Beyond the Assistant Turn: User Turn Generation as a Probe of Interaction Awareness in Language Models

TL;DR

This paper introduces user-turn generation as a new method to probe whether language models encode awareness of ongoing interactions, revealing a dimension of behavior not captured by traditional benchmarks.

Contribution

It proposes a novel probe for interaction awareness in LLMs, demonstrating that models can generate grounded user follow-ups, which is often latent in standard evaluation.

Findings

Interaction awareness is decoupled from task accuracy.

Higher temperature sampling reveals latent interaction awareness.

Post-training improves models' follow-up generation rates.

Abstract

Standard LLM benchmarks evaluate the assistant turn: the model generates a response to an input, a verifier scores correctness, and the analysis ends. This paradigm leaves unmeasured whether the LLM encodes any awareness of what follows the assistant response. We propose user-turn generation as a probe of this gap: given a conversation context of user query and assistant response, we let a model generate under the user role. If the model's weights encode interaction awareness, the generated user turn will be a grounded follow-up that reacts to the preceding context. Through experiments across $11$ open-weight LLMs (Qwen3.5, gpt-oss, GLM) and $5$ datasets (math reasoning, instruction following, conversation), we show that interaction awareness is decoupled from task accuracy. In particular, within the Qwen3.5 family, GSM8K accuracy scales from $41\%$ ($0.8$B) to $96.8\%$ ($397$B-A$17$B), yet genuine follow-up rates under deterministic generation remain near zero. In contrast, higher temperature sampling reveals interaction awareness is latent with follow up rates reaching $22\%$. Controlled perturbations validate that the proposed probe measures a real property of the model, and collaboration-oriented post-training on Qwen3.5-2B demonstrates an increase in follow-up rates. Our results show that user-turn generation captures a dimension of LLM behavior, interaction awareness, that is unexplored and invisible with current assistant-only benchmarks.