The Other Side of the Conversation: What User-Turn Generation Reveals About Language Models

Introduction

We have built an entire evaluation culture around a single turn. A user asks something, the model answers, a verifier checks the answer, and we move on. This is the assistant turn, and for years it has been the unit of measurement in nearly every benchmark we care about: GSM8K, GPQA, IFBench, and the rest. The implicit assumption is that if a model produces correct responses, it has learned something deep about conversation. The paper "Beyond the Assistant Turn: User Turn Generation as a Probe of Interaction Awareness in Language Models" by Shekkizhar, Cosentino, and Earle at Salesforce AI Research challenges this assumption in a clean and uncomfortable way.

The idea is simple enough to state in one sentence. Take a model, give it a user query and an assistant response, then force it to continue generating under the user role rather than the assistant role. Since the weights do not change, whatever comes out is a behavioral readout of what the model believes a plausible user would say next. If the model has internalized any awareness of conversational dynamics, the generated user turn should be a grounded follow-up that reacts to the assistant's answer. If it has not, you get garbage: identity drift, repeated previous turns, or the model simply continuing as the assistant again.

What the Probe Measures and Why It Matters

The authors formalize user-turn generation as u = M_θ([q; a]), where the model conditions on a query q and assistant response a, then generates under the user header. They distinguish two settings. In the self-generated setting, the assistant response is the model's own output, so the probe tests whether the model can anticipate the user reaction to its own answer. In the held-out setting, the assistant response comes from a real conversation, so the probe tests follow-up generation grounded in an externally provided response.

This distinction matters more than it might first appear. The self-generated setting is essentially a test of self-consistency in a model's world view. A model that produces a confident, correct math solution and then, under the user role, generates a plausible "Thanks, but can you explain step three?" is demonstrating something close to partner modeling. It is representing not just the content of its answer but the effect that answer has on a counterpart. This is the kind of capability that theory-of-mind work (Sap et al., Gandhi et al.) has tried to elicit through indirect probes. User-turn generation gives us a direct behavioral measure.

The headline empirical finding is that interaction awareness is decoupled from task accuracy. Within the Qwen3.5 family, GSM8K accuracy scales from 41% at 0.8B parameters to 96.8% at 397B-A17B. Yet under deterministic greedy decoding, genuine follow-up rates stay near zero across that entire range. The model that almost never makes an arithmetic mistake is also the model that, when asked to continue as a user, produces something that is not a real conversational continuation. Scaling solves the assistant turn. It does nothing for the user turn.

The Temperature Result and Its Interpretation

The most analytically interesting result is what happens when you raise the sampling temperature. At T=0, most models produce near-zero follow-ups. As temperature increases toward 1.0, genuine follow-up rates climb substantially: Qwen3.5-27B goes from 0% to 22% on GSM8K, and follow-up rates on IFBench reach 43.7% for some models. The capability is latent. It exists somewhere in the model's distribution, but it is not the modal continuation. Greedy decoding, which selects the highest-probability token at each step, walks the model straight into degenerate behavior.

This is a subtle and important distinction. There are two ways a model can fail at user-turn generation. One is a genuine capability gap: the weights simply do not encode any representation of plausible user reactions. The other is a suppression problem: the capability is present in the distribution but assigned low probability mass relative to degenerate continuations like identity drift. The temperature experiments suggest that for Qwen3.5 and GLM, the second story is closer to the truth. For gpt-oss, follow-up rates stay near zero even at high temperature, which points to a real capability gap or at least a very different training distribution.

The authors validate that the probe measures something real through controlled perturbations and a collaboration-oriented post-training experiment on Qwen3.5-2B, which increases follow-up rates. The perturbation work is essential. Without it, one could argue that high-temperature follow-ups are just noise that happens to look conversational. The fact that targeted perturbations move the metric in predictable directions, and that post-training can deliberately raise it, suggests the metric tracks a genuine property of the weights rather than an artifact of sampling randomness.

My Take: Sampling Problem or Capability Gap

Here is where I want to push beyond what the paper claims. The temperature result reframes the entire question. If latent interaction awareness exists but stays suppressed under greedy decoding, then for some model families partner modeling is partly a sampling and decoding problem, not purely a capability gap. That distinction matters enormously for how we approach self-play and multi-agent training.

Consider the implication for reinforcement learning pipelines. Many self-play and agent-to-agent training setups generate rollouts with low-temperature or greedy decoding for stability. If the conversational continuations the model needs to learn from are precisely the ones suppressed at low temperature, then we are training on the degenerate part of the distribution and reinforcing identity drift. The Mon-Williams et al. observation that partner models only emerge when the training environment imposes sufficient pressure takes on a new dimension here: the pressure may need to act not just on what the model can represent, but on what it surfaces under realistic decoding.

This leads to the experiment I most want to see, which I flagged in my earlier note on this paper. Does the high-temperature follow-up rate predict downstream multi-agent collaboration quality? If a model with 22% latent follow-up rate at T=1.0 outperforms a model with near-zero follow-up rate even at high temperature when both are dropped into a collaborative task, then the metric is measuring something causally relevant to collaboration. If it does not, then user-turn generation is an interesting curiosity but not a useful predictor. The decoupling from task accuracy makes this a genuinely open empirical question rather than a foregone conclusion.

There is also a deeper interpretability angle. The fact that the same weights produce coherent assistant turns and degenerate user turns under greedy decoding tells us that the chat template and role headers are doing heavy lifting. The model has effectively been trained to treat the user role header as a signal to do something other than continue the conversation faithfully. Post-training optimizes the assistant role relentlessly while leaving the user role distribution to drift. This is an artifact of how RLHF and instruction tuning allocate their optimization pressure, and it is fixable.

Limitations and Open Questions

A few caveats are worth stating honestly. First, "genuine follow-up rate" is necessarily defined by some classifier or judgment procedure, and the boundary between a grounded follow-up and a plausible-looking but ungrounded one is fuzzy. The validity of the whole framework rests on how reliably that distinction is drawn. Second, the held-out versus self-generated distinction may produce quite different conclusions, and the broader narrative should be careful not to conflate them. Third, gpt-oss behaving differently from Qwen and GLM means we cannot assume the temperature story generalizes; it may be family-specific.

The open questions are what make this worth pursuing. Does latent interaction awareness predict collaboration quality? Can targeted post-training raise greedy-decoding follow-up rates without harming assistant accuracy? And most importantly, is there a fundamental tension between optimizing the assistant turn and preserving the user turn, or can a model be good at both? The answer determines whether the next generation of conversational and multi-agent systems treats both sides of the dialogue as first-class objects of training, rather than measuring only half of what a conversation actually is.