Skill evolution 88 papers

Every note, newest paper first

1. 23% of Agent Skills Already Bundle Executable Code, Not Just Prompts

   2026-07-21 — Skillware is Fan and Lan's name for what an agent skill actually is once you stop treating it as a prompt: a three-layer object. The Skill Artifact is just the natural-language task spec. Wrapped…

2. A co-evolving skill library lifts tool-use accuracy from 27.7% to 32.0% — with fewer tool calls, not more

   2026-07-15 — SPyCE trains multimodal agents that think with images by distilling every successful multi-step trajectory into a two-tier hierarchical skill library , rather than collapsing it into a scalar…

3. Ten anchored examples recover 88-110% of an oracle metric's gains

   2026-07-14 — Self-evolving agent loops assume a reliable evaluator already exists to grade each attempt. This paper drops that assumption and evolves the evaluator itself. The metric takes shape as an expression…

4. Flat retrieval breaks once a skill library hits the tens-to-hundreds range Bridge: evolution × retrieval

   2026-07-11 — This survey audits 124 papers on agent skill systems published between 2023 and 2026 (2 from 2023, 19 from 2025, 103 from 2026, cutoff May 31, 2026) and builds three shared tools for comparing them.…

5. Skill retirement flatlines past a ~45% false-pass rate — and no amount of data brings it back

   2026-07-08 — Self-evolving agents that keep accumulating skills need a curator — a mechanism that retires a skill once its observed pass rate drops to a set threshold, which is what keeps a growing library from…

6. Evolving the improver — not just the skill — accounts for all of ALFWorld's gain and half of SealQA's

   2026-07-06 — MetaSkill-Evolve doesn't stop at letting an agent revise its own skills — it lets the agent revise the machinery that does the revising. Each search branch pairs a task skill with a meta-skill :…

7. A nano model beats a flagship one, 0.7758 vs 0.7620, by editing its own harness

   2026-07-03 — Skill optimization here becomes a form of zeroth-order optimization : treat a skill document as a text parameter that execution feedback nudges, no gradient required. The authors map existing…

8. Skill evolution beats verifier-only refinement by 4 points, no refinement by 10

   2026-07-02 — ComfyClaw treats ComfyUI workflow construction as typed graph editing, not prompt rewriting. An agent inserts and connects nodes, tunes samplers, attaches LoRAs, and applies regional conditioning;…

9. Strip the tutorial video from a skill wiki, lose 9.5 points

   2026-06-28 — Most agent skill libraries come from one medium — hand-written instructions, or traces an agent left behind after solving a task once. Resource2Skill goes wider: it mines the tutorials, code…

10. Self-Authored Skills Crashed One Model's Accuracy From 78% to Under 1% Bridge: evolution × benchmarks

    2026-06-23 — EvoClawBench asks a narrow, closed-loop question: can an agent turn evidence from its own task runs into a reusable skill that makes a fresh run better, after paying the cost of authoring it? The…

11. Letting a robot invent its own practice tasks adds 20.6 points of task success

    2026-06-17 — RATs (Robotics Agent Teams) gives a robot a practice stage before it's ever handed a task — a multi-agent Code-as-Policy system built around three coordinating roles. A task proposer invents…

12. Skill selection, not RL, drives the gain — 8.3 of 10.4 points

    2026-06-15 — OpenClaw-Skill builds a searchable tree of skills instead of a flat library. Its construction method, Collective Skill Tree Search (CSTS), breaks a long-horizon agentic task into an ordered chain…

13. SkillWiki Converts Structured Sources to Governed Skills at 96% — Legacy Skill Files Lag 32 Points Behind Bridge: evolution × benchmarks

    2026-06-15 — SkillWiki treats an agent skill as a governed, versioned asset, not a stored prompt or memory snippet — the same move Wikipedia made for articles and Git made for code. The system ingests five kinds…

14. Diff a task run with the skill against one without it: 73.9% reward, zero labels

    2026-06-12 — SkillAudit evolves an agent skill with no labeled outcomes to train against. The core mechanism, paired trajectory auditing , runs the same task twice per iteration — once with the candidate skill…

15. Auto-evolved skill specs beat hand-written ones by 7 points, at twice the constraint density

    2026-06-11 — E-commerce visual assistants take one uploaded photo and have to route it to wildly different behaviors — product search, style advice, an encyclopedia lookup, a plain utility call — and each one…

16. Certainty tags turn 0–9% accuracy into 88–100%

    2026-06-10 — Notes2Skills runs lab notebooks through two stages before any of it reaches an agent. Stage 1, epistemic directive extraction (EDE), scans free-text notes and tags each statement as FACT, JUDGMENT,…

17. Curated skills add 16 points of task success — up to 51.9 in healthcare Bridge: evolution × benchmarks

    2026-06-09 — This survey traces how agent skills — packaged, reusable procedures that let LLM agents execute domain tasks — move from one-off authoring to systematic evaluation and revision. Skill-improvement…

18. LLM-Written Agent Skills Add Zero Pass-Rate Gain — Self-Diagnosis Closes Up to 67% of the Gap to Human Authors Bridge: evolution × benchmarks

    2026-06-09 — SkillAxe treats skill quality as a diagnosis problem, not a rewriting problem. It runs each task twice — once with a skill injected, once without — and scores the gap along four axes. An LLM judge…

19. No search stack — still +15.8 GAIA points over a commercial deep-research agent

    2026-06-07 — SkillHone treats agent skill maintenance as a logging problem, not an editing one. Every development step — probe run, diagnosis, candidate revision, accept or reject — gets written down as a…

20. Testing a Skill Before It's Banked Lifts WebShop Success From 72.7% to 78.1% Bridge: evolution × retrieval

    2026-06-07 — The working assumption behind skill-augmented agents — write a reusable procedural note into a retrievable skill bank , retrieve it on a later task — is that a stronger model writes a more…

21. Decomposing agent traces into workflow, semantics, and attachments beats prompted summaries by 10.5%

    2026-06-05 — Most Skill-authoring pipelines still turn a pile of agent traces into a single block of prose. RWSA (Routing–Workflow–Semantics–Attachments) is a representation built to stop that: it splits a…

22. Matching skills to scene context takes 3D-agent tool use from 39% to 78% Bridge: evolution × retrieval

    2026-06-05 — MLLM agents doing 3D spatial reasoning — object counting, distance estimation, route planning — call on external tools for detection, depth estimation, and 3D reconstruction. But they default to the…

23. Compiling skills into LoRA weights beats prompting them in-context by up to 21 points, on 64% fewer prefill tokens

    2026-06-04 — LatentSkill replaces the standard approach of stuffing a skill's text into the prompt at every decision step with a hypernetwork that compiles the skill document directly into weights. Given the…

24. OpenSkill's verifier never sees the answer key, yet agrees with it 61% of the time -- and the skills it certifies beat closed-world baselines by 8.9 points

    2026-06-04 — Most self-evolving agent frameworks quietly assume away the hardest case: deployment with nothing but a task prompt, no example skills, no successful past runs, no verifier to grade attempts.…

25. Self-graded exploration closes a 32-point reasoning gap — no labels needed

    2026-06-04 — DataCOPE builds data-analysis skills without ever seeing a labeled example. Instead of grading trajectories against ground-truth answers, it manufactures its own quality signal out of the agent's…

26. Splitting skill-writing into create, merge, and improve lifts a 27B agent up to 4.5 points -- merge drives generalization, improve drives specification, and neither alone gets you there

    2026-06-04 — Agent skills — reusable natural-language instructions that steer a model's reasoning without touching its weights — usually get built one of two ways: hand-written by a person, or extracted whole…

27. Pasting a guide into the prompt costs points; compiling it into a skill adds up to 25

    2026-06-01 — MMG2Skill formalizes guide-to-skill learning : turning human-written how-tos scraped off the web — app tutorials, wiki recipes, strategy guides — into procedures a vision-language model (VLM) agent…

28. Diagnose, don't rewrite: three revision rounds take a skill from 36% to 62% success

    2026-05-31 — Most skills start life one of two ways: an expert writes them by hand, or an LLM generates them in a single pass from a task description. Both are cold-start methods — neither one looks at what…

29. Turning a visual skill into text throws away up to 59% of its recoverable accuracy Bridge: evolution × benchmarks

    2026-05-31 — Most reusable agent skills today are pure text — instructions, reasoning traces, summarized trajectories. For GUI manipulation, layout work, counting, and other visual-centric tasks, the knowledge…

30. 215 Skills, 165 Contributors, No Fidelity Test

    2026-05-29 — COLLEAGUE.SKILL turns raw traces of a person — chat logs, work documents, email, screenshots, public research material, subtitles, plus a lightweight description — into a skill package through …

31. The right skill text for one model is the wrong one for another — model-aware rewriting recovers up to 25.8 points

    2026-05-29 — Across four Qwen3 backbones (4B, 8B, 14B, and 32B parameters), the same skill text does not transfer cleanly from one model to another. A controlled granularity study on ALFWorld makes the point…

32. The validation gate turns a 40% agent into an 89% one — skill-writing without it changes nothing

    2026-05-28 — GRASP treats agent self-improvement as edits to a small, versioned skill library , not an ever-growing memory dump. When an agent fails, the failure gets a mechanism-specific label, and a proposer…

33. Route skill treatment by task difficulty, not skill category: +13.2 points OOD success

    2026-05-27 — Skill0.5 stops treating every agent skill the same way. The usual binary — full skill externalization , where skills sit in the prompt at ever-growing token cost, versus full internalization, where…

34. MUSE-Autoskill's Self-Written Skills Beat Human-Authored Ones, 85.24% vs. 81.17%

    2026-05-26 — MUSE-Autoskill wraps an LLM agent in a five-stage skill lifecycle — creation, memory, management, evaluation, refinement — instead of treating each skill as a one-off script. New skills are built…

35. Text-based skill edits beat trained baselines by 6.7 points

    2026-05-26 — SkillGrad treats an agent's skill package — its metadata, always-loaded skill body, and conditional resource files — as a parameter to update, not a document to hand-edit. Each task run against the…

36. A skill library trained on 5x5 robot grids beats genetic search at 10x10 — on all 7 EvoGym tasks

    2026-05-25 — Auto-Robotist is a self-evolving LLM agent for evolutionary robot design that turns each generation's simulator results into a structured, natural-language skill library instead of discarding…

37. Learned Curation Halves the Skill Bank, Keeps the 9.69-Point Gain

    2026-05-25 — CODESKILL treats skill-bank curation as something a model learns, not a heuristic bolted onto a frozen agent. A management policy — trained through a three-stage curriculum with GRPO — watches…

38. Failure-only skill revision helps 1 model in 10, hurts the other 8 Bridge: evolution × benchmarks

    2026-05-22 — SkillEvolBench tests something more specific than whether skills help an agent — whether an agent can turn its own one-off experience into a skill worth keeping. The benchmark spans 180 tasks across…

39. One in Four Model-Generated Skills Backfires on the Agent Using It Bridge: evolution × benchmarks

    2026-05-22 — A full-lifecycle test bed for model-generated agent skills — experience generation, skill extraction , skill consumption — runs across five domains (ALFWorld, SpreadsheetBench, SWE-bench-Verified,…

40. One to four gated edits lift agent accuracy by up to 39 points

    2026-05-22 — SkillOpt is a text-space optimizer : it treats an agent skill document — the natural-language file prepended to a frozen model's context — as something you train rather than hand-write. A separate…

41. Three skill optimizers stall at zero on 4 of 6 tasks — Chebyshev scalarization unsticks them Bridge: evolution × benchmarks

    2026-05-19 — Agent skills aren't single prompts — they're multi-field specs bound by hard platform limits: a description field truncated at 1,024 characters for routing, an instruction body capped at 5,000…

42. Executable skills beat prompted skills 51% to 20.5% on web search

    2026-05-18 — HASP turns each stored agent skill into a small executable module called a Program Function (PF): a should_activate() check that watches the agent's current state and proposed next action, paired…

43. Four-Way Attribution Gates Every Update to a Million-Scale Skill Corpus

    2026-05-18 — SkillsVote treats the full lifecycle of an agent skill — collection, recommendation, attribution, and evolution — as one governance problem instead of four separate ones. The starting point is a…

44. Compiling a skill per task beats a static library by up to 13 Pass@1 points Bridge: evolution × retrieval

    2026-05-16 — SkillTTA treats skill-building as test-time adaptation : instead of shipping one skill per task family, it keeps a pool of past training trajectories, both successes and failures, tagged with…

45. Delete one repair action, skill-library success falls from 79.5% to 13.2%

    2026-05-13 — LLM agents that keep a library of reusable skills eventually run into a problem invisible at the single-skill level: skills get added, patched, reused in new contexts, and wired to shifting…

46. Ditch REINFORCE for a regression loss: agent-orchestration accuracy jumps +38 points in-distribution, +21 OOD

    2026-05-13 — SkillFlow trains a single orchestrator (a "Supervisor" that calls a frozen executor and a growing library of reusable skills) with Tempered Trajectory Balance (TTB), a regression-style…

47. Give a small model screenshots of its own skills and its GUI success rate more than doubles — 10.78% to 25.40%

    2026-05-13 — MMSkills packages each reusable procedure as more than a script. Alongside a compact descriptor and text procedure, it attaches state cards — when to apply the skill, when not to, and which visual…

48. Deployment Failure, Not Self-Reflection, Drives Two-Thirds of a 13-Point Skill Gain

    2026-05-11 — SkillEvolver treats skill-authoring as a closed loop, not a one-time artifact — and it never touches model weights; the update target is the skill's own prose and code. A meta-skill , loaded through…

49. Reorganize a team without touching its skills, and you lose 26 points to doing nothing at all

    2026-05-10 — SkillMAS couples two adaptation loops that agent frameworks usually run separately: skill evolution (refining a library of reusable procedures) and MAS restructuring (changing which executor…

50. Train the skill, not the model: frozen GPT-4o-mini jumps from 6% to 42% on GAIA

    2026-05-10 — Skill-R1 leaves the task-solving model frozen and trains something else entirely: a lightweight editor that rewrites the agent skill itself, generation after generation. Each round, the current…

51. 3-6x Fewer Injected Tokens, Still +5.9 Points Over ACE

    2026-05-09 — MIND-Skill turns skill-writing into a closed loop with two roles instead of one static write-up. An induction agent abstracts a successful task trajectory into a skill document; a separate…

52. A weak model with a distilled skill beats its unaided teacher — at 1,000x lower inference cost

    2026-05-09 — Most agent-skill generation pipelines get the order backwards: a model plans a solution before it ever touches the environment, and that plan becomes the skill. That's the wrong sequence — a skill is…

53. Checking what a value is for, not just whether it changed, cuts drift false alarms from 40% to zero Bridge: evolution × benchmarks

    2026-05-09 — Agent skill libraries decay quietly. An API migrates, a pinned dependency bumps a version, and the skill built against the old one keeps executing — misfiring in small ways until something downstream…

54. Self-tested skill edits beat teacher-curated skill banks by up to 9.3 points

    2026-05-09 — SkillMaster gives an LLM agent the right to edit its own skill bank instead of treating skills as content pushed in by an external teacher. After every training episode the agent enters a review…

55. Skill synthesis that checks its own work: +3 to +10 accuracy points, only 6% of skills still backfire

    2026-05-09 — SkillGen turns an agent's own rollouts into a single, human-readable skill file instead of asking a person to write one. It collects successful and failed trajectories from a base agent, then runs …

56. A trained 8B model beats Gemini-2.5-Pro at curating agent skills: 61.2% vs 50.7%

    2026-05-07 — SkillOS splits the self-evolving agent into two roles: a frozen executor that solves tasks and retrieves skills via BM25 retrieval , and a separately trained curator that edits an external …

57. One shared reward signal pushes skill-augmented agents to 97.5% success, beating every tested baseline — skill-augmented or skill-free — by up to 20 points

    2026-05-07 — Skill1 trains one policy, via reinforcement learning, to run all three stages of a skill library lifecycle end to end: it writes a text query and re-ranks candidates to pick a skill, solves the…

58. No labels, no verifier: self-play skill-writing lifts GPT-4.1 from 11.1% to 16.5% on context learning

    2026-04-30 — Ctx2Skill builds a skill file straight from a raw, unlabeled context using a five-role self-play loop — no human-written skills, no external verifier. Each round, a Challenger agent writes probing…

59. Skill pass rates jump from 33.6% to 88.0% — with zero gradient steps

    2026-04-30 — Skills-Coach puts a skill through a four-stage self-improvement loop. First, a task generator reads the skill's own specification and writes a test suite of standard, advanced, and boundary-case…

60. An 8B Model Beats 4 Frontier LLMs by 25%+ — By Mining Its Own Skill Bank

    2026-04-22 — COSPLAY runs two roles inside one system: a decision agent that plays a game by drawing on a skill bank , and a skill-pipeline agent that mines the decision agent's own unlabeled rollouts,…

61. Self-repairing skills add 8.4 points — raw context alone scores worse than no skills at all Bridge: evolution × benchmarks

    2026-04-19 — SkillFlow introduces 166 tasks across 20 workflow families spanning five domains — finance and economics, operations and supply chain, healthcare and life sciences, governance and strategy, and data…

62. 0 of 9 add-skill edits improved pass rate; cost fell up to 31.7% vs. static bundles

    2026-04-10 — SkillMOO treats a coding agent's agent skill bundle as something to search over, not a fixed asset written once. An LLM optimizer agent proposes edits — pruning a skill, substituting one,…

63. A skill improves every time anyone uses it, not just its author

    2026-04-09 — Most agent skills ship once and sit still. The same tool-usage mistake, the same brittle workaround, the same failure mode gets rediscovered independently by every user who hits it, because nothing…

64. Three rounds of automated repair add up to 12pp of consistency — and beat a mature production system by 13.76pp

    2026-04-09 — SkillForge builds and maintains agent skills for cloud technical-support agents at Alibaba Cloud, closing a loop that most skill-authoring setups leave open. A Domain-Contextualized Skill Creator…

65. A skill library built by GLM-4.6 boosts GPT-4.1's task success by 10 points

    2026-04-06 — SkillX turns an agent's raw trajectories into a three-tier skill knowledge base : strategic plans that sequence a task end to end, functional skills that package a tool-based subroutine, and …

66. 71% of Mined Scientific Skills Don't Exist in Any Skill Marketplace

    2026-04-05 — SkillFoundry turns scientific papers, repos, notebooks, and docs into a maintained skill library, not a one-time scrape. It maps a target domain as a domain knowledge tree , then mines the branches…

67. Self-evolved skills beat human-curated ones by 17.6 points on SkillsBench

    2026-04-02 — CoEvoSkills leans on a distinction Anthropic drew when it introduced agent skills : a tool is a single self-contained function, but a skill is a structured bundle of interdependent instructions,…

68. Skills Trained Into the Model Beat Runtime-Loaded Skills by Up to 10.1 Points

    2026-04-02 — SKILL0 trains an agent to stop needing its own skill files. That's skill internalization — procedural knowledge migrates from context into weights instead of getting reloaded every time the agent…

69. A skill grown from a 35B model's traces adds 57.65 points to a 122B agent on table QA it's never seen

    2026-03-26 — Trace2Skill turns a pile of raw agent execution traces into one consolidated skill file, not a growing memory store, not an order-dependent chain of edits. A frozen agent runs a batch of tasks first…

70. Self-generated web skills score below no skill at all — contract-based repair triples success to 28.1%

    2026-03-20 — ContractSkill converts a draft web-agent skill into a contract artifact — a structured record of the goal, preconditions, an ordered list of steps (each carrying a selector, action, optional…

71. A self-writing skill library lifts Humanity's Last Exam accuracy 116% — with zero weight updates

    2026-03-19 — Memento-Skills flips who designs the agent: instead of a human hand-building a task-specific system, a generalist LLM builds and refines its own sub-agents through experience — an agent-designing…

72. Self-authored skills buy up to 2.9 points on out-of-distribution math for a 3% token tax

    2026-03-17 — ARISE trains a single policy to play two roles inside the same reinforcement-learning loop: a Skills Manager and a Worker. Before attempting a problem, the Manager scores every skill sitting in its…

73. Splitting agent memory into skills and experience gains up to 11 points of task success Bridge: evolution × retrieval

    2026-03-12 — XSkill splits a multimodal agent's reusable memory into two separate structures instead of one. A skill library holds stable, task-level workflows and tool templates in Markdown. A companion …

74. Letting computer-use agents explore GUIs first turns near-zero success into 25–37%

    2026-03-09 — General-purpose computer-use agents train on demonstrations from roughly 100 digital environments, then get thrown at whatever unfamiliar interface a real task requires — and the skills don't…

75. Skill folders, not weights: +12.1 points on noisy-retrieval QA

    2026-03-03 — EvoSkill discovers and edits agent skills — structured folders of instructions and helper scripts — by watching a coding agent fail, then testing fixes at the skill level instead of touching…

76. 1,858 reusable skills mined from 22,511 real chats — zero weight updates

    2026-03-01 — AutoSkill is a training-free plug-in layer that turns repeated user requests into versioned skill artifacts instead of treating them as disposable chat history. Two loops run in tandem. A serving…

77. Skill abstraction beats raw-trajectory memory by up to 25 points

    2026-02-09 — SkillRL splits an LLM agent's raw rollouts into two streams and distills each differently: successful trajectories become demonstrations, failed ones get compressed by a teacher model (OpenAI o3)…

78. Offline Skill-Patching Beats DPT-WToM 15.6x on 27% Fewer Tokens

    2026-02-04 — CoWork-X splits agent cognition into two clocks. A Skill-Agent handles every in-episode decision by retrieving from a structured skill library built as hierarchical task networks (HTN) — state…

79. 92.5% Skill Reuse From 816 Tokens

    2026-02-02 — Agents built on chain-of-thought reasoning treat every episode as new — even when the underlying subtask has been solved dozens of times, the reasoning chain gets thrown away the moment the episode…

80. A learned memory-skill bank beats fixed insert/update/delete baselines by 5-7 points, on 6x fewer LLM calls

    2026-02-02 — MemSkill treats the four hard-coded memory operations — insert, update, delete, skip — that most LLM agent memory systems run on as a starting point, not the final word. It organizes them into a…

81. Evolve the context-engineering skill itself, not just the context — up to 54% better than fixed pipelines like ACE and GEPA

    2026-01-29 — MCE reframes context engineering as a two-level search rather than a fixed pipeline. Instead of hand-designing the harness that decides how context gets built — the generation-reflection-curation…

82. 35 iterations, 6 of 6 runs: programmatic skill graphs triple Voyager's hit rate on Minecraft's diamond-tool climb

    2026-01-07 — Evolving Programmatic Skill Networks builds agent skills as executable, typed programs — not prompts or embeddings — arranged in a directed graph that keeps growing and rewriting itself as the agent…

83. Reward reuse, not just creation: +8.9 points scenario completion, -59% tokens

    2025-12-18 — SAGE trains an LLM agent to build and reuse a library of executable skills by extending GRPO (Group Relative Policy Optimization) with two additions. Sequential Rollout chains two similar tasks…

84. A three-tier skill hierarchy — not a bigger skill library — drives a 79.6% jump on long-horizon GUI tasks

    2025-06-12 — MLLM-based GUI agents that plan step-by-step from screenshots run out of usable knowledge once a task stretches across many steps and several apps. The fix: a Hierarchical Multimodal Skills (HMS)…

85. Pre-simulated skill libraries beat live code generation 44.7% to 28.7% -- at zero runtime cost

    2025-04-29 — This framework skips runtime code generation entirely: it pre-builds and validates a library of Adobe Illustrator scripts offline, then serves them by lookup instead of writing ExtendScript on the…

86. Self-written APIs make web agents 32-40% better — weaker models gain even more from them

    2025-04-09 — SkillWeaver turns a web agent loose on a new site and runs a tight loop: the agent discovers candidate skills, executes them repeatedly for practice, then distills the successful attempts into…

87. Verify a skill before you trust it — that's an 11% jump in success rate Bridge: evolution × security

    2025-04-09 — Agent skill induction ( ASI ) lets a web-browsing agent mine its own successful task runs for reusable Python functions, instead of jotting down prose notes. Once a task finishes, the agent drafts…

88. Remove the curriculum and Voyager's item discovery collapses 93%

    2023-05-25 — Voyager pairs GPT-4 with three components instead of one clever prompt. An automatic curriculum proposes the next Minecraft objective by reading the agent's live state against its history of…