TLDR: A new research paper reveals that the single character used to separate in-context examples in Large Language Model (LLM) prompts dramatically affects model performance and evaluation rankings. Performance on benchmarks like MMLU can vary by over 20%, and model rankings can be manipulated by changing this delimiter. This brittleness is pervasive across model families, topics, and does not improve with scale, even affecting advanced models like GPT-4o. The study suggests that good delimiters steer model attention to key tokens. Practical solutions include explicitly specifying the delimiter in the prompt or using recommended characters like newline (‘\n’) or exclamation mark (‘!’).
In the rapidly evolving world of Large Language Models (LLMs), evaluations are crucial for measuring progress and comparing different models. These evaluations often rely on providing demonstration examples to guide the model’s responses into a desired style. While much attention has been paid to the number of examples used, a recent study uncovers a surprisingly critical factor: the single character used to separate these in-context examples.
A research paper titled “A single character can make or break your LLM evals” by Jingtong Su, Jianyu Zhang, Karen Ullrich, Léon Bottou, and Mark Ibrahim, highlights that this seemingly minor choice—whether to use a comma, a new line, a semi-colon, or other characters like ‘#’ or ‘|’—can dramatically alter the quality of an LLM’s response. The findings are significant, showing that performance on benchmarks like MMLU can fluctuate by as much as ±23% depending solely on the chosen delimiter. This level of variation is equivalent to several years of cumulative progress in language models, underscoring the profound impact of this overlooked detail.
The study demonstrates that this sensitivity is not limited to specific models or topics. Across leading model families such as Llama, Qwen, and Gemma, and across diverse topics within MMLU (STEM, humanities, social sciences), the choice of delimiter consistently affects performance. Even more strikingly, the researchers found that one could manipulate model rankings, placing any model in the lead, simply by changing this single separating character. This raises serious questions about the reliability and interpretability of current LLM benchmarks.
Furthermore, the research reveals that this brittleness does not improve with model scale. Comparing Llama-3.1-8B-instruct with its much larger counterpart, Llama-3.1-70B-instruct, the larger model, despite achieving better overall performance, exhibited similar or even greater sensitivity to delimiter choices. For instance, on the Commonsense QA task, the 70B model showed a ±40% fluctuation, compared to ±29.1% for the 8B model. Even closed-source models like GPT-4o were not immune, showing a remarkable performance spread of 45.63% on MMLU with different delimiters.
To understand the underlying mechanism, the researchers probed attention head scores using a dictionary lookup task. They found that well-performing delimiters effectively steer the model’s attention towards key tokens in the input, which are crucial for solving the task. For example, using a newline character (‘\n’) as a delimiter boosted attention scores for target keys by 25% for Llama-3.1-8B-instruct compared to a space character.
The paper also explores practical methods to improve LLM robustness to delimiter choices. A simple yet effective strategy is to explicitly specify the chosen delimiter within the prompt itself. By adding a line like “The following are multiple choice questions (with answers), separated by X” (where X is the delimiter), models consistently showed improved performance and robustness. For example, Qwen2.5-7B-instruct gained +14.2% on MMLU with this modification. Additionally, the study recommends ‘\n’ (newline) and ‘!’ (exclamation mark) as generally strong performing delimiters when explicit specification isn’t used.
Also Read:
- The Science of LLM Rules: Anchors, Attention, and Compliance
- Beyond Retrieval: How Input Length Alone Challenges Large Language Models
In conclusion, this research sheds light on a fundamental yet overlooked aspect of prompt engineering, demonstrating that a single character can have a profound impact on LLM evaluation outcomes and model behavior. It underscores the need for a deeper understanding of how delimiters interact with queries, training data, and model architecture, and offers practical recommendations for more robust LLM usage. For more details, you can read the full paper here.
