Scholar-Skill: Artifacts for Scholarly Work

• Scholar-Skill is a framework that defines portable, agent-facing artifacts embedding instructions, procedures, and references for a range of scholarly tasks.
• It differentiates between general and task-specific skills to mitigate challenges like context overhead and knowledge conflicts in scholarly workflows.
• It integrates artifact contracts, graph-based retrieval, and continual auditing to enhance literature search, evidence extraction, and reproducibility.

Scholar-Skill can be understood as a family of agent skill systems for scholarly work in which a skill is treated as a portable, agent-facing artifact that packages instructions, procedures, references, and output conventions for a class of research tasks. In the recent literature, these systems are applied to literature search, source triage, evidence extraction, verification and synthesis, citation tracking, recovery, reproducibility checks, and related scholarly workflows, while also supporting retrieval, internalization, correction, rollback, and audit of the skill artifacts themselves (Li et al., 7 Jun 2026, Zhou et al., 29 May 2026, Yu et al., 21 Jun 2026).

1. Conceptual foundations

A central premise of Scholar-Skill is that scholarly agency is not reducible to a single monolithic prompt. Recent work separates at least two orthogonal distinctions. First, Skill0.5 divides skills into general skills and task-specific skills. General skills are domain-agnostic cognitive heuristics such as systematic exploration, error recovery, and meta-reasoning; task-specific skills are fine-grained execution rules tied to a domain or task family. Second, COLLEAGUE.SKILL divides a person-grounded scholarly skill into a capability track and a bounded behavior track. The capability track covers responsibilities, workflows, technical standards, review criteria, decision heuristics, and mental models, whereas the bounded behavior track covers expression constraints, tone and register, interaction rules, citation style policies, and correction history (Zhu et al., 27 May 2026, Zhou et al., 29 May 2026).

This dual taxonomy addresses two distinct failure modes. Skill0.5 argues that full externalization of all skills into the prompt causes context overhead, “lost in the middle” effects, and contextual interference, while full internalization risks model-capacity limits, knowledge conflicts, execution hallucination, and parametric conflicts. It also distinguishes these issues from memory-augmented RL methods, which may inject noisy, overly detailed trajectories and thereby yield brittle behavior and degraded out-of-distribution generalization (Zhu et al., 27 May 2026).

In scholarly settings, the task surface is correspondingly broad. SkillHone enumerates search and discovery, source triage, evidence extraction, verification and synthesis, citation tracking, and recovery as core scholarly skill categories. COLLEAGUE.SKILL gives a complementary scholarly decomposition through literature search strategies, paper triage criteria, study design evaluation, argumentation structure analysis, citation management and formatting, reproducibility checks, and reviewer ethics. Taken together, these formulations treat scholarly competence as structured procedural knowledge rather than as latent background knowledge alone (Li et al., 7 Jun 2026, Zhou et al., 29 May 2026).

2. Skill representations and artifact contracts

Skill representations in Scholar-Skill are explicitly artifact-centric. SkillHone defines a skill as “a loadable artifact that packages the instructions, procedures, references, and output conventions an agent uses for a class of tasks,” with a portable bundle interface centered on a SKILL.md entry file and optional scripts, references, templates, and supporting resources. SkillJuror similarly treats a skill as a navigable directory centered on SKILL.md plus optional scripts, references, assets, and metadata, emphasizing that the agent actively chooses which files to open and which helpers to invoke during task execution (Li et al., 7 Jun 2026, Chen et al., 10 Jun 2026).

COLLEAGUE.SKILL formalizes the package as

$S = (A, M, L),$

where $A$ are generated files, $M$ is machine-readable metadata and installation information, and $L$ is lifecycle state. Its dual-track layout uses work.md and work_skill.md for capability, persona.md and persona_skill.md for bounded behavior, and a combined SKILL.md as the invokable entrypoint. manifest.json and meta.json carry installers, provenance, compatibility, distribution metadata, versioning, correction count, and rollback history; the reported schema version is 3 (Zhou et al., 29 May 2026).

Other frameworks emphasize different representational granularity. Skill0.5 stores skills in a hierarchical skill bank

$S = S_G \cup S_S,$

with general skills $S_G$ and domain-tied specific skills $S_S$, both represented as modular textual directives. Skill1 represents each library entry as text with two fields, s.strat and s.desc, supporting strategy execution and scenario-conditioned retrieval. SkillX instead decomposes a reusable skill knowledge base into planning skills, functional skills, and atomic skills, and gives a unified schema with name, document, content, and tools fields; this schema explicitly records inputs, outputs, usage notes, constraints, pre/postconditions, tool invocation patterns, parameter regimes, and failure modes (Zhu et al., 27 May 2026, Shi et al., 7 May 2026, Wang et al., 6 Apr 2026).

These representational choices are significant because they define what can be inspected, corrected, versioned, and retrieved. COLLEAGUE.SKILL makes this explicit by avoiding opaque model tuning in favor of inspectable files and explicit lifecycle state, while SkillHone links each revision to diagnoses, redacted evidence, and outcomes. This suggests that, within Scholar-Skill, the artifact boundary is not only a packaging decision but also the unit of governance and longitudinal maintenance (Zhou et al., 29 May 2026, Li et al., 7 Jun 2026).

3. Learning, internalization, and out-of-distribution generalization

One major line of work asks which parts of scholarly competence should remain external artifacts and which should be absorbed into model parameters. SKILL0 addresses this through in-context reinforcement learning for skill internalization. Training begins with full skill context and progressively withdraws it under a linearly decaying skill budget, while a Dynamic Curriculum retains only skills from which the current policy still benefits. At the end of training, the agent operates in a fully zero-shot setting without runtime skill retrieval. On ALFWorld, SKILL0 improves over the standard RL baseline by +9.7%; on Search-QA, the gain is +6.6%, while maintaining fewer than 0.5k tokens per step (Lu et al., 2 Apr 2026).

Skill0.5 makes a stronger type distinction. It routes tasks into Hard, Medium, and Easy tiers using empirical pass rates under the same prompt used at inference:

$$\mathcal{M}(x_i)= \begin{cases} \text{Hard}, & p_i = 0 \ \text{Medium}, & 0 < p_i \le \eta_t \ \text{Easy}, & p_i > \eta_t \end{cases}$$

and then applies tier-specific optimization. Hard tasks internalize general skills through privileged distillation, Medium tasks use standard GRPO, and Easy tasks use diagnostic probing to penalize shortcut learning and enforce use of retrieved specific skills. Its combined objective is

$$\mathcal{L}=\mathcal{L}_{\text{hard}}+\mathcal{L}_{\text{medium}}+\mathcal{L}_{\text{easy}}.$$

At inference, only specific skills are provided because general skills have been internalized. On ALFWorld, Skill0.5 reports 93.1% ID average and 58.5% OOD average; on WebShop, it reports 40.4% ID and 40.6% OOD, outperforming both memory-based and skill-based baselines in the reported setups (Zhu et al., 27 May 2026).

Skill1 moves from differentiated treatment to unified co-evolution. It trains a single policy to generate a search query, re-rank retrieved candidates, solve the task conditioned on the selected skill, and distill a new skill from the trajectory, all under a single task-outcome objective

$$J(\theta)=\mathbb{E}_{x\sim D,\tau\sim\pi_\theta(\cdot|x)}[r(\tau)].$$

Its low-frequency utility trend credits selection, its high-frequency variation credits distillation, and raw task outcome credits utilization. The framework reports 97.5 average success on ALFWorld and 82.9 WebShop success, with ablations showing degradation when selection credit, distillation credit, or the library itself is removed (Shi et al., 7 May 2026).

A plausible implication for Scholar-Skill is that scholarly systems benefit from type-specific learning rules. General scholarly heuristics—query decomposition, evidence triangulation, factual verification, or “know when to search”—fit the internalization regime described by Skill0.5 and SKILL0, whereas volatile domain-bound procedures such as citation normalization, venue-specific formatting, or tool-API fallbacks remain better aligned with external skill utilization (Zhu et al., 27 May 2026, Lu et al., 2 Apr 2026).

4. Retrieval, organization, and runtime composition

When skills remain external, runtime performance depends not only on which skills are retrieved but also on how they are organized and compiled. SkillGraph replaces flat libraries with a directed graph whose nodes are skills and whose typed edges encode prerequisite, enhancement, and co-occurrence relations. Retrieval expands from seed skills through backward prerequisite recovery and forward beam search, then produces an ordered subgraph by topological sorting over dependency edges. In the reported configuration, retrieval is capped at $A$0 skills. On ALFWorld, removing graph-aware retrieval drops success from 90.6 to 59.4; on WebShop, removing graph structure reduces success from 84.4 to 72.7 (Li et al., 12 May 2026).

SkillRAE focuses on the compilation layer between retrieval and execution. It builds a multi-level skill graph over communities, skills, and reusable subunits, and then constructs a compact context

$A$1

under a token budget. Online retrieval combines top-down community matching and bottom-up subunit evidence, exports highlighted subunits from selected skills, and uses rescue-aware compact compilation to affiliate additional high-value cues back to the selected skills rather than exposing free-floating snippets. With a budget of 384 tokens, SkillRAE reports a SkillsBench reward mean of 29.26 versus 26.20 for curated skills and 22.04 for SkillRouter; removing context compilation lowers the reported score to 22.59 (Meng et al., 11 May 2026).

SkillJuror isolates a different variable: organization with content held fixed. It compares Progressive Disclosure, in which a concise root file points to supporting resources on demand, with a normalized flat baseline containing the same knowledge in a single self-contained root. In an 82-task SkillsBench study, Progressive Disclosure raises distinct Skill resources touched per trajectory from 1.18 to 3.85, mean effective uptake events from 1.33 to 3.92, and verifier-passing trials from 172/410 to 189/410, a gain of 17 matched trials or about +4.1%. The reported effect is task-dependent: it helps when supporting resources guide implementation, checking, or repair, and is weaker when success depends on exact output conventions, numerical thresholds, or long artifact-generation pipelines (Chen et al., 10 Jun 2026).

SkillX offers yet another runtime architecture. It uses pseudo-plan rewriting to guide retrieval, then performs step-wise retrieval of functional and atomic skills, deduplicates across steps, and injects the final compact skill set once into the system prompt. Its vector store uses FAISS HNSW over Qwen3-Embedding-8B vectors, with broad retrieval of Top-100 nearest skills, thresholding at cosine similarity $A$2 and within 0.08 of the best match, near-duplicate removal above cosine 0.95, and diversity-aware selection through MMR with weight 0.75, returning up to 8 skills (Wang et al., 6 Apr 2026).

For Scholar-Skill, these results indicate that scholarly performance is shaped by retrieval topology, subunit exposure, and document layout, not merely by the presence of additional instructions. The same procedural content can induce different runtime behavior depending on whether it is flattened, hierarchically routed, graph-ordered, or compiled into a constrained advisory packet (Meng et al., 11 May 2026, Chen et al., 10 Jun 2026, Li et al., 12 May 2026).

5. Continual evolution, auditing, and skill-centered assessment

Scholar-Skill systems are not static repositories; several frameworks make persistence and auditability first-class. SkillHone formalizes continual skill evolution as

$A$3

with role-separated optimization and evaluation subagent teams plus a runtime dispatcher. Each iteration appends a decision record

$A$4

where $A$5 is a diagnosis, $A$6 a candidate revision, $A$7 redacted evaluation evidence, and $A$8 an outcome in $A$9. Because later optimization dispatches retrieve relevant prior records from $M$0, they can avoid re-trying failed fixes, revisit rejected alternatives under updated evidence, and target only the problematic part of a regressed change. In the reported raw open-web evaluations, SkillHone reaches 64.6 average on GAIA versus 48.8 for the deep-research agent, and 66.4 on WebWalkerQA-EN versus 63.2 (Li et al., 7 Jun 2026).

COLLEAGUE.SKILL implements a related but more artifact-oriented correction lifecycle. Natural-language feedback is classified into capability or behavior edits; capability corrections patch work.md, whereas behavior corrections are normalized into records of the form {scene, wrong, correct} appended to the correction log in persona.md. Version numbers increment, previous states are archived, and rollback points are maintained in meta.json. This makes correction history part of the skill artifact rather than an external training log (Zhou et al., 29 May 2026).

SkillAudit shifts attention from fixed task-suite benchmarking to skill-centered assessment. It auto-generates capability-aligned utility tasks and safety probes directly from a skill package, executes them in isolated sandbox environments, and aggregates utility, efficiency/cost, and safety into a final report tuple. Its pass-rate gain per utility pair is

$M$1

and its safety score subtracts confidence-weighted penalties from 100 with a floor of 10. In a scan of 226 top-ranked real-world skill packages across 23 occupational categories, 17 skills, or 7.5%, were flagged as risky under the paper’s $M$2 boundary. Under Codex / GPT-5.4, the reported aggregate mean PRG is 0.183, while the overall efficiency-cost gain is negative, $M$3, indicating that enabling a skill often improves task success while increasing time and token consumption (Yu et al., 21 Jun 2026).

SkillJuror complements this with process metrics rather than only end outcomes: mean distinct resources $M$4, mean effective uptake $M$5, ERU rate, verifier-passing trials, and efficiency normalized by passes. This is important for scholarly systems because procedure changes may alter how agents search, validate, or repair before aggregate benchmark outcomes move materially (Chen et al., 10 Jun 2026).

6. Scholarly workflows, data infrastructure, and open constraints

The scholarly use cases proposed across the literature are unusually concrete. SkillHone’s scholarly taxonomy includes search and discovery, source triage, evidence extraction, verification and synthesis, citation tracking, and recovery. COLLEAGUE.SKILL adds scholar-specific heuristics such as seed-paper search with forward and backward citation chaining, paper triage decisions based on scope alignment and methodological validity, study design evaluation, claim–evidence–warrant mapping, citation rules that prefer DOIs and persistent URLs while avoiding secondary citations when primary is available, reproducibility actions such as rerunning notebooks in a clean environment, and reviewer-ethics rules such as avoiding identity guesses and declaring conflicts. Its bounded behavior track encodes a formal, precise, cautious register, with hedging patterns such as “the evidence suggests” and “a plausible interpretation is” when evidence is limited (Li et al., 7 Jun 2026, Zhou et al., 29 May 2026).

These workflows depend on external scholarly infrastructure. The Semantic Scholar Open Data Platform describes the Semantic Scholar Academic Graph with 200M+ papers, 80M+ authors, 550M+ paper-authorship edges, and 2.4B+ citation edges, plus structurally parsed text, TLDR summaries, citation intent and influence signals, and SPECTER embeddings. It exposes this information through Graph, Datasets, Recommendations, and Peer Review APIs, while S2ORC provides open-access full text with structural annotations. The paper explicitly recommends Graph API paper/author search, citation-graph expansion, author and venue navigation, TLDR-assisted triage, and embedding-based retrieval as integration patterns for a scholarly assistant (Kinney et al., 2023).

Google Scholar remains a complementary but operationally distinct source. It is described as an academic web search engine that automatically crawls and indexes academic-like documents from a very wide range of web sources and is roughly 2.7–2.8× the size of Web of Science and Scopus by source records in the chapter’s 2017 estimates. It indexes quickly and captures early citations, but the same source documents emphasize its limits for research assessment: no API, a 1,000-result cap, unstable hit counts, limited filters, no controlled vocabularies for authors or institutions, metadata errors, non-peer-reviewed inclusions, and susceptibility to manipulation. For Scholar-Skill, this makes Google Scholar useful as a broad supplementary source, especially in Humanities and Social Sciences, but methodologically riskier as a sole ground-truth substrate (López-Cózar et al., 2018).

Reported constraints remain substantial. Skill0.5 is validated in text-based embodied and web environments and notes extension to code agents, multimodal settings, and longer horizons as future work; it also depends on the quality of privileged general skills and on pass-rate routing that may misclassify tasks under stochasticity. SkillHone currently evolves a single skill in isolation and reports English-language benchmarks only. COLLEAGUE.SKILL identifies trace quality, editor bias in corrections, privacy risks for confidential materials, and domain drift as central maintenance problems. SkillAudit notes limited skill coverage, single-run conditions, rubric and judge bias, sandbox mismatch, and configuration-dependent exploitability. These constraints indicate that Scholar-Skill is best understood as a governed systems problem spanning artifact design, continual evaluation, data provenance, privacy, and changing scholarly toolchains rather than as a one-shot prompting technique (Zhu et al., 27 May 2026, Li et al., 7 Jun 2026, Zhou et al., 29 May 2026, Yu et al., 21 Jun 2026).