Instruction-following tax
The instruction-following tax, also referred to as the "curse of instructions," describes a non-linear inverse relationship between the number of active behavioral constraints in a system prompt and the model's compliance Verified Answer #1. Empirical evidence from mid-2026 indicates that as the density of constraints increases, the probability that a large language model (LLM) will satisfy all requirements diminishes significantly Verified Answer #1.
Patterns of Performance Degradation
Research using benchmarks such as ManyIFEval and IFScale has identified three distinct patterns of compliance decay across frontier LLM architectures Verified Answer #1.
- Threshold Decay: This pattern features near-perfect adherence until a critical density threshold is reached, followed by a sudden and sharp collapse in performance Verified Answer #1. It is commonly observed in advanced reasoning-focused models like o3 and Gemini 2.5 Pro Verified Answer #1.
- Linear Decay: In models such as GPT-4.1 and Claude 3.7 Sonnet, instruction-following accuracy experiences a steady and predictable decline proportional to the increase in constraint load Verified Answer #1.
- Exponential Decay: Characterized by rapid initial performance drops that eventually plateau at lower accuracy levels, this pattern is often observed in models like GPT-4o and Llama 4 Scout Verified Answer #1.
Underlying Mechanisms of Failure
The degradation of instruction-following is attributed to specific cognitive and architectural constraints within LLMs Verified Answer #1.
Attention Dilution and Primacy Effect
Models often exhibit a non-uniform allocation of attention, prioritizing instructions located at the beginning of a system prompt Verified Answer #1. This "primacy bias" causes the model to ignore or deprioritize constraints placed later in high-density prompts Verified Answer #1.
Multiplicative Probability of Failure
Benchmark data suggests that the overall success of a model in following a set of multiple instructions can be modeled as a multiplicative probability of failure for each individual constraint Verified Answer #1.