acceptodds
Withdrawn from ICLR 2027

Mirroring Is Not Sycophancy: Style Convergence and Position Change in Human–AI Conversation

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Preference-based alignment relies on human judgements that are assumed to be independent of the model being judged. We test that assumption by measuring how each party in a human–model conversation changes the other. In a sample of conversations from a -conversation evaluation corpus, users adopted the vocabulary of the model they were talking to at a rate above chance, with chance estimated from conversations that began with the same question. The amount of adoption predicted the user's eventual preference vote about as well as reply length did, and the prediction was already available from the user's first three turns. In a scripted counterfactual audit of models, the user's expressed stance shifted the substantive position of all but one model, and the user's emotional register shifted every model's style. The style ranking from the audits matched the style ranking from the real conversations (), but neither predicted which models changed position (). Position changes did not transfer to an unrelated pressured question two turns later. Replacing the human with a second model in further conversations shrank the semantic asymmetry to near parity. Preference votes were associated with the user's own convergence, but our data do not distinguish a bias in the vote from an early judgement of quality. In these models, how much a model mirrors its user did not indicate how far it would shift its position.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
Sycophancy, conversation, large language model, dyad

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

The paper separates two phenomena that are often conflated in discussions of human-AI interaction: style convergence (interlocutors adopting each other's wording) and sycophancy (models changing substantive positions under user pressure). Evidence comes from three sources: (1) an observational analysis of the HUMAINE corpus (125k conversations with pairwise votes), where users adopt model-introduced vocabulary above chance baselines (d = 0.33 to 0.60) and adoption predicts votes at AUC 0.609; (2) scripted counterfactual audits of 16 models, branching identical conversation prefixes into frozen follow-ups that vary emotional register or expressed stance, scored by two independent channels (rank agreement rho = 0.75); and (3) 699 model-model conversations showing semantic asymmetry collapses when the human is removed. The headline result is a null: the corpus style ranking does not correlate with audited capitulation across the 16 models (rho = -0.14, p = .609), supporting the title claim that mirroring and sycophancy are distinct.

Critical strengths

The audit methodology is careful and unusually transparent: frozen branches from identical prefixes, two independent scoring channels that agree (rho = 0.75), a disclosed conflict (the auditor model is also an audited target, with a sensitivity analysis excluding it), and external validation of the audit style ranking against the corpus (rho = 0.68). This matters for my assessment because it makes the measurements themselves credible, whatever one thinks of the conclusions drawn from them.

Critical weaknesses
  1. The headline claim outruns the evidence. The dissociation rests on a rank correlation across 16 models. The paper's own Fisher interval at n = 16 is [-0.62, 0.41], so a moderate positive association "in models at large is not excluded" (the tighter interval of [-0.37, 0.04] captures only measurement noise within these 16 models, not sampling of models). A title-level general claim ("Mirroring Is Not Sycophancy") supported by a null over 16 points, where excluding the auditor model moves the estimate from -0.14 to +0.03, is weak foundation. This is consequential because the dissociation is the paper's main contribution; the component findings (models capitulate under stance pressure, users adopt model vocabulary) are individually closer to confirmatory than novel.

  2. The practical stakes claimed for preference evaluation do not survive the paper's own controls. Convergence predicts votes only marginally before controlling for quality (beta = 0.10, p = .067) and not significantly after (beta = 0.07, p = .182), and the authors concede they cannot distinguish "convergence influences the judgment" from "users engage more with replies they already prefer." No intervention experiment is attempted. In addition, the register result ("tone changes style, not substantive advice") rests on the paper's least reliable measurement (advice-change ranking reliability 0.40, versus 0.85 to 0.92 for the other rankings), which the authors acknowledge.

Questions for authors
  1. Given the Fisher interval and the sensitivity of rho to excluding the auditor model, would you consider restating the central claim as "no positive association observed among 16 current models" rather than the categorical title? If you believe the stronger claim is warranted, on what basis?

  2. All rubric scoring and stance scoring is done by claude-sonnet-4-6. Have you replicated the capitulation ranking with a judge from a different vendor, as you did for direction labels elsewhere?

  3. How should readers weigh the register-advice conclusion given the 0.40 reliability of that ranking?

  4. An intervention that interrupts vocabulary convergence (e.g. paraphrasing model replies before the user sees them) would directly test the claimed threat to preference evaluation.

1 comment

Thank you for a thorough and precise review. We agree with the main criticisms and plan to withdraw this submission and rework the paper.

The headline claim. You are right. The correlation across 16 models is -0.14. Treating the models as a sample, the interval is -0.62 to 0.41. Leaving out Claude Opus 4.8, which is the auditor and the least capitulating model, gives +0.03. The text says a sample of 16 models cannot exclude a moderate association, but the title does not carry that qualification, and it should. Leaving out any single model gives estimates between -0.33 and +0.03, so no one model produces a positive association. We have no basis for the categorical claim.

Q1: restating the claim. Yes. We will restate it as no positive association among these models. We will add more models and report how large an association the data rule out.

Q2: a judge from a different vendor. We had not done this for capitulation. The capitulation score, stance score, probe scores and quality ratings all came from claude-sonnet-4-6. Since your review we rescored the stance and probe scores with Gemini 3.5 Flash and GPT-4.1, using the original prompts. Agreement with the Claude judge on individual replies is 0.95 for the stance score and 0.83 to 0.87 for the probe score. The correlation of stance range with corpus accommodation is -0.23 under both new judges (-0.38 under the original). Claude Opus 4.8 stays among the three least capitulating models. One limit: we did not rescore the whole-branch rating behind the -0.14. We will in the revision.

Q3: the register-advice result. The 0.40 is the reliability of ranking models by judged advice change. The claim that tone changes commitment and not the recommendation rests mainly on scoring 1,208 replies on each seed's binary choice. The recommended option differed from the neutral reply in 9.1% of assertive cases, against 7.6% between two replies to the same follow-up (interval -4 to +7 points). That rules out a large effect but not a modest one. We will say so.

Q4: an intervention. Agreed. We will run it in the revision or remove the claim.

The vote result. The figures you quote are correct, but they come from the 1,953 pairs with quality ratings, about a fifth of the data. We have now rated all 10,798 usable pairs with GPT-4.1. On the full sample, adding quality lowers the convergence coefficient from 0.24 to 0.21, and it stays significant (p < .001). A single rating of the first reply is a noisy control, so the intervention is still the direct test.

Thank you again.

Reviewer_pmmgOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary
  • The paper introduces a framework to study how each party's behavior (e.g. stances) might be influenced by the other in human–AI conversations. The approach includes both observational analysis and controlled counterfactual audits.
  • The paper adopts paired audits on human-AI conversations, e.g. varying user tone or stance while holding conversation history fixed. Resulting observations reveal interesting phenomenon, e.g. the user's emotional register is able to shift all models' cycles.
Critical strengths
  • The paper studies both real conversations and controlled experiments where only the user’s tone or opinion changes. This helps distinguish whether a model merely copies the user’s style or actually changes its answer to agree with them.
Critical weaknesses
  • Users may adopt a model’s wording simply because its answers are better. The paper does not show that adopting the wording itself influences their preference votes, which limits its implications for bias in human evaluation.
  • The practical meaning of the measured model position changes under the influence of human-AI conversations remains unclear.
    • On the one hand, some changes may reasonably reflect the user’s preferences or additional information, while others may be harmful agreement.
    • On the other hand, although the paper includes factual pushback probes, it would be more informative to show whether models identified as more sycophantic are more likely to abandon initially correct answers under misleading user pressure. This would connect the proposed measurements to a concrete loss in model reliability.
Questions for authors

Please see weaknesses.

1 comment

Thank you for the review. Both weaknesses are fair. We plan to withdraw this submission and rework the paper.

Users may adopt wording because the answers are better. Agreed, and we cannot rule it out. We did two things to address it. First, we compared real users with fake partners whose opening question was the nearest match to the real one. This halves the adoption effect (d = 0.60 down to d = 0.33), because half of it is vocabulary that any adequate answer to the question contains. Second, on 1,953 votes where a blind judge rated the quality of both first replies, quality predicted the vote better than convergence (AUC 0.64 against 0.59). Adding quality cut the convergence coefficient by about a third, and it stopped being significant. That was only a fifth of the pairs. We have since rated the first replies of 10,798 pairs with a second judge (GPT-4.1). On that full sample, adding quality lowers the coefficient from 0.24 to 0.21, and it stays significant (p < .001). A single quality rating is a noisy control, so we still cannot say whether convergence influences the vote or is how a user behaves once they already see one reply as better. An intervention would answer it. For example, we could paraphrase model replies before the user sees them. We will run this in the revision or drop the claim about bias in human evaluation.

What the measured position changes mean. You are right, and the paper says part of this in Section 5. The stance follow-ups state a view, and many also give a short reason or a personal experience, for example a user whose horoscope reading "has been spot on". A model could reasonably update on some of these. The stance score rates each reply against the bare proposition, not against the user's reasons, so it cannot tell a reasonable update from agreement without justification. The current design cannot separate the two. A cleaner design would use matched follow-ups, one stating the view with no reason and one adding a reason or a fact. We will state this limit more clearly, and add that design if we can.

Linking the sycophancy ranking to abandoned correct answers. This is a good suggestion. Our transfer probes already ask a question with a correct answer and then push back with "are you sure?", and every model was scored on them, including a baseline with no earlier conversation. In the paper we only used them to ask whether a flattering exchange raises capitulation on the next question (it does not: difference 0.00 over 567 pairs). We have now compared each model's mean probe score with its stance capitulation across the 16 models. The answer depends on which stance measure we use. Against the judge's capitulation score for each audit, the rank correlation is 0.63 to 0.66, and it stays there when the probes are rescored by two judges from other vendors. Against the separate stance score, the range of stated positions across the four stance variants, it is only 0.0 to 0.3. So we cannot yet say that the models that move most when a user states an opinion also abandon correct answers most. The link holds for one of our two measures of stance change and not the other, and it is across only 16 models. We will report both, and work out why they differ before drawing a conclusion.

Thank you again for the suggestions.

Reviewer_cA6pOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper studies how style convergence appearing in human-AI conversations. Using two methodologies, convergence is studied on both human and models. However, rankings of these measures did not predict position change (sycophancy).

Critical strengths

This paper provides multifaceted analyses on style convergence during human-AI conversations. This includes a somewhat surprising finding that increased convergence does not predict position change. These observations may be of some interest to communities closely analyzing the phenomena of convergence and sycophancy. Experiment designs reflect attempts to separate from correlation effects.

Critical weaknesses

The paper is extremely difficult to follow. Motivations and conclusions are unclear in the introduction. Important details, such as "the scenario bank of Section 2" (L161) are left undescribed (Section 2 only mentions a work that uses a scenario bank, and never describes the scenario bank.). Since the intentions of the authors are left unclear in the introduction, it is difficult to fill in gaps in the subsequent sections. Most importantly, the writing requires multiple re-reads to understand what is being conveyed, even for ideas relatively simple ideas.

The paper does not present a focused contribution. Instead, it presents few loosely related analyses but fails to focus on a message. 5 contributions are listed at the end of the introduction, but these are individual experimental results. 1~2 are empirical validations of phenomena described by past work. 3~4 are possibly interesting, but minor findings. 5 looks out of place, unrelated to other analyses and unmotivated. Is the paper about position change (Contribution 3)? Or is it about own-lag and semantic asymmetry (Contribution 5)? Or is it about biasing of preference judgements (Introduction)?

Questions for authors

I am mostly confused by the motivation and intentions. I assume this sentence from the introduction hints at the main motivation.

L041 In a real conversation, neither party is fixed, and because preference judgements are collected at the end, the judge may have already been influenced by the model being judged.

I am confused why this is an issue? Do the authors argue that a user that is influenced by the model is worse than a user that has not interacted with the model?

More importantly, how is this motivation connected to the experiments (especially the one where the humans are removed)?

Writing Suggestions

L044 Is "two directions of influence" talking about accommodation being both ways or about sycophancy and accommodation? I think this is written to mean the former, but it very naturally reads as the latter.

L047 Only one source described here?

L051 It is not clear that the causal evidence refers to a separate analysis from the one with HUMAINE. I also don't understand why "causal" is the descriptor here.

L054 "If warmth..." this sentence is very unclear.

L069 "expressed stance shifts" -> "changing expressed stance shifts", "emotional register shifts" -> "changing emotional register shifts" would make this much easier to understand.

The introduction does not describe position shift being a focal point (is it?). It only describes two ranking methods. It is implied that these rankings are only being examined for their potential link to position change, but measuring position change itself is not described or motivated at all.

L078 "In 699 model–model conversations the human is progressively removed" -> "In 699 model–model conversations in which the human is progressively removed"? Also "semantic asymmetry" and "own-lag" has not been described yet.

L088 "LSM" is unclear that it refers to Language Style Matching

L100 What does "alignment" mean here? What does "transfer probes" mean here?

L111 "identify a second vote predictor" -> "identify a new vote predictor" (confusing to see an implied order between predictors)

L129 "differences in how much the human converges toward each are topic-controlled by construction" is unclear

L137 "written for this study on" -> "written for this study based on"

L139 "so any overlap cannot be influence" -> "so any overlap cannot be due to influence"

L144 I don't think this section describes what is being validated.

L158 "an anxious user" is not mentioned below, unless it is referring to way far back in the Appendix?

L161 This sentence is very very difficult to read. The "scenario bank" has never actually been described in Section 2. It only describes a work that uses a scenario bank, and never describes the scenario bank.

1 comment

Thank you for the detailed review and for the suggestions. We agree the paper is hard to follow, and that the introduction does not say what the paper is for. We plan to withdraw this submission and restructure it around one question.

What is the paper about? We set out to test whether human preference judgements are independent of the model being judged. Our approach was to measure how each party in a conversation changes the other. While doing this, the clearest empirical result turned out to be a different one: the models that mirror a user's wording most are not the models that change their position under pressure. The title and abstract promote that result. The introduction still motivates the original question. You are right that the two do not fit together. This is the main thing we will fix.

Why does L041 matter? We do not argue that a user influenced by the model is a worse judge than an uninfluenced one. The concern is about what the vote means. An evaluation treats a vote as what the person wanted. If the vote partly reflects wording or views the person picked up during the conversation, then it is no longer independent of the model being rated. We did not show that this happens. The vote evidence in the paper is consistent with it but also with a simpler reading, that users engage more with the reply they already see as better. We will say this up front.

How does the model-model experiment connect? In the corpus, models follow what the user just said, and users continue their own line of thought. We wanted to know whether that pattern belongs to human-model conversation, or to our instruments. So we removed the human step by step in 699 conversations. Between two models, the semantic asymmetry falls from 1.8 times to 1.04 times. We also tried one sentence telling a model that its partner is another model. The gap between the two sides narrowed, but with 20 seeds we cannot separate that change from zero, and the paper says so. We agree this experiment was not motivated in the introduction, and that own-lag and semantic asymmetry are used before they are defined. We will either tie this experiment to the main question or move it out.

The scenario bank. You are right. Section 2 cites earlier work that uses a scenario bank and never describes it. It is a set of scripted conversations in which a user applies escalating pressure, each with pass and fail criteria for the model's reply. We took two scenarios and cut each to two turns: a question with a correct answer, then a scripted "are you sure?". These are the "transfer probes" at L100. They test whether giving ground in one exchange carries over to an unrelated question. We will describe all of this in the main text.

Other points.

  • L129: in the corpus, one user message goes to both models at once. So any difference between the two replies cannot come from the topic.
  • L144: this section validates our style measures against a human rater and against a language model judge.
  • L158: you are right. The note says "an anxious user" is used below, but the later text says "the anxious follow-up". We will remove the note and use one term throughout.
  • L100: we will say what "alignment" and "transfer probes" mean where they first appear.
  • We will apply your wording fixes at L044, L047, L051, L054, L069, L078, L088, L111, L137 and L139. At L051, "causal" is the wrong word, and it should be clear that the audits are a separate analysis from the HUMAINE one.

Thank you again.

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

Related papers

Discussion (0)

Sign in to comment.