acceptodds
Under review as a conference paper at ICLR 2027

The Other Side of the Conversation: Learning Behavioral User Models from User-Side Turns

Abstract

Conversational intelligence requires an assistant to not only respond to what is said, but to anticipate how a particular interlocutor is likely to respond. But can LLMs construct such a user model from dialogue alone? We study this question in a controlled setting where a synthetic user is defined by a combination of persona and style, encoded in a known system prompt . We evaluate how well three strategies can approximate by only observing conversations, without having access to : 1) conditioning on demonstrations in context, 2) explicitly reconstructing a natural-language description of the user, and 3) learning a continuous soft prompt from user-generated turns. Our evaluation shows that these conversation-only methods recover a substantial amount of user-specific signal. However, fidelity to the user depends on the criterion we apply: distributional agreement, recovery of persona and style, or how reliably generated text is judged to come from a user. Our framework provides a quantitative way to evaluate user simulation from dialogue and its limits.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
LLM Conversations, LLM Persona, Soft Prompt, Prompt Inversion, Behavioral Model, User Simulation, Theory of Mind

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_WbLeOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper investigates the ability of LLMs to detect the persona of synthetic users that are given to them as instructions inside their user role-playing prompt. The paper experiments with three approaches: in-context learning, explicitly reconstructing a natural-language description of the user, and learning a continuous soft prompt from user-generated turns. The paper then presents results and analysis using the Qwen3.5-9B model.

Critical strengths

The paper addresses the problem of trying to estimate the user's persona traits given their historical interaction with the model. This could be interesting for the sake of better personalization of LLMs to specific user traits. However, the paper doesn't really motivate why this is useful or provide experiments that demonstrate this.

Critical weaknesses

The paper lacks serious clarity in writing, making it difficult to understand what the main motivation and contributions are despite re-reading several parts. Further, the experimental details provided are limited, it is unclear what data is used for evaluating the model. The tested approaches are also a few standard methods that are overly described as something novel. Overall, in its current form, I do not find any new methodological or experimental novelties in this submission.

Questions for authors

No particular questions, please see the weaknesses section for the main issues in the paper.

Reviewer_TicfOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper evaluates how well an LLM can approximate a model of its conversation partner from dialogue alone. The authors compare three strategies: in-context learning, natural language prompt inversion, and soft prompt tuning.

The authors evaluate the strategies on 50 synthetic users defined by persona prompts (10 personas ×\times 5 writing styles), using three metrics: the token-level KL divergence from the true user, agreement on questionnaires, and attribution of generated conversations by an LLM judge. They report that soft-prompt tuning best approximates the user's response distribution. The inference-only strategies, by contrast, produce text that is recognizable as the user without matching that distribution. Based on this, they argue that recognizable imitation of a user and predictive fidelity to the user are distinct.

Critical strengths

The evaluation comprehensively covers various tasks and methods.

Critical weaknesses

Figure presentation can be improved (presentation 2)

The text in the figures is too small to read. Also, captions under the figures are difficult to understand on their own.

The setting favors soft-prompt tuning (soundness 2)

The authors conclude that optimizing on user turns recovers the user better than inference-only strategies. However, the setting makes this result close to expected.

The user and the learner share the same frozen model. Learning a soft prompt that reproduces the user prompt is therefore close to prompt compression on the same model, which is known to work well. The main metric, the token-level KL divergence, is also the quantity that soft-prompt tuning approximates. It is also used to select the soft-prompt checkpoints on the same topics that we report.

The inference-only strategies observe at most 25 conversations, while soft-prompt tuning observes 200. The prompt reconstruction is a single greedy pass of the same 9B model.

The unfair comparison and limited scale of experiments make the generalizability of the findings in this paper questionable (informativeness 1, rating 2)

Due to the previous point, the gap between recognizable imitation and predictive fidelity may partly reflect the unequal budgets and the weak reconstruction rather than a property of user modeling.

The paper focuses on synthetic users only. Due to the lack of human-subject experiments, it is unclear how much of the findings in the paper can translate to realistic users.

In addition, the experiment's scale is limited in model coverage and persona scale. Only relatively small models from the limited families are studied. And 50 synthetic users from a fixed set of personas and styles only provide a controlled proof of concept.

Questions for authors

With the same number of conversations and a stronger prompt reconstruction, do the inference-only strategies still fall behind soft-prompt tuning in KL divergence?

Can you provide a small study with real human users (10-20 participants)? Can you test on models significantly larger and from different families? Do the findings transfer?

Which downstream task should this enable? Can you evaluate one of these concretely rather than using proxy metrics?

Reviewer_qHhyOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies whether a model can construct a behavioral model of an interlocutor from user-side dialogue turns alone. To make recovery measurable, it creates 50 synthetic users by combining 10 persona prompts with five writing-style prompts, all instantiated in a frozen Qwen3.5-9B model. The authors compare in-context learning, natural-language prompt inversion, and per-user soft-prompt learning from observed user turns (CONVSFT). They also introduce privileged soft-prompt distillation references that access the original user prompt and its next-token distribution. Evaluation covers symmetric conversation KL on held-out dialogue topics, questionnaire agreement, learned-prompt geometry, generated-conversation attribution by a judge LLM, and behavioral-change detection.

The central empirical finding is that inference-only methods can reproduce recognizable persona cues, while soft prompts optimized on observed user text more closely match the target conditional distribution and better preserve style. The paper further finds that conversational breadth is generally more valuable than depth in this setup, and that different metrics can reverse the ranking of methods.

Critical strengths
  1. A useful controlled framework for measuring behavioral recovery. Knowing the generating persona–style prompt provides access to the target conditional distribution, enabling direct measurement of behavioral recovery. The factorial persona/style construction and cross-persona/cross-style calibration provide interpretable reference points for assessing recovery under controlled conditions. This makes the synthetic testbed methodologically useful (Section 3 and Figure 2).

  2. Complementary diagnostics reveal distinct aspects of user modeling. Conversation KL, questionnaire agreement, generated-conversation attribution, and change detection examine different aspects of user modeling. The main results show that agreement with elicited traits and recognizable persona cues need not imply distributional fidelity. This is an informative empirical contribution and supports using multiple criteria to evaluate user models (Sections 5–6 and Figures 3–4).

  3. The reference design and main ablations make the comparisons informative. CONVSFT versus CONVdist holds conversational evidence and capacity fixed while changing sampled-token to distributional supervision. CONVdist versus ALPdist probes conversational versus instruction contexts; ALPdist versus ALPoff dist probes response-prefix source. The breadth, depth, and capacity sweeps further examine how recovery depends on evidence and representation. Together these form an informative diagnostic study (Sections 3–4 and Figure 2).

Critical weaknesses

1.The differentiated contribution is narrower than the framing suggests. Profile inference, conditioning on user history, and soft-prompt personalization already have substantial precedent. The paper cites PERSOMA, Recon, and Latent Personal Memory, so this is not a claim that its recent literature coverage is absent. However, directly relevant USP (ACL 2025), HumanLM (2026), and Turing-RL (2026) are missing from the supplied manuscript [R1–R3]. These study inferred profiles or training objectives for individualized simulation. They do not duplicate the paper’s known-target KL evaluation, but they reduce the novelty of the broader motivation and require a sharper account of what new knowledge this particular protocol establishes.

2.The comparisons do not fully support the broader method-class conclusions. The three observational strategies are sensible first controls, and Figure 2(a) already compares KL at equal budgets of 1–25 conversations. However, the questionnaire and generation-attribution comparisons use 25 conversations for text methods and 200 for conversational soft prompts (Sections 5–6). Comparing models given vastly different amounts of data fails to isolate whether the performance gap stems from the representation/optimization strategy or simply from seeing more evidence. Furthermore, a single greedy prompt-inversion instruction that favors a committal description is a limited representative of text-based inference (Appendix C). A stronger evidence-grounded profile or a retrieval/selection-based candidate baseline would better test the Introduction's claim that user-specific optimization is required to capture the target conditional distribution. The empirical distillation references cannot by themselves identify all information revealed by conversations, as the paper correctly acknowledges that they are not theoretical upper bounds. Residual gaps mix elicitation, sampling, representation, and optimization effects.

3.Essential experimental choices require too much cross-referencing. Section 3 defines the observational access restriction, but understanding the complete pipeline requires combining the main results with Appendices B.2–B.4, C, G, and I. The five-token default is evaluated in the 1/5/10-token ablation in Figure 2(d), but the main text should point to this ablation when the parameterization is first introduced. More consequentially, observation budgets and checkpoint rules vary across results. The terminology around interaction length must be unified: Figure 2(a)’s “3 user turns” needs to explicitly map to Appendix C’s “three generated user responses” and clarify how this aligns with Appendix B’s depth convention, which counts the opening dilemma as a turn. A compact main-text table covering observations, targets, privileged access, selection, and evaluation budget would make the study substantially easier to assess.

[R1] Wang et al. 2025. Know You First and Be You Better: Modeling Human-Like User Simulators via Implicit Profiles. ACL. https://aclanthology.org/2025.acl-long.1025/

[R2] Wu et al. 2026. HumanLM: Simulating Users with State Alignment Beats Response Imitation. https://arxiv.org/abs/2603.03303

[R3] Wang et al. 2026. Learning User Simulators with Turing Rewards. https://arxiv.org/abs/2606.19336

Questions for authors
  1. Can the key comparisons be matched on observed evidence? Please evaluate the existing 25-conversation CONVSFT condition alongside ICL and inferred prompts on the questionnaire and generation-attribution metrics, retaining the 200-conversation condition as a separate scaling result. Report generated user responses, and history/target token counts. One stronger frozen-backbone text alternative, such as a structured evidence-grounded profile or an adapted retrieval/selection-based candidate baseline, would strengthen the empirical claims.

  2. What is the new conclusion beyond existing user modeling? Please position the study against USP, HumanLM, and Turing-RL, while distinguishing them from the already discussed Recon and dynamic soft-prompt memory [R1–R5]. Which finding is enabled specifically by access to the true conditional distribution? Can the statements about requiring user-specific optimization and separating revealed from extracted information be restricted to the tested methods and empirical reference gaps? An explanation plus a representative strengthened comparison would help clarify the exact contribution.

  3. Can readers reconstruct the protocol from the main text? Please state the turn-count convention, observation budget, and checkpoint rule next to each principal comparison, and summarize the already available non-oracle selection checks. A single method/access/selection table and direct ablation references would resolve much of the reading burden. Distinguish unexpressed persona information from demonstrated recovery error; a finite minimum sufficient dialogue count has not been established.

[R4] Zhu et al. 2026. Recon: Reconstruction-Guided Reasoning Synthesis for User Modeling. https://arxiv.org/abs/2605.26969

[R5] Das et al. 2026. Latent Personal Memory: Represent personal memory as dynamic soft prompts. https://arxiv.org/abs/2606.20911

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.