A research paper proposes replacing the Markdown files that teach AI agents how to behave with tiny learned vectors baked directly into a model's context.
The technique, called SoftSkill, keeps a language model's weights frozen and instead trains a short "soft" prefix — just 32 tokens of continuous vectors — that steers how the model approaches a task. Tested on Qwen3.5-4B, SoftSkill beat no-skill prompting by 8.3 points on SearchQA, 42.1 points on LiveMath, and 1.3 points on DocVQA. Against SkillOpt, a prior text-based approach, it gained 5.2 points on SearchQA and 12.5 points on LiveMath — while swapping out hundreds or thousands of Markdown tokens for a handful of latent ones. The base model never changes; only the prefix is trained.
Right now, deploying a skill to an AI agent usually means writing a Markdown file that describes what the agent should do and hoping the model interprets it consistently at runtime. SoftSkill short-circuits that translation step, which matters because every extra token in a prompt adds latency and cost, and language models can misread long instruction files in ways that are hard to debug. Compressing behavioral policies into fixed-size vector prefixes makes skills faster, cheaper, and more reproducible.
The paper is honest about where the approach breaks down: for long-horizon agentic tasks — multi-step tool use, autonomous workflows — sparse trajectory data doesn't yet compress procedural behavior reliably, which means the Markdown file isn't going anywhere for complex agents just yet.