Human Evaluation of LLMs

Suppose you have built a language model and you want to know whether its replies in a chat are any good. Checking one reply is simple: you read it and decide. Checking a thousand replies, produced across many different conversations, is a different kind of task. The effort does not stay the same as the pile grows. This article starts from that practical difficulty and works backward to the concept it depends on, which is conversational performance.

The source describes human evaluation as valuable but costly. It gives a concrete scale: reading and assigning ratings to all 1,100 responses would require a significant amount of effort. Keep that number in mind. Most of what follows explains what a person is judging in each response, and why that judgment is hard to repeat consistently across a large set.

Follow a single exchange when conversational performance is being evaluated. A person writes a message. The model produces a reply. A human evaluator reads that reply in the context of the conversation and assigns a rating. The diagram traces that path. Notice that the evaluator is the step that turns a reply into a judgment. The judgment is produced by a person reading, not by the model.

The same four-step path repeats for every reply in the set. The reply is read, its surrounding conversation is brought into view, and a rating is given. Because a person performs each of these steps, the cost of the whole set is the cost of many separate readings added together.

Read the definition in two parts. The first part names the goal: human-like communication. The second part names what makes that communication possible: understanding context, nuance, and intent. The phrase human-like does not mean the model must sound like a particular person. It means the exchange has the qualities of talk between people, where the listener grasps what was meant and not only the literal words.

The source lists three things the definition depends on. Context is the surrounding material that gives a message its meaning, such as what was said earlier in the conversation. Nuance is the subtle difference in tone or meaning that a flat reading would miss. Intent is what the person is actually trying to get done. A reply can be fluent and still fail on any of these three.

The three elements are not separate checklists. Each one shapes how the others are read. A short message such as a yes or a no can carry a very different meaning depending on the earlier turns, so context changes what the words mean. Nuance and intent then decide what the reply should actually accomplish.

The source goes beyond understanding and names observable skills that a conversational model shows. These include providing relevant and coherent responses, maintaining consistency, and adapting to different topics and styles of interaction. These are the qualities an evaluator looks for when reading a reply, so it helps to separate them out.

Take a concrete exchange to test the criteria. Imagine a user says they are planning a trip, names two cities, and then asks a few messages later which of the two is closer. This example is illustrative and was written for this article; it is not taken from the source.

The source states that human evaluation, while providing valuable insights, can be relatively laborious and time-consuming, especially when dealing with a large number of responses. The cost grows with the number of responses. Each reply has to be read, understood in its conversation, and rated, and the work does not get smaller as the pile gets bigger.