Skip to content
Back to search

CoEvoSkills

CoEvoSkills: Self-Evolving Agent Skills via Co-Evolutionary Verification — COLM 2026

Safety scan

A static text scan, not a guarantee. Select the badge to see what was scanned and every finding.

Checked today
Badge for your README
findsafeskills safety rating

Markdown

[![findsafeskills safety rating](/api/listings/zhang-henry--coevoskills/badge.svg)](/listings/zhang-henry--coevoskills)

HTML

<a href="/listings/zhang-henry--coevoskills"><img src="/api/listings/zhang-henry--coevoskills/badge.svg" alt="findsafeskills safety rating" /></a>

Install

i

Install instructions from the README

As written by the repo's author. findsafeskills has not verified these steps.

1. Install

Requirements:

- Linux or macOS - Python 3.12 - Docker - [`uv`](https://docs.astral.sh/uv/) - Credentials for the model provider you want to use

git clone https://github.com/Zhang-Henry/CoEvoSkills.git
cd CoEvoSkills

uv sync --locked
cp .env.example .env
docker info

Fill in one Agent-provider block in `.env`. Login commands and environment variables for Claude Code subscription, Anthropic API, Vertex AI, Amazon Bedrock, Gemini, OpenAI/Codex, and Azure OpenAI are documented in [`docs/provider_setup.md`](docs/provider_setup.md).

Some benchmark tasks also declare task-specific API variables independently of the Agent provider. Before a full 85-task run, follow the task-credential check in the provider setup guide. In particular, `pg-essay-to-audiobook` needs `OPENAI_API_KEY` for its canonical verifier even when the Agent uses Claude.

2. Run one released Skill

The following stages one finalized Skill with its background document removed, then gives the task and Skill to a fresh Claude Code Agent:

export RUN_ID=skill-only-smoke

uv run python scripts/prepare_tasks.py \
  --condition skill-only \
  --skill-set validated \
  --tasks 3d-scan-calc

MODEL=anthropic/claude-opus-4-6 \
AGENT=claude-code-skill-only \
MAX_PARALLEL=1 \
  scripts/run_condition.sh skill-only --only-tasks 3d-scan-calc

Set `--skill-set all` to stage all released Skills, including candidates. Use `--skill-set validated` for the finalized subset reported by the release. This status check also applies when `--tasks` names an explicit subset, so a candidate cannot be staged accidentally as a validated Skill.

The preparation command physically removes the background-document directory. The fresh Agent receives only the task and the released Skill. Choose the Agent adapter and model route below:

| Agent route | `AGENT` | Example `MODEL` | |---|---|---| | Claude Code (subscription, Anthropic API, Vertex AI, or Bedrock) | `claude-code-skill-only` | `anthropic/claude-opus-4-6` | | Codex with ChatGPT subscription | `codex-subscription` | `openai/gpt-5.4` | | Codex with OpenAI API | `codex-skill-only` | `openai/<available-codex-model>` | | Other LiteLLM-backed Agents (exploratory transfer) | `terminus-2` | `<provider>/<model-id>` |

Strict release reproduction uses `claude-code-skill-only`, `codex-skill-only`, or `codex-subscription`. These adapters enforce the no-Doc barrier, make the released Skill read-only, and attest its digest before and after the run. The generic `terminus-2` route is useful for exploratory cross-model transfer, but it is not treated as equivalent strict attestation.

For example, to stage the same released Skill in an isolated Codex subscription run:

codex login
codex login status

export CODEX_CHATGPT_AUTH_FILE="$HOME/.codex/auth.json"
export CODEX_CLI_BINARY="$(command -v codex)"
export CODEX_REASONING_EFFORT="low"
export RUN_ID=codex-skill-only-smoke

uv run python scripts/prepare_tasks.py \
  --condition skill-only \
  --skill-set validated \
  --tasks 3d-scan-calc

MODEL=openai/gpt-5.4 \
AGENT=codex-subscription \
MAX_PARALLEL=1 \
  scripts/run_condition.sh skill-only --only-tasks 3d-scan-calc

For an OpenAI API-backed Codex run, export `OPENAI_API_KEY`, replace `AGENT` with `codex-skill-only`, and set `MODEL` to a Codex model available to that API project. For Gemini, Azure OpenAI, or another LiteLLM-supported model, export that provider's credentials and run with `AGENT=terminus-2` and its provider-qualified model name. Exact authentication examples are in [`docs/provider_setup.md`](docs/provider_setup.md).

When adding a new native Agent adapter, preserve the evaluation contract: load the released package from `/app/environment/skills/*/SKILL.md`, expose no background document, keep the Skill immutable during the run, and execute the canonical task verifier after the Agent exits.

3. Evolve a Skill from scratch

What was scanned

The files the safety scan read, as written by the repository's author and shown as plain text. Lines with a finding are marked.

README.md · README · 469 lines

View this file on GitHub (read at the repository's current version: the scan did not record which commit it read)

  1. <div align="center">
  2. # 🧬 CoEvoSkills
  3. ### Self-Evolving Agent Skills via Co-Evolutionary Verification
  4. **Accepted at the Conference on Language Modeling (COLM) 2026.**
  5. [![Paper](https://img.shields.io/badge/Paper-PDF-b31b1b.svg)](https://arxiv.org/pdf/2604.01687)
  6. [![arXiv](https://img.shields.io/badge/arXiv-2604.01687-b31b1b.svg)](https://arxiv.org/abs/2604.01687)
  7. [![Website](https://img.shields.io/badge/Website-Project%20Page-4c8bf5.svg)](https://zhang-henry.github.io/CoEvoSkills/)
  8. [![Benchmark](https://img.shields.io/badge/Benchmark-SkillsBench-2ea44f.svg)](https://github.com/benchflow-ai/skillsbench)
  9. <em>A self-evolving framework that lets LLM agents construct reusable,<br>
  10. multi-file Skill packages without access to protected ground-truth answers.</em>
  11. If you find CoEvoSkills useful, please consider [citing our work](#citation).
  12. </div>
  13. ---
  14. ## Citation
  15. If you find this work useful, please consider citing:
  16. ```bibtex
  17. @misc{zhang2026coevoskillsselfevolvingagentskills,
  18. title={CoEvoSkills: Self-Evolving Agent Skills via Co-Evolutionary Verification},
  19. author={Hanrong Zhang and Shicheng Fan and Henry Peng Zou and Yankai Chen and Zhenting Wang and Jiayu Zhou and Chengze Li and Wei-Chieh Huang and Yifei Yao and Kening Zheng and Xue Liu and Xiaoxiao Li and Philip S. Yu},
  20. year={2026},
  21. eprint={2604.01687},
  22. archivePrefix={arXiv},
  23. primaryClass={cs.AI},
  24. url={https://arxiv.org/abs/2604.01687},
  25. }
  26. ```
  27. ## Overview
  28. Agent Skills are structured folders of instructions, scripts, and resources
  29. that help an Agent perform specialized workflows. CoEvoSkills learns these
  30. packages through a repeated **generate → verify → refine** loop.
  31. <div align="center">
  32. <img src="assets/tool_skill_diff.png" alt="Tool versus Skill" width="82%"/>
  33. <br/>
  34. <sub><b>Figure 1.</b> A tool exposes one operation; a Skill packages a reusable workflow and its supporting resources.</sub>
  35. </div>
  36. CoEvoSkills couples two evolving components:
  37. - **Skill Generator** — creates and revises a schema-valid, multi-file Skill.
  38. - **Surrogate Verifier** — independently builds executable checks and returns
  39. dense diagnostic feedback without seeing protected benchmark answers.
  40. A fresh ground-truth Agent periodically tests the current Skill. It receives
  41. the task and Skill only—never the background document or the evolution
  42. conversation. A canonical reward of `1.0` ends evolution; otherwise the opaque
  43. failure feedback starts the next revision.
  44. <div align="center">
  45. <img src="assets/framework.png" alt="CoEvoSkills framework" width="94%"/>
  46. <br/>
  47. <sub><b>Figure 2.</b> Co-evolution of the Skill Generator and Surrogate Verifier, with isolated ground-truth testing.</sub>
  48. </div>
  49. ## Highlights
  50. - **Structured Skills:** evolves executable, multi-file Agent Skill packages
  51. rather than a single prompt or function.
  52. - **Information isolation:** background documents contain domain knowledge,
  53. while protected answers and verifier internals remain unavailable during
  54. evolution.
  55. - **Independent transfer test:** every finalized Skill is evaluated by a fresh
  56. Agent in Skill-only mode.
  57. - **Cross-model support:** the framework supports Claude, OpenAI/Codex, Gemini,
  58. Vertex AI, Amazon Bedrock, Azure OpenAI, and LiteLLM-compatible routes.
  59. ## Repository contents
  60. Everything needed for the 85-task release is included in this repository. The
  61. three generated-data locations have deliberately different lifecycles:
  62. `artifacts/` is the immutable release bundle, `workspaces/` contains mutable
  63. run-specific task copies, and `outputs/` contains run-specific evidence.
  64. | Path | What it contains |
  65. |---|---|
  66. | `tasks/` | The exact 85 benchmark task packages used by this release: instructions, runtime inputs, Docker environments, and canonical verifiers. |
  67. | `artifacts/background_docs/` | One answer-free domain reference set per task in the bundled release. These documents are available during evolution only. |
  68. | `artifacts/skills/` | The immutable, schema-valid Skill package bundled for each task. Do not use this directory as an evolution workspace. |
  69. | `artifacts/skill_status.tsv` | Whether each bundled Skill obtained canonical reward `1.0` in a fresh Skill-only run or remains a candidate. |
  70. | `artifacts/evaluations/release-skill-only-results.tsv` | The canonical, trace-free result table for the current finalized Skills. |
  71. | `libs/terminus_agent/` | Skill Generator, Surrogate Verifier, and fresh Skill-only Agent implementations. |
  72. | `meta_skills/skill-creator/` | The Skill package format and authoring rules supplied to the evolution Agent. |
  73. | `scripts/prepare_tasks.py` | Creates an isolated evolution or Skill-only copy of selected tasks. The bundled originals are not modified. |
  74. | `scripts/run_condition.sh` | Runs a prepared condition and writes all generated output under one run directory. |
  75. | `run_exp.py` | Docker/Harbor task scheduler and result collector. |
  76. | `workspaces/<RUN_ID>/` | Mutable, isolated task copies. Evolution writes new Skills and verifier state here. Generated locally and ignored by Git. |
  77. | `outputs/<RUN_ID>/` | Immutable snapshots, task sandboxes, logs, and summaries from one run. Generated locally and ignored by Git. |
  78. The task packages are adapted from SkillsBench commit
  79. `a7028dfd37cfff86acaf248656cdbd9ad0179592`. CoEvoSkills-specific changes are
  80. already present in `tasks/`; their attribution is recorded in `NOTICE`.
  81. The two large archives under `tasks/sec-financial-report/environment/` are the
  82. SEC Form 13F bulk datasets used by that benchmark task. They contain the Q2 and
  83. Q3 2025 filing tables, including `COVERPAGE.tsv`, `INFOTABLE.tsv`, and the
  84. submission and summary tables. They are task inputs—not generated logs or model
  85. outputs.
  86. ## Quick start
  87. ### 1. Install
  88. Requirements:
  89. - Linux or macOS
  90. - Python 3.12
  91. - Docker
  92. - [`uv`](https://docs.astral.sh/uv/)
  93. - Credentials for the model provider you want to use
  94. ```bash
  95. git clone https://github.com/Zhang-Henry/CoEvoSkills.git
  96. cd CoEvoSkills
  97. uv sync --locked
  98. cp .env.example .env
  99. docker info
  100. ```
  101. Fill in one Agent-provider block in `.env`. Login commands and environment variables
  102. for Claude Code subscription, Anthropic API, Vertex AI, Amazon Bedrock, Gemini,
  103. OpenAI/Codex, and Azure OpenAI are documented in
  104. [`docs/provider_setup.md`](docs/provider_setup.md).
  105. Some benchmark tasks also declare task-specific API variables independently of
  106. the Agent provider. Before a full 85-task run, follow the task-credential check
  107. in the provider setup guide. In particular, `pg-essay-to-audiobook` needs
  108. `OPENAI_API_KEY` for its canonical verifier even when the Agent uses Claude.
  109. ### 2. Run one released Skill
  110. The following stages one finalized Skill with its background document removed,
  111. then gives the task and Skill to a fresh Claude Code Agent:
  112. ```bash
  113. export RUN_ID=skill-only-smoke
  114. uv run python scripts/prepare_tasks.py \
  115. --condition skill-only \
  116. --skill-set validated \
  117. --tasks 3d-scan-calc
  118. MODEL=anthropic/claude-opus-4-6 \
  119. AGENT=claude-code-skill-only \
  120. MAX_PARALLEL=1 \
  121. scripts/run_condition.sh skill-only --only-tasks 3d-scan-calc
  122. ```
  123. Set `--skill-set all` to stage all released Skills, including candidates. Use
  124. `--skill-set validated` for the finalized subset reported by the release. This
  125. status check also applies when `--tasks` names an explicit subset, so a candidate
  126. cannot be staged accidentally as a validated Skill.
  127. The preparation command physically removes the background-document directory.
  128. The fresh Agent receives only the task and the released Skill. Choose the Agent
  129. adapter and model route below:
  130. | Agent route | `AGENT` | Example `MODEL` |
  131. |---|---|---|
  132. | Claude Code (subscription, Anthropic API, Vertex AI, or Bedrock) | `claude-code-skill-only` | `anthropic/claude-opus-4-6` |
  133. | Codex with ChatGPT subscription | `codex-subscription` | `openai/gpt-5.4` |
  134. | Codex with OpenAI API | `codex-skill-only` | `openai/<available-codex-model>` |
  135. | Other LiteLLM-backed Agents (exploratory transfer) | `terminus-2` | `<provider>/<model-id>` |
  136. Strict release reproduction uses `claude-code-skill-only`,
  137. `codex-skill-only`, or `codex-subscription`. These adapters enforce the
  138. no-Doc barrier, make the released Skill read-only, and attest its digest before
  139. and after the run. The generic `terminus-2` route is useful for exploratory
  140. cross-model transfer, but it is not treated as equivalent strict attestation.
  141. For example, to stage the same released Skill in an isolated Codex subscription
  142. run:
  143. ```bash
  144. codex login
  145. codex login status
  146. export CODEX_CHATGPT_AUTH_FILE="$HOME/.codex/auth.json"
  147. export CODEX_CLI_BINARY="$(command -v codex)"
  148. export CODEX_REASONING_EFFORT="low"
  149. export RUN_ID=codex-skill-only-smoke
  150. uv run python scripts/prepare_tasks.py \
  151. --condition skill-only \
  152. --skill-set validated \
  153. --tasks 3d-scan-calc
  154. MODEL=openai/gpt-5.4 \
  155. AGENT=codex-subscription \
  156. MAX_PARALLEL=1 \
  157. scripts/run_condition.sh skill-only --only-tasks 3d-scan-calc
  158. ```
  159. For an OpenAI API-backed Codex run, export `OPENAI_API_KEY`, replace `AGENT`
  160. with `codex-skill-only`, and set `MODEL` to a Codex model available to that API
  161. project. For Gemini, Azure OpenAI, or another LiteLLM-supported model, export
  162. that provider's credentials and run with `AGENT=terminus-2` and its
  163. provider-qualified model name. Exact authentication examples are in
  164. [`docs/provider_setup.md`](docs/provider_setup.md).
  165. When adding a new native Agent adapter, preserve the evaluation contract: load
  166. the released package from `/app/environment/skills/*/SKILL.md`, expose no
  167. background document, keep the Skill immutable during the run, and execute the
  168. canonical task verifier after the Agent exits.
  169. ### 3. Evolve a Skill from scratch
  170. Evolution starts with the task's background document and an empty Skill
  171. directory. Each ground-truth intervention launches a fresh Skill-only Agent.
  172. ```bash
  173. export RUN_ID=evolution-canary
  174. uv run python scripts/prepare_tasks.py \
  175. --condition evolution \
  176. --tasks 3d-scan-calc
  177. MODEL=anthropic/claude-opus-4-6 \
  178. AGENT=terminus-2-evolution \
  179. VERIFIER_MODEL=anthropic/claude-opus-4-6 \
  180. GT_ORACLE_MODEL=anthropic/claude-opus-4-6 \
  181. GT_ORACLE_AGENT=claude-code-skill-only \
  182. MAX_GT_ITERATIONS=5 \
  183. MAX_PARALLEL=1 \
  184. scripts/run_condition.sh evolution --only-tasks 3d-scan-calc
  185. ```
  186. Start with one task. For the full benchmark, choose a new `RUN_ID`, rerun the
  187. preparation command without `--tasks`, then run without `--only-tasks` and
  188. increase `MAX_PARALLEL`. Reusing an existing `RUN_ID` requires `--force` on the
  189. preparation command and intentionally replaces that mutable workspace.
  190. To continue an interrupted evolution without mixing its old and new evidence,
  191. choose a new output `RUN_ID`, point `TASKS_DIR` to the existing workspace, and
  192. set `EVOLUTION_MODE=continue`:
  193. ```bash
  194. export RUN_ID=evolution-canary-continuation
  195. export TASKS_DIR="$PWD/workspaces/evolution-canary/evolution"
  196. export EVOLUTION_MODE=continue
  197. MODEL=anthropic/claude-opus-4-6 \
  198. AGENT=terminus-2-evolution \
  199. VERIFIER_MODEL=anthropic/claude-opus-4-6 \
  200. GT_ORACLE_MODEL=anthropic/claude-opus-4-6 \
  201. GT_ORACLE_AGENT=claude-code-skill-only \
  202. scripts/run_condition.sh evolution --only-tasks 3d-scan-calc
  203. ```
  204. ### 4. Reproduce the 62 finalized Skill-only evaluations
  205. `--skill-set validated` resolves to the 62
  206. `validated_skill_only_full_score` rows in `artifacts/skill_status.tsv` (62/85,
  207. or 72.9%). The remaining 23 Agent-generated Skills are retained as candidates
  208. and are excluded from the finalized score.
  209. ```bash
  210. export RUN_ID=skill-only-reproduction
  211. uv run python scripts/prepare_tasks.py \
  212. --condition skill-only \
  213. --skill-set validated
  214. MODEL=anthropic/claude-opus-4-6 \
  215. AGENT=claude-code-skill-only \
  216. MAX_PARALLEL=10 \
  217. scripts/run_condition.sh skill-only
  218. ```
  219. Each run is isolated under `outputs/<RUN_ID>/`:
  220. ```text
  221. outputs/<RUN_ID>/
  222. run.log Combined dispatcher output
  223. jobs/ Per-task Agent and Docker artifacts
  224. results/ CSV summaries
  225. ```
  226. Create a trace-free aggregate for the run with:
  227. ```bash
  228. uv run python scripts/summarize_skill_only_results.py \
  229. --results-root "outputs/$RUN_ID/results" \
  230. --jobs-root "outputs/$RUN_ID/jobs" \
  231. --status-file artifacts/skill_status.tsv \
  232. --output-tsv "outputs/$RUN_ID/summary.tsv" \
  233. --output-summary "outputs/$RUN_ID/summary.json" \
  234. --model anthropic/claude-opus-4-6 \
  235. --agent claude-code-skill-only
  236. ```
  237. ### 5. Run tasks in batches
  238. To evaluate all 85 bundled Skills, including the candidate packages, prepare
  239. the complete Skill set and omit `--only-tasks` from the runner:
  240. ```bash
  241. export RUN_ID=skill-only-all-85
  242. uv run python scripts/prepare_tasks.py \
  243. --condition skill-only \
  244. --skill-set all
  245. MODEL=anthropic/claude-opus-4-6 \
  246. AGENT=claude-code-skill-only \
  247. MAX_PARALLEL=10 \
  248. scripts/run_condition.sh skill-only
  249. ```
  250. To evolve Skills from empty Skill directories for all 85 tasks, prepare the
  251. evolution condition without `--tasks` and run it without `--only-tasks`:
  252. ```bash
  253. export RUN_ID=evolution-all-85
  254. uv run python scripts/prepare_tasks.py \
  255. --condition evolution
  256. MODEL=anthropic/claude-opus-4-6 \
  257. AGENT=terminus-2-evolution \
  258. VERIFIER_MODEL=anthropic/claude-opus-4-6 \
  259. GT_ORACLE_MODEL=anthropic/claude-opus-4-6 \
  260. GT_ORACLE_AGENT=claude-code-skill-only \
  261. MAX_GT_ITERATIONS=5 \
  262. MAX_PARALLEL=4 \
  263. scripts/run_condition.sh evolution
  264. ```
  265. For a smaller batch, pass the same comma-separated task list to both steps:
  266. ```bash
  267. export RUN_ID=selected-tasks-r1
  268. export TASK_LIST=3d-scan-calc,protein-expression-analysis,dialogue-parser
  269. uv run python scripts/prepare_tasks.py \
  270. --condition skill-only \
  271. --skill-set all \
  272. --tasks "$TASK_LIST"
  273. MODEL=anthropic/claude-opus-4-6 \
  274. AGENT=claude-code-skill-only \
  275. MAX_PARALLEL=3 \
  276. scripts/run_condition.sh skill-only --only-tasks "$TASK_LIST"
  277. ```
  278. Before a benchmark-wide run, configure the selected provider and every
  279. task-specific credential described in
  280. [`docs/provider_setup.md`](docs/provider_setup.md). Start with a small canary,
  281. then raise `MAX_PARALLEL` to fit the host's CPU, memory, Docker capacity, and
  282. provider quota.
  283. Evolution also leaves a mutable working copy at
  284. `workspaces/<RUN_ID>/evolution/<task>/environment/skills/evo-*/SKILL.md`. The
  285. run-specific snapshot under `outputs/<RUN_ID>/results/.../skills/` is the safer
  286. artifact to archive or compare because a later continuation may change the
  287. workspace copy.
  288. Use a new `RUN_ID` and rerun `prepare_tasks.py` for every independent replicate.
  289. Raw model output stays local and is not part of the release. The runner also
  290. locks mutable evolution workspaces, so two processes cannot silently evolve
  291. the same task copy at once.
  292. ## Extend the framework
  293. ### Add a task before it has a Skill
  294. A new evolution task needs a SkillsBench-compatible task package and an
  295. answer-free background document; it does **not** need a placeholder Skill or a
  296. row in `skill_status.tsv`:
  297. ```text
  298. my_tasks/new-task/task.toml
  299. my_tasks/new-task/environment/
  300. my_release/background_docs/new-task/domain-reference.md
  301. ```
  302. ```bash
  303. export RUN_ID=new-task-evolution-r1
  304. uv run python scripts/prepare_tasks.py \
  305. --condition evolution \
  306. --base my_tasks \
  307. --release-root my_release \
  308. --tasks new-task
  309. MODEL=<provider>/<model-id> \
  310. AGENT=terminus-2-evolution \
  311. VERIFIER_MODEL=<provider>/<model-id> \
  312. GT_ORACLE_MODEL=<provider>/<model-id> \
  313. GT_ORACLE_AGENT=<skill-only-agent> \
  314. MAX_GT_ITERATIONS=5 \
  315. MAX_PARALLEL=1 \
  316. scripts/run_condition.sh evolution --only-tasks new-task
  317. ```
  318. To evaluate a fixed Skill, add `skills/new-task/<skill-name>/SKILL.md` and a
  319. `skill_status.tsv` row under the selected release root, then prepare the
  320. `skill-only` condition. A future version can live under `releases/v2/` and be
  321. selected with `--release-root releases/v2`; no framework code needs to change.
  322. ### Add an Agent adapter
  323. API-backed models supported by LiteLLM only require a new model string. A new
  324. native Harbor Agent can be registered at runtime instead of editing the built-in
  325. registry:
  326. ```bash
  327. export RUN_ID=custom-agent-smoke
  328. uv run python scripts/prepare_tasks.py \
  329. --condition skill-only \
  330. --skill-set validated \
  331. --tasks 3d-scan-calc
  332. MODEL=<provider>/<model-id> \
  333. AGENT=my-agent \
  334. scripts/run_condition.sh skill-only \
  335. --only-tasks 3d-scan-calc \
  336. --agent-import-path my-agent=my_package.agents:MyAgent
  337. ```
  338. The same registration can be used as `GT_ORACLE_AGENT=my-agent` during
  339. evolution. The class must accept Harbor's `logs_dir` and `model_name`
  340. constructor arguments and be importable in the runner environment.
  341. ## Main results
  342. <div align="center">
  343. <img src="assets/main_results.png" alt="Main SkillsBench results" width="88%"/>
  344. <br/>
  345. <sub><b>Figure 3.</b> Main SkillsBench comparison reported in the paper.</sub>
  346. </div>
  347. ### Cross-model transfer
  348. <div align="center">
  349. <img src="assets/cross_model_transfer.png" alt="Cross-model transfer" width="88%"/>
  350. <br/>
  351. <sub><b>Figure 4.</b> Skills transfer to additional Agent backbones without retraining.</sub>
  352. </div>
  353. ### Per-domain performance
  354. <div align="center">
  355. <img src="assets/domain_breakdown.png" alt="Per-domain results" width="88%"/>
  356. <br/>
  357. <sub><b>Figure 5.</b> Pass rate across the 11 SkillsBench domains.</sub>
  358. </div>
  359. ### Evolution trajectory
  360. <div align="center">
  361. <img src="assets/evolution_trajectory_preview.png" alt="Evolution trajectory" width="88%"/>
  362. <br/>
  363. <sub><b>Figure 6.</b> Performance over successive co-evolution iterations.</sub>
  364. </div>

Ratings and comments

Sign in to rate this skill.

Sign in

Free to use. If it’s helped, pass it on.