Many language model users ask vague questions that make it difficult to understand their intentions. For example, a question like “What book should I read next?” It is highly dependent on individual taste. The same is true. “How do antibiotics work?” Answers to questions should vary depending on whether the person is familiar with the subject. The current evaluation methods are often unable to take into account the context of the user, leading to inconsistent judgements. A response that praises coffee may seem good, but it could actually be harmful or unhelpful to someone who has a medical condition. The quality of the model’s responses can’t be accurately assessed without understanding the intent and needs of users.
Prior research has focused on generating clarification questions to address ambiguity or missing information in tasks such as Q&A, dialogue systems, and information retrieval. The methods are designed to better understand user intentions. Similarly, studies on instruction-following and personalization emphasize the importance of tailoring responses to user attributes, such as expertise, age, or style preferences. Similarly, some works examined the adaptability of models to various contexts. They also proposed methods for improving this adaptability. Language model-based evaluations have also gained popularity due to their effectiveness, but they are biased. This has led to efforts to improve them through more clear criteria.
Contextualized evaluations have been proposed by researchers from the University of Pennsylvania and University of Maryland College Park. The method uses synthetic context in the form of question-answer pair follow-ups to help clarify unclear queries when evaluating language models. They found that adding contextual information can impact the evaluation outcome significantly, even changing model rankings. This reduces the reliance on surface features such as style and reveals potential biases within default responses of models, especially in WEIRD contexts (Westernized, Educated Industrialized Rich Democratic). This work shows that different models are sensitive to various user contexts.
The researchers created a simple framework that allows them to assess how models respond when they are given more specific, context-specific queries. They first selected queries that were not well-specified from benchmark datasets, and then enriched these by adding question-answer pairs to simulate context-specific user situations. Then, they collected responses using different language models. Human and model-based evaluaters were asked to compare the responses of both the initial query and context. The context was used to determine how the model ranking, evaluator consensus, and judgment criteria were affected by context. This setup is a useful way to see how models deal with real-world ambiguity.
Adding context, such as user intent or audience, greatly improves model evaluation, boosting inter-rater agreement by 3–10% and even reversing model rankings in some cases. GPT-4 for instance outperformed Gemini-1.50-Flash, but only with context. Context is essential for evaluating the effectiveness of a message. The default generations are often biased towards Western audiences, general audiences, or formal settings, which makes them less useful for users of different backgrounds. The current benchmarks ignore the context risk, which produces unreliable outcomes. Evaluations should pair context rich prompts and matching scoring rubrics to ensure real-world relevance.
The majority of user questions to language models lack context, such as the user’s intent, or their expertise. Evaluations are subjective and therefore unreliable. In order to combat this issue, the researchers propose contextualized evaluations. These are queries that include relevant questions and responses. The added context can help shift focus away from superficial traits and towards meaningful criteria such as helpfulness. It may even be able to reverse the model ranking. The study also reveals underlying assumptions; many models default to WEIRD, (Westernized Educated Industrialized Rich Democratic) assumptions. Although the study is limited in its context types, and partly automated scoring was used to score it, this research makes a compelling case for future evaluations that are more aware of contextual factors.
Click here to find out more Paper, Code, Dataset You can also find out more about the following: Blog. The researchers behind this project deserve all the credit. SUBSCRIBE NOW Subscribe to the AI Newsletter


