ConvApparel: Why Your AI User Simulator Might Be Too Polite (And How to Fix It)

You know the drill. You test your conversational AI on an LLM-powered user simulator, everything looks great, then you put it in front of real humans and it falls apart. The agent forgets a constraint, gives a weird answer, or just can’t handle someone who’s genuinely annoyed.

The problem isn’t necessarily your agent. It’s your simulator.

Google Research just dropped a paper on this exact issue, introducing ConvApparel, a dataset and evaluation framework that quantifies what they call the “realism gap” in LLM-based user simulators. Having spent years building and breaking these systems, I can tell you this is a problem I’ve felt in my bones, and it’s refreshing to see someone actually measuring it.

The simulator problem nobody talks about

LLM-based user simulators are everywhere now. They’re cheaper and faster than human testing, and they scale like crazy. But they’re also terrible actors. Ask an LLM to play a frustrated user, and it’ll probably write a polite paragraph about being “somewhat dissatisfied” instead of just saying “this sucks” and walking away. They have encyclopedic knowledge of domains real users don’t, and they display patience that would make a saint jealous.

Think about it: most LLMs are trained to be helpful assistants. Asking them to roleplay as an impatient, forgetful, or inconsistent human is like asking a ballet dancer to play a clumsy oaf. They can try, but it’s not natural. And if you train your agent only on these overly polite simulators, it’ll fail spectacularly when real humans show up with real frustration.

The counterfactual test that matters

Here’s where ConvApparel gets clever. The team at Google set up a dual-agent data collection protocol. Real humans were randomly routed to either a helpful “Good” agent or an intentionally unhelpful “Bad” agent. This captured the full spectrum of human behavior, from satisfaction to profound annoyance, and they validated it with population-level stats, human-likeness scoring, and something called counterfactual validation.

Counterfactual validation is the real star here. The idea is simple: how would a simulated user react if it encountered a frustrating system that looks nothing like the helpful ones it learned from during training? If your simulator has actually learned plausible human behavior, it should adapt. If it’s just blindly repeating training patterns, it’ll fail.

I’ve seen this play out in practice. You train a simulator on polite interactions, then test a new agent policy that’s intentionally bad, and the simulator just keeps being polite. It’s useless. Counterfactual validation calls this out directly.

ConvApparel in practice

The dataset focuses on Conversational Recommender Systems (CRSs), which are decision-support systems that need to handle complex, multi-turn tasks. The team established a baseline for human behavior by having real people interact with both good and bad agents, then used that baseline to evaluate existing simulators.

Unsurprisingly, most simulators failed the counterfactual test. They couldn’t handle out-of-distribution assistant behavior. But ConvApparel isn’t just a benchmark; it provides a path forward. By training simulators on this richer dataset that includes both positive and negative interactions, you can produce simulators that actually behave like real people, even when things go wrong.

Why this matters for your next project

If you’re building any kind of conversational agent, this should be on your radar. The days of assuming your LLM-based simulator is good enough are over. ConvApparel gives you a concrete way to measure whether your simulator is actually realistic, and more importantly, it provides data to train better ones.

I’m not saying every team needs to adopt ConvApparel wholesale, but the methodology is sound. If you’re serious about building robust conversational agents, you need to stress-test your simulators with counterfactual scenarios. Throw a bad agent at your simulator and see if it reacts like a real human would.

Otherwise, you’re just building a system that works perfectly in a simulator and falls apart in the real world. And nobody has time for that.