Discourse-Role Labels as Presentation-Time Variables for Context Use in Language Models
Quick Take
Discourse-role labels significantly influence language model behavior, with misleading adoption rates varying by 56-84 percentage points across models like GPT-5.5 and Llama-3-8B-Instruct. Labels like 'Instruction:' and 'Reference:' increase reliance on incorrect options, while 'Example:' suppresses it. This highlights the need for context-utilization benchmarks to control for presentation choices.
Key Points
- Misleading adoption rates vary by 56-84 percentage points across tested models.
- Labels like 'Instruction:' and 'Reference:' lead to higher reliance on incorrect answers.
- 'Example:' label consistently suppresses misleading adoption.
- Boundary probes reveal context shapes the effect of labels on adoption.
- A manual audit confirms stability of short-answer contrasts under conservative adjudication.
Article Content
From source RSS / original summaryarXiv:2606. 04109v1 Announce Type: new Abstract: Context-augmented language model systems often wrap supplied content with labels such as Reference:, Evidence:, Instruction:, Note:, or Example:, but the effect of these labels on reader-model behavior remains underexplored.
We introduce a paired fixed-content probe over 500 MMLU-Pro items: each item receives the same misleading answer-bearing assertion under different discourse-role labels, and adoption is measured by whether the model outputs the injected wrong option. Across GPT-5. 5, DeepSeek V4 Pro, Llama-3-8B-Instruct, and Qwen2. 5-7B-Instruct, Misleading Adoption Rate shifts by 56-84 percentage points.
Binding or source-like labels such as Instruction: and Reference: produce high adoption, whereas Example: consistently suppresses it. Paired tests, bootstrap intervals, final-instruction ablations, and Qwen final-step log-probability probes support a label-conditioned candidate preference.
Boundary probes show where the effect weakens or persists: arithmetic tasks reduce adoption, passage-shaped external context preserves smaller label gaps, short-answer evaluation rules out option-letter copying, and nested-label conflicts suggest that illustrative framing can delimit adoption scope. A 200-case single-author manual audit confirms that the short-answer contrasts are stable under conservative adjudication.
The resulting claim is bounded but practical: context-utilization and reader-side RAG benchmarks should report and control wrapper labels, because presentation choices can change measured reliance on supplied context.
Reader Mode unavailable (could not extract clean content).
Want this in your inbox every morning?
Daily brief at your local 8am — bilingual EN/中文, free.
More from arXiv cs.CL
See more →Time to REFLECT: Can We Trust LLM Judges for Evidence-based Research Agents?
The REFLECT benchmark reveals that current LLM judges are unreliable, achieving below 55% accuracy in evaluating reasoning and evidence use, highlighting the need for improved evaluation methods for deep research agents.