ConvApparel: Why Your AI User Simulator Probably Sucks

Google Research just dropped ConvApparel, and honestly, it’s about time someone called out the elephant in the room.

For years, we’ve been building conversational AI agents that can handle multi-turn tasks, ask clarifying questions, and even proactively assist. But anyone who’s actually deployed one knows the dirty secret: they fall apart in long interactions. They forget constraints. They generate irrelevant nonsense. And fixing this? Well, the ‘gold standard’ is live human testing, which is expensive, slow, and doesn’t scale worth a damn.

So the research community did what it always does: turned to LLM-based user simulators. You tell an LLM to roleplay as a human user, and boom, instant test subject. Except there’s a problem. A big one.

These simulators suffer from what the paper calls a ‘realism gap.’ They’re too patient. They have encyclopedic knowledge of the domain they’re supposedly exploring. They don’t get frustrated like real humans do. Think of it like a flight simulator that never throws in a sudden storm or a bird strike. You’re training pilots for a world that doesn’t exist.

ConvApparel is Google’s attempt to measure and bridge that gap. It’s a new dataset of human-AI conversations, paired with an evaluation framework that actually quantifies how realistic a simulator is. The clever bit? They used a dual-agent collection protocol. Real participants were randomly routed to either a helpful ‘Good’ agent or an intentionally unhelpful ‘Bad’ agent. This captures the full spectrum of human behavior, from satisfaction to profound annoyance. No cherry-picking polite interactions here.

The real innovation, though, is ‘counterfactual validation.’ Here’s the problem: a simulator trained on data from a helpful agent might just learn to mimic that behavior. But what happens when you throw it at a new, untested agent that behaves completely differently? If the simulator just repeats its training patterns, it’s useless. Counterfactual validation asks: how would a simulated user react if it encountered a frustrating system that looked nothing like the helpful ones it learned from? If it can’t adapt plausibly, you know it’s just memorizing, not understanding.

I’ve seen this pattern before in other domains. Remember when reinforcement learning agents would exploit quirks in their training environments? Same thing here. If your user simulator has never seen a bad agent, it will never learn to react like a frustrated human. And if you train your conversational agent only on these polite simulators, it will fail spectacularly when real users show up.

The ConvApparel dataset itself focuses on Conversational Recommender Systems (CRSs), which is a smart choice. These are complex, multi-turn interactions where an AI agent acts as a decision-support system. It’s not just answering a question; it’s guiding a user through a personalized recommendation process. That’s where the realism gap hurts most.

What I appreciate about this work is that it doesn’t just identify the problem. It provides a concrete methodology for testing and improving simulators. The three-pillar validation strategy (population-level statistics, human-likeness scoring, and counterfactual validation) is a solid framework. It moves beyond surface-level mimicry and actually probes for understanding.

Is it perfect? No. The dataset is still limited to a specific domain (recommender systems). And the evaluation framework, while clever, requires careful implementation. But it’s a significant step forward.

If you’re building conversational AI, stop pretending your LLM-based simulators are realistic. They’re not. ConvApparel gives you the tools to measure the gap and, more importantly, start closing it. Your users will thank you.