Heterogeneous Scientific Resources
Apr 5, 2026Shuaike Shen, Wenduo Cheng, Mingqian Ma +3Agent Skill Learning AI Agents for Scientific Discovery
Ray and Stephanie Lane Computational Biology Department, School of Computer Science, Carnegie Mellon University · Machine Learning Department, School of Computer Science, Carnegie Mellon University
EvoSkill: Automated Skill Discovery for Multi-Agent Systems
Mar 3, 2026Salaheddin Alzubi, Noah Provenzano, Jaydon Bingham +2Agent Skill Learning LLM Agent Self-Improvement
Sentient · Virginia Tech
Scientific Agent Skills: A Library of Procedural Knowledge for Research Agents
Aug 30, 2026Timothy Kassis, Vinayak Agarwal, Yuhuan He +2AI for Science LLM Agent Skill Retrieval
K-Dense, Inc.

Apr 5, 2026J/K move · Enter open · S save

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

Shuaike Shen, Wenduo Cheng, Mingqian Ma, Alistair Turcan+2

Ray and Stephanie Lane Computational Biology Department, School of Computer Science, Carnegie Mellon University · Machine Learning Department, School of Computer Science, Carnegie Mellon University

Agent Skill Learning AI Agents for Scientific Discovery LLM Agent Self-Improvement

Open paper AskPDF arXiv

Modern scientific ecosystems are rich in procedural knowledge across repositories, APIs, scripts, notebooks, documentation, databases, and papers, yet much of this knowledge remains fragmented across heterogeneous artifacts that agents cannot readily operationalize. This gap between abundant scientific know-how and usable agent capabilities is a key bottleneck for building effective scientific agents. We present SkillFoundry, a self-evolving framework that converts such resources into validated agent skills, reusable packages that encode task scope, inputs and outputs, execution steps, environment assumptions, provenance, and tests. SkillFoundry organizes a target domain as a domain knowledge tree, mines resources from high-value branches, extracts operational contracts, compiles them into executable skill packages, and then iteratively expands, repairs, merges, or prunes the resulting library through a closed-loop validation process. SkillFoundry produces a substantially novel and internally valid skill library, with 71.1% of mined skills differing from existing skill libraries such as SkillHub and SkillSMP. We demonstrate that these mined skills improve coding agent performance on five of the six MoSciBench datasets. We further show that SkillFoundry can design new task-specific skills on demand for concrete scientific objectives, and that the resulting skills substantially improve performance on two challenging genomics tasks: cell type annotation and the scDRS workflow. Together, these results show that automatically mined skills improve agent performance on benchmarks and domain-specific tasks, expand coverage beyond hand-crafted skill libraries, and provide a practical foundation for more capable scientific agents.

Abstract

Computer-Use Agents

Explore similar work

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

EvoSkill: Automated Skill Discovery for Multi-Agent Systems

Scientific Agent Skills: A Library of Procedural Knowledge for Research Agents