Title: Treating Discourse-Role Labels as Presentation-Time Variables for Context Utilization in Language Models

Abstract

While context-augmented language model systems frequently enclose provided content with specific discourse-role markers—such as "Reference:", "Evidence:", "Instruction:", "Note:", or "Example:"—the impact of these labels on how models process information remains largely unexamined. To investigate this, we developed a paired fixed-content probe utilizing over 500 items from MMLU-Pro. In this setup, each item is subjected to an identical assertion containing a misleading answer, but the assertion is preceded by varying discourse-role labels. We measured model behavior by tracking the "Misleading Adoption Rate," defined as the frequency with which the model outputs the injected incorrect option.

Our analysis, spanning GPT-5.5, DeepSeek V4 Pro, Llama-3-8B-Instruct, and Qwen2.5-7B-Instruct, reveals that the Misleading Adoption Rate fluctuates dramatically, shifting by 56 to 84 percentage points depending on the label used. Labels that imply binding authority or source credibility, such as "Instruction:" and "Reference:", lead to high rates of adoption. Conversely, the "Example:" label consistently suppresses this adoption.

We corroborate the hypothesis of label-conditioned candidate preference through paired tests, bootstrap intervals, ablations focusing on final instructions, and log-probability probes at the final step of Qwen. Boundary probes further clarify the limits of this effect: adoption decreases in arithmetic tasks, while passage-shaped external contexts maintain smaller label-induced gaps. Short-answer evaluations eliminate the possibility that results stem from simple option-letter copying, and tests involving nested-label conflicts indicate that illustrative framing can effectively delimit the scope of adoption. Additionally, a manual audit of 200 cases by a single author verifies that the observed contrasts in short-answer formats remain stable under conservative adjudication.

Although our findings are bounded, they offer practical implications: benchmarks for context utilization and reader-side Retrieval-Augmented Generation (RAG) should report and control for wrapper labels, as presentation choices significantly influence measured reliance on supplied context.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC