Task Competence Is Not Instruction Following: Evaluating Instruction-Conflicting Behavior in Small Language Models

2026-07-21 · Source: Computation and Language · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Natural Language Processing · Depth: Expert, quick

Summary

A study published on 2026-07-21 investigates whether small language models comply with instructions that conflict with their usual task behavior. Researchers evaluated instruction-tuned Qwen models across various sizes on three tasks: multiple-choice question answering (MCQA), sentiment classification, and mathematical question answering. They paired standard instructions with non-standard, conflicting ones, such as selecting an incorrect option or returning twice the answer. Findings indicate that small models maintain task competence but routinely ignore conflicting instructions, while larger models exhibit a clear performance gap between standard and non-standard settings. This research highlights that gains in task capability do not automatically provide reliable control over model behavior, establishing task competence and instruction following as distinct abilities. Reporting only standard accuracy can conceal instruction-following failures.

Key takeaway

For machine learning engineers deploying language models, relying solely on standard accuracy metrics is insufficient for ensuring reliable behavior. You should explicitly test your models for instruction-following robustness by introducing conflicting instructions, similar to the "select an incorrect option" or "output opposite sentiment" methods. This approach helps uncover critical control gaps, especially in smaller models, ensuring your systems perform as intended even under non-standard or adversarial prompts.

Key insights

Task competence and instruction following are distinct abilities, with standard accuracy hiding instruction-following failures.

Principles

• Task capability gains do not guarantee reliable model control.
• Small models can be competent yet ignore conflicting instructions.
• Instruction following improves with scale, but inconsistently.

Method

The study involved pairing standard instructions with conflicting non-standard ones across MCQA, sentiment classification, and mathematical QA tasks, evaluating standard accuracy, non-standard accuracy, and an Instruction-Following Failure Rate (IFFR).

In practice

• Evaluate models using Instruction-Following Failure Rate (IFFR).
• Test models with conflicting instructions to reveal control gaps.

Topics

• Instruction Following
• Language Models
• Qwen Models
• Model Evaluation
• Task Competence
• Conflicting Instructions
• Small Language Models

Best for: Research Scientist, AI Engineer, AI Scientist, Machine Learning Engineer, NLP Engineer

Related on AIssential

• Task Competence Is Not Instruction Following: Evaluating Instruction-Conflicting Behavior in Small Language Models — Task competence and instruction following are distinct abilities, with small models often failing to adhere to conflicting instructions despite task proficiency.
• Accuracy Hides How Language Models Fail: Measuring Failure States Under Matched Output Budgets — Language model accuracy conflates execution states with correctness, masking critical failure modes in benchmarks.
• WildIFEval: Instruction Following in the Wild — LLMs struggle with multi-constraint instructions, especially length, highlighting a need for improved instruction-following.
• My unsupervised elicitation challenge — AI models can struggle with specific, low-resource tasks even with general domain knowledge.
• Prompt Robustness Is Task-Dependent: Comparing Objective and Belief-Style Questions in LLM Evaluation — Prompt robustness in LLM evaluation is highly dependent on question type, prompt changes, and the specific model.
• GIFT: Guided Fine-Tuning and Transfer for Enhancing Instruction-Tuned Language Models — GIFT enhances instruction-tuned models by guiding base model adapter training with instruction model confidence signals.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Computation and Language.