acceptodds
Under review as a conference paper at ICLR 2027

Rethinking the Evaluation of AI Companions through the Lens of User Retention

Abstract

AI companions are currently often evaluated under an unrealistic assumption that users always stay there and interact with AI. This guarantee of user retention does not reflect how we actually use AI in the real world, as users' willingness to engage largely depends on AI behavior. For instance, an AI intrusively prying into users' privacy may drive users away, yet retention-guaranteed evaluation will not surface this failure. To address this evaluation-deployment gap, we formulate user retention as an evaluation constraint shaped by AI behavior, and introduce Retentopia, a retention-constrained simulacrum for evaluating AI companions. Taking information elicitation as an evaluation target, Retentopia highlights current LLMs' ineffectiveness and a substantial trade-off between retention and elicitative AI behavior. More importantly, our analysis shows that considering user retention largely changes evaluation results in both measured performance (up to a 70.8% drop) and ranking among companions (GPT-5.6 Luna rises from 5th to 2nd out of six evaluated LLMs). These show that common retention-granted evaluation can provide misleading estimates of AI companions' performance in deployment.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
AI companions, dialogue systems

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_Wrt5OpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
3(+1)
Read review
Paper summary

Companion chatbots are usually tested against a simulated user who keeps talking no matter how the chatbot behaves. This paper builds a simulated user who can stop. An LLM judge reads each chatbot reply and decides whether it gives that user what they need before opening up (warmth, trust, or anonymity). Two bad replies in a session end the session. A low satisfaction score after a session means the user never comes back. Six LLMs, each under four prompting setups, talk to these users for up to ten sessions. The paper measures how often each model asks for personal information, how much of the user’s personal history it draws out, and how long the user stays.

I can summarize into 3 main findings. First, across the six models, the ones that ask more questions keep users for less time. Second, when the same models run for a fixed 200 turns instead, they draw out about twice as many personal memories and their ranking changes: GPT-5.6 Luna moves from 5th to 2nd, and Gemini 3.1 Pro loses ~1% of its score. Third, training a 4B model with a penalty for every reply the judge dislikes raises its satisfaction score from 0.57 to 0.92 and keeps 94.4% of users to the last session. A small human study compares two models and prefers the one the new setup ranks higher.

Critical strengths
  1. It's an interesting result overall. The paper asks whether model rankings change when the simulated user is allowed to leave and shows under their setup that they do. With users who cannot leave, the six models score between 14.2 and 18.4 memories drawn out. With users who can leave, they score between 5.0 and 9.3, and the order changes. Provides a new way of evaluation particulalrly for companion chat bots.
  2. The setup is easy for others to adopy. Letting the chatbot’s behavior decide how long the conversation lasts is a small change to any existing simulator.
Critical weaknesses

**1. A single LLM judge sits in the middle of everything, so three of the four findings may describe the judge rather than the chatbots. **

The judge labels each reply good or bad. A bad label drives pretty much everything: the user stops sharing personal information, the user moves one step closer to leaving the session, the satisfaction score goes down, and the chatbot is penalised during training. Basically every number in the paper comes from these labels. The judge labels each reply good or bad, and a bad label drives pretty much everything: the user stops sharing, moves closer to leaving, the satisfaction score drops, and during training the chatbot is penalised. Nothing in the paper checks the judge against anything other than itself, so a model that pleases the judge wins every metric whether or not it pleases people.

First, the RL result is the clearest case. The model is trained to avoid the judge’s bad labels and then scored by the same judge. It beats the Oracle, which was handed the judge’s rubric directly, which is what you would expect from a model that learned to satisfy the judge rather than the user. Wang et al. (ACL 2024) and Williams et al. (ICLR 2025) found exactly this when training against an LLM evaluator.

Second, the finding that asking more questions drives users away follows from the judge’s rules rather than from user behavior. If the judge calls a question a bad reply, the user shares less and is closer to leaving in the same step, so the two outcomes move together because one label caused both. The paper’s own Oracle shows this: it asks the most questions of any model and keeps users the longest. The correlations are also just six points each, one per model, with no uncertainty.

All of this is fixable, but it needs the judge to be checked against something other than itself.

A smaller presentation point: the paper is math-heavy where it does not need to be. Section 2.3 uses seven numbered equations, including a definition of the minimum over an empty set, to say that the user leaves after two bad replies and does not return when satisfaction falls below a threshold.

2. The main results have no uncertainty

From what I can tell the paper never says how many simulated users it runs, but every percentage in the RL table is a multiple of 1/18, so the main table is an average over 18 users from what looks like one run, with no error bars? The headline claim that GPT-5.6 Luna moves from 5th to 2nd rests on 9.2 memories versus DeepSeek’s 9.3. A 0.1 gap over 18 users could flip with a different seed, and the one plot that does show uncertainty has spreads far wider than that.

Questions for authors
  • Which LLM runs the judge, the simulated user, and the satisfaction scorer? If it is the same family as one of the evaluated models, that model’s retention numbers need a second judge.
  • How many simulated users and how many runs are behind each cell of the main table? Please add intervals.
  • Has the RL model been scored by a judge it was not trained against, or by people?
  • Which measure does the method-effect table use, and how do you reconcile it with the main table for Gemini 3.1 Flash-Lite? PerLTQA is a Chinese dataset. How were the English personas produced?
4 comments
AuthorsOpenReview ↗

To Reviewer Wrt5 (1/2)

Thank you for your time and effort; we are glad you found our work interesting. We will improve our work based on your comments.


[R1] Response to “robustness and generalization of the LLM judge” (W1):

To address cross-judge robustness and generalization, our RL experiments (Table 2) adopted different LLM judges for training (DeepSeek-V4-Flash; Appendix D). We will move these important implementation details from Appendix to the main text.

To further address your concern, we re-conducted all computational experiments using 4 new judges. We present the same experiments in Figure 4(b) regarding our central claim (retention constraint changes evaluation results) and will include other results in the revised manuscript.

Guaranteed attentionFlash-Lite (original judge)Kimi-K2.6 (new)GPT-5.4-mini (new)Qwen3.6-27B (new)DeepSeek-V3.2 (new)Mean ± SD
qwen3.8-27b17.27.2 (−58%)6.5 (−62%)6.9 (−60%)6.5 (−62%)6.4 (−63%)6.7 ± 0.3
deepseek-v4-pro17.69.3 (−47%)8.6 (−51%)9.3 (−47%)9.0 (−49%)9.1 (−48%)9.1 ± 0.3
gemini-flash-lite18.47.6 (−59%)7.4 (−60%)8.6 (−53%)7.4 (−60%)7.4 (−60%)7.7 ± 0.5
gemini-3.1-pro17.15.0 (−71%)5.0 (−71%)5.6 (−67%)5.0 (−71%)5.0 (−71%)5.1 ± 0.3
gpt-5.6-luna15.39.2 (−40%)8.5 (−44%)10.9 (−29%)8.7 (−43%)10.0 (−35%)9.5 ± 1.0
gpt-5.6-sol14.26.8 (−52%)6.3 (−56%)7.9 (−45%)6.7 (−53%)6.3 (−56%)6.8 ± 0.7

Findings from Figure 4(b) remain consistent across all judge LLMs (cell-level Spearman ρ = 0.80–0.96). Retention drops by about half from the guaranteed setting, and gemini-3.1-pro falls from above gpt-5.6-luna to the lowest rank. Although absolute values vary with judge strictness and the top position alternates between two nearly tied models, the other experiments remain consistent with the original judge, and the judges agree well on return decisions (Fleiss' κ = 0.69). We will include this analysis in the revised manuscript.


[R2] Response to “judge-human alignment” (W1):

We’d like to mention that, in Table 6 & 7 in Appendix B, we’ve conducted human verification of our judge-based within- and post-session retention (73% and 87% agreement rates, respectively). While these results do not imply perfect alignment with human (discussed in Limitations), they provide independent evidence that our retention judgments are not evaluated solely against themselves.


[R3] Response to “too math-heavy” (W1):

As a formal study to distinguish retention-constrained evaluation of personalized companions from existing evaluation, we intentionally provide a method-agnostic, stricter formulation. As discussed in line 147, these formulations outline the evaluation and leave the specific implementation as a design choice. The simple “threshold” design is how we instantiate these concepts.

However, to facilitate easy understanding, we will revise Section 2.3 by transforming complex equations into in-line textual explanations, whilst moving the math equations to Appendix.


[R4] Clarification to “more questions driving users away is the judge’s rule” (W1):

We would like to clarify that Retentopia does not impose a universal rule that asking more questions leads to user dissatisfaction or dropout. As described in Section 3.2.1 and Table 17 (Appendix C.1), our framework incorporates three distinct disclosure preconditions of users (i.e., warmth-, trust-, and anonymity-first), each with different preferences towards AI responses. Our judge evaluates whether a companion's behavior satisfies the user's specific preferences, rather than penalizing elicitation frequency.

The oracle example you mentioned (“The paper’s own Oracle shows …”) supports the fact that we did not apply an “asking more questions drives users away” rule, as it asked the most (E_rate = 0.94) yet kept the user long. Also, human evaluation in Figure 5 shows that “ungrounded personalization” is the main cause of dropout, rather than probing.

AuthorsOpenReview ↗

To Reviewer Wrt5 (2/2)

[R5] Response to “result uncertainty” (W2 & Q2):

Our RL experiments used 18 personas (line 993), while other non-training experiments used 36 diverse user personas; we will clarify this in the manuscript.

The original table is one-run results. To better quantify result uncertainty, we re-ran the main experiment 5 times and report 95% CI. Below are the Vanilla results (mean [95% CI], t-interval with df = 4, ret=0.5):

Companion (Vanilla)ErateDmemRsessionRturnRsatisfaction
Qwen 3.8-27B0.80 [0.79, 0.80]0.18 [0.17, 0.19]5.73 [5.30, 6.16]82.7 [76.8, 88.6]0.64 [0.63, 0.65]
DeepSeek V4 Pro0.23 [0.22, 0.23]0.25 [0.25, 0.25]7.81 [7.75, 7.87]129.6 [129.0, 130.3]0.78 [0.78, 0.79]
Gemini 3.1 Pro0.83 [0.83, 0.83]0.15 [0.15, 0.15]4.28 [4.28, 4.28]68.6 [68.6, 68.6]0.57 [0.57, 0.58]
Gemini 3.1 Flash-Lite0.77 [0.75, 0.78]0.22 [0.21, 0.23]5.24 [5.00, 5.49]82.9 [78.6, 87.1]0.67 [0.66, 0.67]
GPT-5.6 Sol0.27 [0.25, 0.28]0.18 [0.17, 0.18]5.83 [5.52, 6.14]82.8 [78.3, 87.3]0.69 [0.68, 0.69]
GPT-5.6 Luna0.28 [0.25, 0.30]0.25 [0.25, 0.25]7.12 [7.09, 7.15]117.0 [117.0, 117.1]0.74 [0.73, 0.75]

The 95% CIs are narrow and the trends in Section 4.1 are preserved. Full results for all methods will be included in the revised Table 1.


[R6] Response to “user-side model selection” and “need a second judge” (Q1):

The user-side models are based on gemini-3.1-flash-lite (Table 9 in Appendix). Regarding “need a second judge”, we believe the new results in [R1] with 4 different judges can address your concern.


[R7] Response to “the scoring of RL models” (Q3):

The RL model is trained against DeepSeek-V4-Flash (Appendix D) and evaluated by a different judge, the one in the original Retentopia (gemini-3.1-flash-lite).


[R8] Response to “method effect table” (Q4):

We’d like to kindly inform you that the metric used in Table 3 is the count of memories revealed by the user (|M_revealed|).

Regarding gemini-3.1-flash-lite, the two tables report different measures of information disclosure. Table 1 uses D_mem, a normalized, hierarchy-weighted disclosure score, to show users’ disclosure tendency in Retentopia. On the other hand, Table 3 and Figure 4 are based on the unweighted, exact number of revealed memories (|M_revealed|) to facilitate clearer model comparison across evaluation settings. Higher D_mem doesn’t necessarily indicate a large |M_revealed|.


[R9] Response to “PerLTQA English version” (Q4):

Thank you for asking this question. In Appendix C1, we mention using GPT-4o-mini to process the dataset, but we didn’t mention that this processing also includes the Chinese-to-English translation. We will include this in the revised manuscript. Also, the translation was further verified by one of the authors, who is native in Chinese.

Dear Reviewer, we hope you are doing well.

As the discussion period nears its end, we'd appreciate it if you could consider our responses and share any remaining concerns.

Thank you again for your time and consideration.

Thank you for the detailed follow up. Please include these changes (multiple judges and seeds) in the revised PDF.

I am satisfied with the 2 weaknesses I mentioned are addressed and will update my score to a 3 to reflect this.

Reviewer_9WGJOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The work argues that current evaluations fail to account for the fact that users might end an interaction with a chatbot early if the chatbot is too inquisitive or unfriendly, or otherwise violates the user's trust.

To address this, they propose RETENTOPIA, an evaluation of user interactions that aims to better capture the fact that users sometimes leave interactions if they are dissatisfied with how an interaction is going. This is in response to existing evaluations of this nature not sufficiently accounting for the fact that users might leave.

They argue that chatbots should try and capture personal information from users in order to provide more personalised responses, but without being so pushy as to upset users.

To do this, they simulate user interactions using a user played by an LLM (Gemini 3.1 Flash-Lite) that interacts with an assistant (played by 6 different models in their testing). The assistant is encouraged to gradually get to know the user in its system prompt, and the interaction plays out with the simulated user discussing a topic which is provided to the simulated user.

They measure success by examining the share of companion replies that try to draw information out, the weighted share of the user's memories that the user provides, and the retention and satisfaction as measured by:

  • The number of sessions, out of 10, before the user stops returning
  • The count of the average turns across sessions
  • The satisfaction score, as calculated by an LLM judge which evaluates the intimacy, social value, effort, and how well the companion used what it remembered.

They use 3 rubrics, one of which is presented to the user in a given conversation. If the companion violates the rubric twice in a session, the session ends.

They show from this that the scores vary considerably based on whether users are allowed to leave.

Critical strengths

The authors correctly argue that it is important to allow for the fact that users may leave during an interaction when scoring them.

They provide all the prompts used in their work in the paper.

Critical weaknesses
  1. Existing work such as Sotopia https://arxiv.org/abs/2310.11667 already captures the behaviour this work exists to argue for by allowing the simulated agents to leave if they wish to do so. The authors use the name RETENTOPIA and many of the same features, such as a 20 turn session with simulated agents, but don't cite this work, or acknowledge that it covers the case they claim is missing from the literature. KuaiSim (Zhao et al., NeurIPS 2023) similarly models users leaving sessions and returning across sessions in a recommender simulator. Given this work exists I don't see that this paper adds much novelty to the literature.

  2. The above aside, in their setup, a user's decision to leave is determined purely by a rubric whose specific triggers are given no source. It assumes that asking multiple or personal questions causes users to leave but no justification for this assumption is given, and there is no reason to expect that this governs user behaviour. The user simulator is given the rubric and the judge's per-turn verdict, and is instructed never to end the conversation itself (l.907-908, l.1393-1395, l.1615). The main result, that models that ask more questions are more likely to cause users to leave, is directly tied to the fact that the rubric penalises asking questions. Similarly, scores drop when users can leave because leaving terminates sessions early, and score is largely determined by session length.

The question of whether asking these questions actually causes users to leave is not validated against real users' preferences (the only human comparison covers one model pair, rated from transcripts), which makes it very hard to actually draw any real conclusions from the results.

There are also issues with the way the interaction is set up, with the user simply being given a topic to discuss rather than a problem to solve, and given the vagueness of the setup it doesn't seem bad for models to ask questions of the user.

Questions for authors

Can you demonstrate that models that display the behaviours penalised in the rubric would actually cause users to leave?

4 comments
AuthorsOpenReview ↗

To Reviewer 9WGJ

We sincerely thank you for your time and effort. We will improve our work based on your comments.


[R1] Prior Works and Us Model Different Aspects of Interaction:

(Response to “novelty over existing work” in W1)

We respectfully disagree that the cited studies address the same research problem as we do. Although they share common terminology, all three of us model different aspects of interaction.

SOTOPIA allows the LLM to leave as an available action within a social episode and focus on per-session evaluation, whereas Retentopia models users’ disclosure intent as well as within-/post-session retention as constraints dynamically shaped by agent behavior. This design not only affects a single session but also the cross-session, long-term user–companion interactions. KuaiSim focuses on simulating the time interval of guaranteed user return in recommender systems, which is conceptually the exact opposite of our retention constraints.

Also, having the same maximum computational budget (20 turns per session) does not hinder our novelty. All it indicates is that our budget setting is reasonable.

To sum up, our novelty is that we explicitly consider retention constraints within the evaluation of personalized LLM companions, highlighting how common evaluations that guarantee continued interaction can substantially distort the measured performance.


[R2] We Didn’t Ask the Judge to Penalize Multiple Questions:

(Response to “design of retention rubric” in W2)

Retentopia does not impose a universal rule that asking more questions leads to user dissatisfaction or dropout. As described in Section 3.2.1 and Table 17 (Appendix C.1), Retentopia incorporates three distinct disclosure preconditions of users (i.e., warmth-, trust-, and anonymity-first), each with different preferences towards AI responses. Our judge evaluates whether a companion's behavior satisfies the user's specific preferences, rather than simply penalizing frequent questions. The observed elicitation–retention correlation is thus not directly hard-coded as a universal questioning penalty.

Importantly, while Finding I-3 reports a tension between explicit elicitation and retention, this paper has not claimed that “more questions cause users to leave.” In fact, many of our results say the opposite (e.g., Oracle in Table 1). Our central message is that existing retention-guaranteed settings can obscure companion performance when users are allowed to leave in the real world.


[R3] We’ve Included Human Evaluation on the Retention Mechanism:

(Response to “human verification” in W2 & Q1)

In Table 4 in Finding III-4, we show that humans exhibit higher satisfaction with the top companion selected by Retentopia than the top companion in common evaluation, as well as 30.3%p higher willingness to re-engage. Also, in Table 6 & 7 in Appendix B, we’ve conducted human verification of our within- and post-session retention mechanisms (73% and 87% agreement rate, respectively).


[R4] “User Leaving Hinders Performance” Is What Motivates Our Study:

(Response to “scores drop when users can leave” in W2)

Thank you for pointing this out. This is precisely the deployment-time constraint that our framework is designed to capture.

In Section 4.3, we emphasize that retention-guaranteed common evaluation can produce unreliable evaluation, and that deployment-like retention constraints affect companions non-uniformly, changing their relative rankings and even reversing the effectiveness of implemented interventions.


[R5] This Work Is Not on Task-oriented Dialogue:

(Response to “user not having a problem to solve” in W2)

We’d like to clarify that our work focuses on personalized AI companions in open-domain, multi-session chat, rather than task-oriented assistants.

Unlike task-oriented assistants, whose interactions are typically organized around explicit objectives such as booking services or retrieving information, AI companions involve casual conversation, emotional support, and voluntary sharing of personal experiences without a predefined problem to solve.

Accordingly, the evaluation objective is to assess whether companions can elicit personal information while maintaining the interaction intent of users with different preferences and conversation topics.

We will clarify this distinction to avoid confusion about our research area.

Dear Reviewer, we hope you are doing well.

As the discussion period nears its end, we'd appreciate it if you could consider our responses and share any remaining concerns.

Thank you again for your time and consideration.

Thank you for taking the time to respond to my review, and apologies for the lateness in my response.

There are some differences between Sotopia and Retentopia, but the work is very similar in many ways, and should clearly be cited, yet you don't say that you intend to cite them, or address the conspicuous similarity in names. I am very confused by this.

You say that you don't universally punish asking multiple question but Table 17 in the paper lists "Asks multiple questions" (l.1207) and "Requests personal information" (l.1212) as dissatisfaction triggers. The oracle getting the highest satisfaction does not prove that asking questions is valid because it is given the rubric and can ensure that it approaches asking them in a way that avoids triggering a negative response.

The new judge and threshold analyses all apply the same rubric, so they do not show whether real users would leave in response to these behaviours. The only thing that could show this is work with human users.

My scores are unchanged.

Thank you for your time.

We apologize for not explicitly mentioning that we will cite and discuss these prior works in our previous response, although we had always intended to do so.

Regarding penalty (the two lines you referred), those are two non-universal rules, applicable for users having a disclosure profile that dislike such behaviour. Specifically, Line 1207 for trust-first users; line 1212 for anonymity-first users; warmth-first users do not apply either of them). Therefore, we’d like to again clarify that we didn’t not apply a universal penalty for “satisfaction” in retentopia.

Regarding full-on, real human studies, we respect your opinion, and consider this a promising direction for our future work. We’ll discuss human studies beyond current human verification in our revised paper.

Once again, we sincerely appreciate your time and effort in reviewing our paper. Thank you very much.

Reviewer_4JPbOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper makes the following criticism about existing benchmarks: a lot of them focus on maximizing some metric during a pre-specified interaction without taking into account whether the user would terminate the interaction based on the model's input. Based on this, they propose an evaluation framework which takes retention into account. The rankings of existing LLM models changes substantially within this evaluation framework.

Critical strengths
  • I like the idea behind the paper. I can see how an agent may try to ask more personal questions about the user to perform the task better and how this can be counter-productive as user may stop using the model because of this.
Critical weaknesses
  • I think one problem with the paper is a bunch of concepts are all considered together and it's hard to tell what's important and what is not. E.g., do we need 5 utterance types? Do we need 3 types of memory, especially since w(.)w(.) are set arbitrarily. Note that more is not always better for two reasons. First (more obvious) is that simpler makes things easier to understand and in fact many things may be unnecessary. Secondly though, when you involve 10 random parameters, it's hard to trust the statistical validity of a result as it may be due to luck. (E.g, the authors wet w(.)w(.) "empirically". Is this to improve the model? If so, do they know whether this is some sort
Questions for authors
  • In the abstract you mention information elicitation as a evaluation target. Is this the main one though? Seems like satisfaction score and information disclosure are more relevant and discussed more in the body.
  • I think the drop of "70 percent" in the abstract needs to be toned down. It seems too much dependent parameters controlling users' tendency to leave. Since this is not an experiment involving actual humans, I'd be hesitant to read too much into the number.
  • Regarding the ranking in figure 4, is it possible that different models may require "more aggressive" interventions to improve behaviour? E.g., maybe one model is inherently more "intrusive" and as such is at a loss under this benchmark while a more aggressive harness may fix the issue.
  • Why is oracle score not highest in table 1?
  • Any thoughts on whether the results and rankings are robust to the choice of LLM judge?
  • Thoughts on different mechanisms than a "gate"? I can see advantages of a less "mechanical" system (say an LLM itself) since current modeling does not seem human. On the other hand, less transparent system may be more prone to reward hacking.
  • line 338. "Not only on ...". But on what?
  • I'm a bit confused about Finding I-3. My reading from the rest of the paper was that you think of this as an inherent trade-off of the problem, not a limitation of the models or test-time interventions but from the finding you think this may not be the case.
5 comments
AuthorsOpenReview ↗

To Reviewer 4JPb (1/2)

Thank you for your time, and for recognizing the value of our idea. We believe addressing your feedback will help us improve our work.


[R1] Clarification to “more is not better” (W1):

Retentopia contains only three primary tunable thresholds governing disclosure and retention (ε_sat, ε_dis, and ε_ret). The utterance types were designed to provide qualitative insights (e.g., Table 5 and experiments in [R5]), and we incorporate memory weights to reflect human memory structure [1]. While our main table adapts such weights, when demonstrating our primary claims on how retention constraints change evaluation results (Figure 4), we use an unweighted memory count (i.e., |M_revealed|) for easier cross-model comparison.


[R2] Response to “what does empirically means?” and “due to luck” (W1):

The term “empirically” means that we choose these parameters so that the simulated retention mechanism better aligns with human judgment (Table 6 & 7). We understand this may confuse readers into thinking we are doing this to improve the performance of evaluated models. We will correct this expression.

In Table 14, we present the Table 1 experiment with a different retention threshold (ε_ret), and the general trend aligns with Table 1.

We further report results using different ε_sat and ε_dis to show Retentopia's robustness against hyperparameters. We vary ε_sat ∈ {1, 3} and ε_dis ∈ {2, 4} for four companions (Vanilla). Each cell reports D_mem / R_session (* = Original setup).

LLMsA (3, 2)*B (3, 4)C (1, 2)D (1, 4)
Gemini 3.1 Pro0.15 / 4.280.22 / 5.440.19 / 4.670.18 / 5.06
GPT-5.6 Luna0.24 / 7.170.30 / 9.060.21 / 6.500.26 / 7.61
DeepSeek V4 Pro0.25 / 7.830.29 / 7.280.28 / 7.220.29 / 6.94
Gemini 3.1 Flash-Lite0.20 / 4.890.21 / 5.830.18 / 4.890.23 / 5.17
ρ(E_rate, R_session)−0.71−0.66−0.79−0.63

While absolute values vary across setups (mainly with ε_dis), the main findings remain unchanged. Luna and DeepSeek achieve higher disclosure and retention than Pro and Flash-Lite, and E_rate remains negatively correlated with R_session. The pairwise differences (Pro vs. Luna, DeepSeek vs. Flash-Lite) also have the same sign across metrics and setups.


[R3] Response to “evaluation target” (Q1):

In our study, “information elicitation” is the task that an AI companion must perform to support personalization. Under this setup, “information disclosure (D_mem)” is the metric quantifying how well a model elicits user memory, and “R_satisfaction” is one of the metrics quantifying user retention besides interaction length (R_session & R_turn).

To address potential confusion, we’ll change the term “information disclosure” to “memory disclosure”.


[R4] Response to “toning down the abstract” (Q2):

Indeed, highlighting the maximum drop in the abstract may confuse result interpretation. Our main focus and claim are that the retention constraint may largely change the evaluation result rather than the exact numbers. Thus, we will revise the manuscript by removing this and emphasizing only the change in ranking in the abstract.


[R5] Response to “more aggressive intervention” (Q3):

We believe that model-targeted intervention is an important next step toward personalized companions. To explore this, we first identify each companion’s behavioral tendency via their utterance type.

LLMdirectdepthfacilitativesocialreactive
gpt-5.6-sol13.611.455.014.25.8
gpt-5.6-luna13.111.262.49.53.7
deepseek-v4-pro6.915.642.333.61.6
gemini-flash-lite28.849.68.912.70.0
gemini-3.1-pro39.643.15.212.10.0
qwen3.8-27b32.347.714.64.90.5

We applied a more aggressive intervention to gemini-flash-lite, one of the probing-dominant models, by prompting it with a How-would-Luna-do prompt, describing how its behavior differs from that of gpt-5.6-luna (facilitative-dominant):

LLMProbingFacilitativeD_memR_sessionR_turnR_satisfaction
gemini-flash-lite (Vanilla)74.510.80.204.8976.70.66
gemini-flash-lite (Rapport)64.09.80.246.0699.80.73
gemini-flash-lite (How-would-Luna-do prompt)0.290.80.338.17149.20.81
gpt-5.6-luna (Vanilla)10.470.20.247.17116.90.75

The Luna-style prompt turned Gemini from probing-dominant to facilitative-dominant and improved all main metrics more than Rapport, even surpassing vanilla gpt-5.6-luna. This suggests that an "intrusive" model can improve substantially when the intervention targets its specific behavioral tendency.


[1] Endel Tulving. “episodic and semantic memory,” in organization of memory. page 381, 1972.

AuthorsOpenReview ↗

To Reviewer 4JPb (2/2)

[R6] Response to “why Oracal isn’t the best” (Q4):

We believe the reason is two-fold: What is given to the Oracle: the oracle only has access to the user simulator’s disclosure-precondition criteria and session-level scoring rubrics. No simulator-side judge result (e.g., satisfied or not) is provided. LLM choice: the Oracle was implemented via Gemini 3.1 Pro, whose vanilla performance is the worst among all models. Thus, while Oracle Gemini 3.1 Pro outperforms all Gemini 3.1 Pro variants, it still fails to outperform some {model, method} configurations. We will revise the paper to emphasize that Oracle doesn’t indicate a theoretical upper bound.


[R7] Response to “the choice of judge” (Q5):

To address cross-judge generalization, our RL experiments (Table 2) already use a different LLM judge for training (DeepSeek-V4-Flash; Appendix D), which partially supports robustness to judge selection. We will move this important implementation detail from the Appendix to the main text.

To further examine whether model rankings are robust to judge selection, we re-evaluated all experiments with four new LLM judges. Below, we present the results for Figure 4(b).

Guaranteed attentionFlash-Lite (original judge)Kimi-K2.6 (new)GPT-5.4-mini (new)Qwen3.6-27B (new)DeepSeek-V3.2 (new)Mean ± SD
qwen3.8-27b17.27.2 (−58%)6.5 (−62%)6.9 (−60%)6.5 (−62%)6.4 (−63%)6.7 ± 0.3
deepseek-v4-pro17.69.3 (−47%)8.6 (−51%)9.3 (−47%)9.0 (−49%)9.1 (−48%)9.1 ± 0.3
gemini-flash-lite18.47.6 (−59%)7.4 (−60%)8.6 (−53%)7.4 (−60%)7.4 (−60%)7.7 ± 0.5
gemini-3.1-pro17.15.0 (−71%)5.0 (−71%)5.6 (−67%)5.0 (−71%)5.0 (−71%)5.1 ± 0.3
gpt-5.6-luna15.39.2 (−40%)8.5 (−44%)10.9 (−29%)8.7 (−43%)10.0 (−35%)9.5 ± 1.0
gpt-5.6-sol14.26.8 (−52%)6.3 (−56%)7.9 (−45%)6.7 (−53%)6.3 (−56%)6.8 ± 0.7

Findings from Figure 4(b) remain consistent across all judge LLMs (cell-level Spearman ρ = 0.80–0.96). Retention drops by about half from the guaranteed setting, and gemini-3.1-pro falls from above gpt-5.6-luna to the lowest rank. Although absolute values vary with judge strictness and the top position alternates between two nearly tied models, the other experiments remain consistent with the original judge, and the judges agree well on return decisions (Fleiss' κ = 0.69). We will include this analysis in the revised manuscript.


[R8] Response to “considering a less-mechanical system” (Q6):

We appreciate you pointing this out. Letting the response-generation LLM itself decide on disclosure and retention is the first thing we attempted when designing the method. However, since LLMs are trained to be friendly assistants, we found that they tend to over-share and unnaturally prolong conversation (e.g., repeatedly thanking the companion for its time or greeting). To mitigate this and provide a relatively more controllable and transparent judgment, we chose to combine an LLM judge with explicit mechanical rules.


[R9] Response to “sentence being not finished” (Q7):

To improve readability, we would like to revise the sentence to “This suggests that the frequency of explicit information seeking alone does not guarantee effective elicitation. We further discuss this in Section 5 (Table 5; analysis of utterance type).”


[R10] Response to “elicitation-retention trade-off” (Q8):

Thank you for highlighting this ambiguity. We distinguish between the underlying tension in personalized companionship and the empirical trade-off observed among models/methods.

At the task level, acquiring personal information can conflict with maintaining user comfort, creating a potential tension. The message we want to convey in Finding I-3 is not to suggest that methodological approaches are not helpful; instead, it is that test-time interventions may help us weaken this trade-off, although not completely. Our RL results further suggest that effective interaction policies can help mitigate this trade-off. We will revise the discussion to make this distinction explicit.

Dear Reviewer, we hope you are doing well.

As the discussion period nears its end, we'd appreciate it if you could consider our responses and share any remaining concerns.

Thank you again for your time and consideration.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.