Measuring Instruction Capacity: How Language Models Degrade Under Load

Introduction

Language models often appear to lose instruction-following quality as prompts become more demanding. Yet “how many instructions can a model handle?” is not an easy question to answer. Real instructions differ in complexity, may interact with one another, and are usually evaluated with a mixture of human judgment and task-specific metrics. The paper Instruction Capacity: Behavior and Degradation of Language Models under Controlled Load proposes a more controlled way to study this limit.

Turning instructions into a measurable load

The proposed instrument gives a model a text with no intended meaning and a list of rules. Each rule replaces one single-token word with another word. The rules are designed to be atomic and independent, so the task does not depend on resolving semantic conflicts or understanding a complicated objective. The model must retain the mapping and apply it at the appropriate locations.

The construction makes automatic scoring possible. Every expected replacement is known in advance, including its location and rule number. Each replacement can therefore be treated as a Bernoulli trial with a correct or incorrect outcome. Instead of relying only on the quality of a whole answer, the researchers estimate the probability of a correct operation under different load conditions.

What the measurements reveal

The analysis examines several sources of load rather than collapsing them into one score:

• As the number of rules grows, replacement quality follows a regular relationship that is linear in the logarithm of rule count.
• Increasing replacement density produces a saturation pattern and eventually contributes to the collapse of the answer.
• The position of a rule in the instruction list matters, with the reported behavior described by a power law.
• Text length and the position of a replacement in the answer interact as constraints rather than acting as fully separate factors.
• Whole-answer degradation follows its own regularities, linked to the load carried by the text and progress through the answer, not merely to the number of rules.

For GPT-4o mini, the study analyzed 2,958 requests and 642,430 replacement trials. The authors fit smooth laws to these dependencies and use them to define “instruction capacity”: the number of rules a model can retain at a specified share of correct replacements. The estimate also includes uncertainty, rather than presenting a single unexplained threshold. The reported capacity is approximately 37–38 rules when the target share of correct replacements is 0.95. At full replacement density, it falls to approximately 18–19 rules.

Why it matters—and what it does not measure

The main contribution is methodological. “The model fails when there are too many instructions” becomes a reproducible measurement question with a defined load unit and an explicit statistical model. This could help compare models, study prompt organization, or decide when a larger task should be split into smaller subtasks in an agentic system. It may also reveal whether degradation is driven more by the number of rules, their ordering, or the position of operations in the generated answer.

The metric should not be interpreted as a universal measure of intelligence or a general-purpose instruction limit. It describes performance on a deliberately controlled replacement task. Different encodings, text distributions, and instruction types may produce different capacities. The study is nevertheless unusually auditable: its dataset, instrument description, and code are open, and the reported passport values can be recomputed from the published data.

Source: Hugging Face Daily Papers