Skip to content

Latest commit

 

History

History
48 lines (27 loc) · 1.95 KB

File metadata and controls

48 lines (27 loc) · 1.95 KB

Baselines

SkillLearnBench provides four skill-generation baselines, each representing a different strategy for how an agent acquires reusable skill documents.


b1 — One-Shot

The agent is given a single task instance and asked to write 1–5 SKILL.md files before solving the task. No feedback loop is involved. The skills are written in a single pass, and the agent then attempts the task using those skills.

This is the simplest baseline and serves as the primary reference point for more iterative methods.


b2 — Self-Feedback

After an initial skill-generation pass, the agent reflects on its own output and revises the skills over multiple rounds. No external signal is used — all feedback comes from the agent's own self-critique.

The number of rounds is configured via max_rounds in method.toml (default: 2).


b3 — Teacher-Feedback

A two-agent setup:

  • Teacher (external LLM): generates an initial skill set and provides feedback when the student fails
  • Student (evaluation agent): attempts the task using the provided skills; on failure, the teacher's feedback is incorporated into a revised skill set

This tests whether externally guided revision produces higher-quality skills than self-reflection alone.


b4 — Skill Creator

Identical to b1 in structure (single-pass, no feedback), but the agent is additionally given Claude's official skill-creator format specification. This tests whether adherence to a formal skill schema improves downstream task performance.


Config Naming

Generated skill configs are named <method>-<model>, e.g. b1-one-shot-claude-sonnet-4-6. This is the directory name used in output/skill_generation_results/ and passed to --skill-path during evaluation.

Special configs:

Config Description
human_authored Ground-truth skills written by task authors
no_skill No skills injected (null baseline); pass none to --skill-path