Instruction Corporatism

From the AI Research Compass wiki

Instruction Corporatism: Holds that capability is organized and unlocked by instruction tuning, so every task should be cast as a natural-language instruction.

Beliefs

Instruction Corporatism believes in a well-ordered society of tasks. Translation, summarization, classification, code repair: each is a guild, and each guild is represented by instructions written in plain language. Pretraining supplies raw capacity, but it is instruction tuning that turns a text predictor into something that does what it is asked. Collect thousands of tasks, phrase them as instructions, mix them carefully, and a single model learns to generalize to instructions it has never seen.

It sits near the Language edge and near the top of the chart: one model, whose behavior is shaped by a large, curated system of task formats absorbed into its weights rather than handled by an engineered pipeline. Its paradigm level is orthodox. It was once the new idea that made zero-shot prompting practical; today it is core infrastructure in nearly every post-training recipe.

Arguments

  • Fine-tuning on a broad mixture of tasks phrased as instructions improves zero-shot performance on held-out tasks, a result replicated across many model families.
  • Instructions are a universal interface: any task a person can describe can be added without changing the architecture.
  • Models can generate their own instruction data, letting the corporatist system expand cheaply.

Critiques

  • Preference-learning camps argue instruction data teaches format more than judgment, and that feedback-based training matters more for real helpfulness.
  • Instruction mixtures inherit the biases of whoever chose and wrote the tasks; benchmark contamination is a constant worry.
  • Reasoning-focused rivals say that following instructions is not the same as solving hard problems, which needs search or long deliberation.

Representative works

Relationships

How to get it

Land on this cell in the full or pro test: x ≈ −0.73, y ≈ +0.70, paradigm ≈ −0.76.

A first draft. Corrections and better arguments are welcome through a pull request. Edit this entry on GitHub Edit its data