Why LLMs Might Not Get the Last Word: A New Look at User-Turn Generation

Most standard benchmarks for large language models (LLMs) stop short. They evaluate a model's ability to respond as an assistant and halt there. This method skims over a important capability: does the model have any inkling of what should happen next in a conversation?

Rethinking the Evaluation

Enter user-turn generation, a novel probe aiming to bridge this evaluative gap. By allowing a model to generate responses from the user's perspective in a given conversation, researchers are seeing whether the model encodes a sense of interaction awareness. If it does, the response won't merely be contextually accurate but will also meaningfully advance the dialogue.

Experiments conducted across 11 open-weight LLMs, including Qwen3.5 and gpt-oss, reveal a fascinating disconnect. Task accuracy and interaction awareness aren't as intertwined as one might think. For instance, within the Qwen3.5 family, although GSM8K task accuracy skyrockets from 41% to nearly 97%, the capacity for genuine follow-up languishes near zero when deterministic generation is applied.

Probing Deeper into Interaction

Surprisingly, when the temperature setting in language generation is increased, essentially adding randomness, the models demonstrate hidden depths of interaction awareness. Follow-up rates leap to 22%, unveiling a latent understanding of conversational dynamics. This isn't just academic. It introduces a new dimension to evaluating LLMs, one that existing benchmarks fail to capture.

Controlled perturbations further validate this new probe, confirming that it measures an authentic property of the models. Moreover, collaboration-centric post-training on Qwen3.5-2B models shows promise in boosting follow-up rates, reinforcing that these systems can be guided toward greater conversational sophistication.

The Future of LLM Evaluation

So, why should we care? The AI-AI Venn diagram is getting thicker. If we're to trust these models in autonomous decision-making roles, understanding their interaction awareness isn't just beneficial. it's essential. Do we want machines that merely respond, or ones that can think a step ahead in a dialogue?

In essence, this research challenges how we perceive AI communication skills. It suggests that current benchmarks may not fully encapsulate a model's conversational prowess. As LLMs continue to evolve, it's clear that their ability to understand and predict conversational flows will be a benchmark of their utility and intelligence. We're building the financial plumbing for machines, and that requires a strong understanding of interaction dynamics.