Measuring Intent Comprehension in LLMs

People judge interactions with large language models (LLMs) as successful when outputs match what they want, not what they type. Yet LLMs are trained to predict the next token solely from text input, not underlying intent. Because written language is an imperfect proxy for intent, and correlations between phrasing and desired outcomes can break down in training data, models that rely too heavily on surface cues may respond inconsistently to semantically equivalent prompts. This makes it essential to evaluate whether LLMs can reliably infer user intent—especially in high-stakes settings where robustness and generalization are critical. We introduce a formal framework for assessing intent comprehension in LLMs: whether a model demonstrates robust understanding of user intent by producing consistent outputs across semantically equivalent prompts while differentiating between prompts with distinct intents. Our evaluation approach is based on a variance decomposition of model responses into three components: variability due to user intent, user articulation, and model uncertainty. Models that understand what users want, and are not overly sensitive to textual cues, should attribute most output variance to intent differences, rather than articulation style. Applying this framework across diverse domains, we find that, within the five LLaMA and Gemma models we evaluate, larger models typically assign a greater share of variance to intent, indicating stronger comprehension of intent, although gains are uneven and often modest with increasing model size. These results motivate moving beyond accuracy-only benchmarks toward semantic diagnostics that directly assess whether models understand what users intend.

People often ask a language model for the same thing in different words. At an airport, "What is the fastest way to Terminal B?" and "My flight is boarding soon, is there a shortcut to Terminal B?" express the same goal: reaching Terminal B as quickly as possible. Asking instead how crowded Terminal B is expresses a different goal. For language models to be useful and robust, they should track what we mean rather than how we happen to say it. A model that understands a request should change its answer when the underlying intent changes, and hold its answer steady when only the wording changes and the intent stays fixed. This paper proposes a way to test whether models actually behave this way. We compare a model's answers across two kinds of prompt pairs: pairs that preserve the goal but vary the phrasing, and pairs in which the goal itself changes. Our method then decomposes the variation in answers into three parts: variation driven by the user's goal, variation driven by wording, and leftover uncertainty. We apply this test to several language models on everyday estimation questions—travel times, household costs, health habits, delivery logistics, and social planning. We find that larger models generally respond more to genuine changes in goal, but that they can also be more sensitive to wording, and that performance varies by topic. The framework gives researchers and practitioners a diagnostic for checking whether a model follows what people mean rather than how they happen to say it.