wuyang
69 lines
1.7 KiB
Markdown
69 lines
1.7 KiB
Markdown
# Industry: Microsoft Research - SkillOpt
|
|
|
|
---
|
|
type: industry
|
|
company: Microsoft
|
|
team: Microsoft Research
|
|
title: "SkillOpt: Agent skills as trainable parameters"
|
|
url: https://www.microsoft.com/en-us/research/blog/skillopt-agent-skills-as-trainable-parameters/
|
|
source_name: Microsoft Research
|
|
source_type: research-blog
|
|
source_quality: official
|
|
published_at: 2026-06-30
|
|
collected_at: 2026-07-08
|
|
status: analyzed
|
|
topics:
|
|
- agent
|
|
- prompt-optimization
|
|
- skill-learning
|
|
implementation_signals:
|
|
- skill-files
|
|
- validation-gating
|
|
- bounded-edits
|
|
- eval-loop
|
|
product_area:
|
|
- agent-framework
|
|
models:
|
|
-
|
|
tools:
|
|
-
|
|
benchmarks:
|
|
- six-benchmark-evaluation
|
|
related_papers:
|
|
-
|
|
related_jobs:
|
|
- 2026-07-08-deepseek-agent-hiring-wave-beijing-hangzhou
|
|
related_experiments:
|
|
-
|
|
related_projects:
|
|
-
|
|
evidence_level: high
|
|
relevance: high
|
|
---
|
|
|
|
## One-line Takeaway
|
|
|
|
Agent skill/prompt 文件可以被当成外部可训练参数，用评估循环优化，而不是靠人工随手改。
|
|
|
|
## What They Built or Claimed
|
|
|
|
Microsoft Research 提出 SkillOpt，将 skill editing 转成训练过程，在不改模型权重的情况下提升 Agent 行为可靠性。
|
|
|
|
## Technical Signals
|
|
|
|
- architecture: frozen target model + external skill file optimization。
|
|
- tool use:
|
|
- memory:
|
|
- evaluation: 六个 benchmark、七个目标模型、三种执行模式。
|
|
- safety: bounded text edits、validation gating、rejected-edit feedback、slow/meta updates。
|
|
- deployment: 可审计技能文件。
|
|
- data:
|
|
|
|
## Evidence Quality
|
|
|
|
官方研究博客给出实验覆盖和机制细节，证据强。
|
|
|
|
## Gaps for Us
|
|
|
|
- 可以把本项目的 `skills`、prompt、playbook 也纳入 eval-driven 优化。
|