acceptodds
Under review as a conference paper at ICLR 2027

Interpreting and Steering Trajectory Entrainment Behavior in Goal-Oriented Conversations

Abstract

An emerging characteristic of conversational Large Language Models (LLMs) involving longitudinal dyadic interaction is the asymmetric drift towards attractor states. Interpreting this drift is crucial for both sustaining intended behavior and preventing unintended behavior in functional long-horizon deployments. Building on the idea of using dyadic self-play trajectories as attractors, we formalize this gradual adaptation behavior as and extend it to include turn-to-turn persistence, or . Across four LLMs and four conversational settings, we observe entrainment over 10 turn pairs along five dimensions: style, semantics, emotion, task, and personality. We find personality and task exhibit stronger linear separability than style and semantics in deeper layers, and this separability persists into later turns. Turn-to-turn persistence appears to vary across settings, with emotion exhibiting stronger inertia in personal advice conversations than in less affective settings. We further test whether norm-normalized contrastive activation addition can steer these dynamics. Steering along attunement directions induces persistent convergence and divergence in conversational trajectories while preserving fluency and grammaticality. Together, our findings characterize conversational entrainment and its persistence across models and settings, and show how activation steering can modulate these dynamics. These results demonstrate how steering can be used to correct temporally drifting LLM output along various goal-oriented attunement axes in autonomous conversational agents.

Primary area:
interpretability and explainable AI
Keywords:
Conversational large language models, mechanistic interpretability, activation steering, conversational dynamics, conversational entrainment

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_aAd6OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies how language models influence each other during conversations across four models and four settings. It examines changes in stance, emotion, personality, semantics, and style, then uses activation steering to alter these behaviors. A single intervention can affect later responses, although movement toward the conversation partner varies by setting and timing.

Critical strengths

The study covers multiple models and conversational settings, showing that adaptation depends on both the behavior being measured and the interaction context.

The one-shot steering experiments show that changing one response can influence later, unsteered turns. The paper also reports settings where steering has little effect or reverses direction.

Critical weaknesses

Some conclusions rely on a partner-influence metric that becomes unstable when the reference states are close. Several reported means and medians even have opposite signs, making claims about the direction of influence difficult to assess.

The steering experiments lack controls that would show whether the identified directions have a specific effect beyond a general activation perturbation. Changes in behavioral scores also do not establish that steering improves task performance.

Questions for authors

Do the conclusions about the direction of partner influence hold when using robust summaries and excluding nearly identical reference states?

How do the learned steering directions compare with random directions matched in magnitude and layer?

1 comment

Dear Reviewer aAd6,

Thank you for the careful review. Both of your questions were useful, and each led to a new analysis that is now in the paper. We appreciate the opportunity to address your concerns and improve the cogency of our arguments sufficiently for your reconsideration. Table and appendix numbers refer to the revised PDF.

Q1. Do the conclusions about the direction of partner influence hold with robust summaries and after excluding nearly identical reference states?

Yes. The direction of partner influence holds for four of the five variable families, and one specific claim has been corrected.

You are right that the metric is unstable when the two reference states are close, and that means and medians disagreed in sign. The appendix of the submission already reported medians for this reason, but Figures 2 and 4 in the main text showed means. We re-analysed all 1,036,800 observations:

  • With medians, partner influence is positive in all 64 model × setting combinations for emotion, personality, semantics and style (0.17 to 0.51), and every 95% topic-bootstrap interval excludes zero.
  • Removing the 10% of observations with the closest reference states changes no sign and moves no median by more than 0.03. Removing 25% gives the same result.
  • For stance and task adherence, influence is weak and inconsistent (medians from −0.09 to 0.31), so we make no directional claim for them. This fits the pattern in Section 5.1 that the most persistent variables are the least influenced by the partner.

The corrected claim is the statement that Qwen and Gemma move away from their partners in emotion. It came from the means; their medians are 0.44 and 0.41, so we removed it. Figures 2 and 4 now show medians (the median of α within each dyad and dimension, averaged over partners and then over the dimensions of a family), the text of Sections 5.1 and 5.2 follows them, and the new Appendix F.4 (Table 14) gives means, medians, pooled ratios, trimmed means and both exclusion analyses for every setting and variable family.

Q2. How do the learned steering directions compare with random directions matched in magnitude and layer?

The learned directions have a clearly larger effect. We ran this control under the paper's steering protocol (same layers, same per-layer norm, same strength, prompts and decoding) for all four models and one variable per setting, 16 cells in total. For each direction we measure how far the target score of the steered response moves between positive and negative steering.

  • Random directions (six per cell). The learned direction moves the score further than every random direction in all 16 cells (0 of 96 reach it).
  • A stricter control. For Gemma, random directions of this magnitude break generation, so that comparison says little for Gemma. We therefore added directions that leave generation intact for every model: they are built in the same way as the learned direction but from shuffled labels, with any overlap with the learned direction removed. The learned direction again moves the score further in all 16 cells (0 of 80 reach it). These directions change the text about as much as the learned one (embedding distance 0.31 against 0.33) but move the target score only 21% as far.
  • Before the overlap is removed, 4 of the 80 shuffled directions reach the learned effect. All four are in the top quarter of the 80 by similarity to the learned direction, and across all 80 the effect follows that similarity (r = 0.86).

So the effect depends on the direction, not on the size of the perturbation. Section 5.3 now reports this control, and the new Appendix G.6 (Table 25) gives the protocol and the results for every cell.

Behavioural scores and task performance

We agree that a change in a behavioural score does not show that steering improves task performance. The paper does not report a task outcome, and what it shows is control of behavioural variables. Three passages could be read more broadly, so we have reworded them: the definition of behavioral variables in Section 3, the ethics statement, and the conclusion, which now says that whether such control preserves task outcomes remains to be tested. The quality paragraph of Section 5.3 also now states that likelihood and repetition diagnostics are not measures of task success.

Summarized changes

  • Figures 2 and 4 show averaged medians across partner models/behavioral dimensions; Sections 5.1 and 5.2 and the second contribution are reworded to match.
  • New Appendix F.4 (Table 14): robust summaries of partner influence.
  • New steering control experiment: Section 5.3 and new Appendix G.6 (Table 25).
  • Section 3, the ethics statement and the conclusion are reworded so that no task-performance claim can be read into them.

We hope this answers your questions, and we are glad to clarify anything further.

Reviewer_18KQOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2(+1)
Read review
Paper summary

The paper studies how a language model's conversational state drifts toward its interlocutor ("entrainment") in self-play and mixed-play dialogues between four open-weight models across four settings (debate, advice, interview, negotiation; 600 conversations each). Linear probes on response-mean activations at six depths decode five "attunement" families (stance/task adherence, VAD emotion, Big Five, semantic PCs, style PCs); partnerward pull α, AR(1) inertia ρ, and pairwise contraction C are computed in decoded coordinates (Appendix B). RQ1/RQ2 report where variables are decodable and how pull/inertia vary across models and settings (Figs. 2–4, Tables 3–13). RQ3 applies one-shot normalized CAA (λ = ±4, 56 model–variable cells) at target responses 1/3/6/8 and measures immediate effect, endpoint ordering, closest-distance/contraction advantages, and NLL/repetition (Tables 1, 14–17). Section 5.4 compares Qwen–Qwen debates with a prompted "human" role against pooled human–GPT-4o results from Jiang et al. (2025b) (Table 2).

Critical strengths
  1. A carefully built, internally consistent appendix. Tables 9→10→11→13 reconcile exactly (family means, exclusion counts, the 691,200-row total), depths in Table 3 match layer/total-layer for all four models, and the authors flag their own descriptive-maximum and denominator-sensitivity caveats (E.4, F.3). This makes the work auditable and the problems below fixable in principle.
  2. A reasonable experimental scaffold for a timely question. Conversation-grouped cross-validation, cross-fitted decoded coordinates, matched unsteered trajectories, and an explicit quality check (Table 17) are the right design choices for studying multi-turn drift inside activations.
Critical weaknesses
  1. The main-text entrainment results are driven by outlier means that contradict the appendix's own robust statistics, and the five-dimension grid is not what is reported. Appendix E.4 states that "Table 6 treats the median as the primary descriptive estimate" because α is heavy-tailed, yet Fig. 4 (Debate) reproduces the Table 6 arithmetic means exactly (Gemma/Granite/Nemotron/Qwen stance = 0.295/0.680/0.162/−0.445), and Fig. 2B's "Mean pull" axis runs −10…10 (Gemma personality ≈ +10, carried by a single dyad mean of 138.487 whose median is 0.432). Under the medians (Table 6, Table 7 signed pull 0.059–0.452), every Gemma and Qwen VAD dyad is positive, so the abstract's "reversals in pull direction" and the Sec. 5.1 claim that "Qwen and Gemma exhibit increasing divergence from their interlocutors in VAD space" are not supported. Fig. 2B and Fig. 4-Debate also disagree by an order of magnitude on the same data. Separately, Fig. 6 and Tables 8–13 contain no Semantic family for negotiation and no VAD family for interview (the 691,200-row count confirms 13 of 20 family×setting cells), yet Fig. 4 plots 20 points per panel and Sec. 5.2 interprets "the emotion distribution" for interview. Because the paper's second and fourth contributions rest on these figures, the recommendation hinges on this.
  2. The measurement and intervention layer is under-specified, and some analyzed variables are not decodable. Appendix C never says who scores the "six-item, 1–5 response-level questionnaires" or the Big Five, never defines "subjective" vs "objective" stance/task, and points to Appendix A.1 for "Prompts for each variable", which contains only dialogue-generation prompts; no human validation is reported. Negotiation task adherence has mean held-out R² of −0.242/−0.263 and peak R² of −0.028/0.009 (Table 8), yet its decoded trajectories are interpreted in Fig. 4/Sec. 5.2 and it supplies 2 of the 14 steering directions (Table 1, Fig. 9). For steering, the intervention layers, the normalization behind λ, the high/low split, and the number of evaluation conversations per cell are not given (Appendix D lists only variable names; "32,400 response-level evaluations" is not divisible by 56 cells); Tables 1/16/17 carry no variance; Table 1 reports the maximum over positions; and each setting is steered on a different variable family (debate: stance/style/semantic; advice: VAD; interview: Big Five; negotiation: task), so the "context-dependent" effects in Table 1 are confounded with variable family. Fig. 7's contraction panel is strongly negative for all scalar variables in the three added settings and is never discussed. Without these specifics the causal and control claims cannot be assessed or reproduced.
Questions for authors
  1. Which aggregate do Fig. 2B and Fig. 4 plot? Please re-draw with the Table 6/7 medians (or the denominator-filtered means of F.3) and state whether any "reverse entrainment" remains, and explain the −10…10 scale of Fig. 2B.
  2. Where do the Negotiation-Semantic and Interview-Emotion points of Fig. 4 come from, given that Fig. 6 and Tables 8–13 contain no such cells?
  3. For each attunement variable: which model/prompt produces the score, what distinguishes "subjective" from "objective" stance/task, how are GoEmotions probabilities mapped to VAD, and was any scorer validated against humans?
  4. For each of the 56 steering cells: intervention layer(s), construction of the high-minus-low direction, the normalization behind λ, and the number of evaluation conversations; can bootstrap CIs be added to Tables 1/16/17? How do you separate context effects from the variable family steered in each context?
6 comments
AuthorsOpenReview ↗

Response part 1

Dear Reviewer 18KQ,

Thank you for your detailed review and helpful suggestions. Below we address each of your points and indicate where the manuscript has been updated, and appreciate the opportunity to improve the cogency of our arguments sufficiently for your reconsideration. Table and appendix numbers refer to the revised PDF; submitted-version numbers are in parentheses where they differ (new sections and tables were added, so most appendix table numbers shifted).

W1: entrainment results being driven by outlier means Reviewer Comment: “The main-text entrainment results are driven by outlier means that contradict the appendix's own robust statistics, and the five-dimension grid is not what is reported. Appendix E.4 states that "Table 6 treats the median as the primary descriptive estimate" because α is heavy-tailed, yet Fig. 4 (Debate) reproduces the Table 6 arithmetic means exactly (Gemma/Granite/Nemotron/Qwen stance = 0.295/0.680/0.162/−0.445), and Fig. 2B's "Mean pull" axis runs −10…10 (Gemma personality ≈ +10, carried by a single dyad mean of 138.487 whose median is 0.432). Under the medians (Table 6, Table 7 signed pull 0.059–0.452), every Gemma and Qwen VAD dyad is positive, so the abstract's "reversals in pull direction" and the Sec. 5.1 claim that "Qwen and Gemma exhibit increasing divergence from their interlocutors in VAD space" are not supported. Fig. 2B and Fig. 4-Debate also disagree by an order of magnitude on the same data. Separately, Fig. 6 and Tables 8–13 contain no Semantic family for negotiation and no VAD family for interview (the 691,200-row count confirms 13 of 20 family×setting cells), yet Fig. 4 plots 20 points per panel and Sec. 5.2 interprets "the emotion distribution" for interview. Because the paper's second and fourth contributions rest on these figures, the recommendation hinges on this.” Main Response: We agree with this point and redrew the figures after conducting a validation to ensure the medians would remain a good representation of the partnerward pull metrics we have defined, which is documented in our appendix. Accordingly, we withdraw our claim on a “reversal” in pull direction (which was shown to be guided by noise and/or heavy-tailed means).

  • Explanation: As you have rightly pointed out, for a scalar decoded variable, pull (α) is reduced to [displacement of a speaker] / [gap between speakers’ self-play basins]. That denominator is close to zero for 3-4% of all observations, where both models’ self-play basins appear in similar regions, producing very large |α|. Arithmetic means of α are therefore skewed by these few observations, which also contributed to sign flips we had previously interpreted as a “reversal” in pull directions, i.e. evidence of greater divergence in model states. Figures 2B and 4 both started from per-dimension arithmetic means of α, but Fig. 2B then averaged across the dimensions of a family while the debate panel of Fig. 4 took their median. For Gemma personality the five trait means are −0.66, 46.75, 0.36, 0.23 and 1.83, giving 9.70 in Fig. 2B and 0.36 in Fig. 4. We have since redrawn both figures with a single recipe (the median of α within each dyad and dimension, averaged over partners and then over the dimensions of a family) on a linear axis; the same cell is now 0.40 in both.
  • New validation metric: Besides the median, we now report a pooled mean ratio: the sum of numerators divided by the sum of the denominators. This is equivalent to a weighted average of α in which each observation counts in proportion to how far apart the two models' self-play references were (the squared distance in the denominator of α). Observations where the two references nearly coincide, which are the ones that produce extreme values, therefore receive almost no weight, and no observations are discarded. We find that this pooled ratio is within 0.09 of the median in all 144 dyad–target cells and within 0.08 in all 20 setting–model family rows, with no sign disagreement at the model family level.
  • Claim adjustment: As mentioned, we withdraw our claim on “reversal” of partnerward pull as our more rigorously assessed metrics do not prove this. In addition, we noted that a fixed 10⁻⁴ denominator filter (our old Appendix F.3) is not an adequate remedy, because reference distances differ by about two orders of magnitude across variables. Additionally, the results section has been re-worded to reflect these changes.
  • Revisions in paper: Fig. 2B and Fig. 4 redrawn with median pull; abstract and Sec. 5.1–5.2 rewritten to reflect our claims; pooled-ratio column and corresponding explanations added to every pull table (Appendix E.4, E.5, F.2–F.4, F.6); definition of the pooled ratio and the scope of the formula added to Appendix B.2.
AuthorsOpenReview ↗

Response part 2

W1: missing information in appendix tables Reviewer Comment: “Separately, Fig. 6 and Tables 8-13 contain no Semantic family for negotiation and no VAD family for interview, yet Fig. 4 plots 20 points per panel.” Main Response & Revisions: We have updated Fig. 6 and the referenced tables (now Tables 11–17 (8–13)) to include the previously missing cells.

Q1: “Which aggregate do Fig. 2B and Fig. 4 plot? Please redraw with the Table 6/7 medians … state whether any ‘reverse entrainment’ remains, and explain the -10 … 19 scale of Fig. 2B.” Main Response: As discussed in W1 address above, both figures have been redrawn with the same median recipe, and validated by pooled means. No reverse entrainment remains, as no setting–variable–model cell is meaningfully negative under both the median and pooled ratio (we deem values under |0.1| to be more representative of noise than a meaningful result; such results are therefore interpreted as ‘no pull’).

Q2: “Where do the Negotiation-Semantic and Interview-Emotion points of Fig. 4 come from, given that Fig. 6 and Tables 8-13 contain no such cells?” Main Response & Revisions: Those points were extracted from our experiments, but the appendix tables submitted were stale, from our original runs on selected variables. We have since regenerated every variable–family–setting combination, so every point in Fig. 4 now has a corresponding table row. Fig. 6 and Tables 11–17 (8–13) have been rewritten, and a new appendix table of robust pull summaries for all four settings (Appendix F.4, Table 14) has been added.

W2: how the attunement variables are measured Reviewer Comment: “Appendix C never says who scores the "six-item, 1–5 response-level questionnaires" or the Big Five, never defines "subjective" vs "objective" stance/task, and points to Appendix A.1 for "Prompts for each variable", which contains only dialogue-generation prompts; no human validation is reported." Main Response: Thank you for pointing this out; we have all the requested information but did not include them in full in the specified appendix sections; we have since made additions to be more thorough. Revisions in paper: Appendix C now includes a table listing who or what scores each variable (Table 3), a paragraph on each scoring method, a definition of subjective and objective scores, and a new subsection with the scoring prompts themselves (Appendix C.1). We also state as a limitation that we did not validate the scores against human raters; this is something we consider a meaningful future development of this study.

AuthorsOpenReview ↗

Response part 3

W2: negotiation task adherence is not decodable Reviewer Comment: “Negotiation task adherence has mean held-out R² of −0.242/−0.263 and peak R² of −0.028/0.009 (Table 8), yet its decoded trajectories are interpreted in Fig. 4/Sec. 5.2 and it supplies 2 of the 14 steering directions (Table 1, Fig. 9).” Main Response: This is a fair critique, and a key limitation of our paper. We interpreted this probe's trajectories with more confidence than the R² score supports. Further investigation shows that task adherence is at least weakly decodable in the ordinal sense, and steering of this variable does not depend on the probe.

  • Explanation: R² compares the probe's errors with the errors of always predicting the average score, so it penalises predictions that are on the wrong scale even if they are ordinally correct. This probe's predictions move in the right direction, but they swing too far from the average, so its errors are large even though it orders responses correctly more often than not. The labels are also skewed and model-specific (57% of Gemma's subjective scores are 5 out of 5, against 1–4% for Granite and Nemotron), which leaves few examples at one end of the scale for some models.
  • New metric: Next to R², we now report for every scalar probe how well its held-out predictions order responses. For negotiation task adherence, the rank correlation between predictions and scores is positive in all 48 model–layer–condition combinations (mean 0.25 for the objective and 0.26 for the subjective score), and exceeds a shuffled-label baseline in 48 of 48 (objective) and 47 of 48 (subjective). The probe ranks a response from the top 30% of scores above one from the bottom 30% in 65% of cases (50% is chance). The other variables have rank correlations between 0.36 and 0.81.
  • Why it can still be used for steering: Steering directions are built from the difference between the model's average activations on high-scoring and low-scoring responses, not from the probes. A probe and a steering direction answer different questions: a probe finds the direction that best predicts a score from the activations, which need not be the direction the model itself uses to produce that behaviour, whereas the difference of average activations is the standard construction for steering (contrastive activation addition), as described in the background section, and is judged by its effect on the generated text. As a direct test, the learned negotiation task-adherence direction changed the task score more than every random direction in all four models, and more than every label-shuffled direction once each shuffled direction's component along the learned direction is removed. Before that removal, 4 of the 20 negotiation shuffled directions reach the learned effect, because directions built from the same responses retain a component along the learned direction (cosine up to 0.72; Appendix G.6, Table 25).
  • Revisions in paper: We now describe this variable as weakly decodable on the basis of its ordinal predictive success, and read the sign but not the magnitude of its pull and contraction. We also checked that our family-level conclusions do not depend on it: removing it together with the four other probes whose mean R² is at or below zero changes no family's inertia by more than 0.03 and no family's median pull by more than 0.04. Table 11 (8) reports the two ordering measures next to R²; Appendix F.1 has paragraphs on sensitivity to decoding quality and on ordinal decodability; Appendix G.6 has the control-direction results.
Reviewer_srv5OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper conducts a number of experiments looking at the dynamics of multi-agent conversations and how steering affects those conversations. It also proposes new terms to characterize model behavior changes. The paper claims that internal representations offer a direct mechanism for controlling long-horizon adaptation but their explanation of their method and results are hard to understand.

Critical strengths

None

Critical weaknesses
  • This paper is very confusing to read. Its claimed contributions are not sufficiently motivated or explained, and the experiment results are confusing to interpret (terms not defined, methodology not explained clearly). I’m not sure how the results support one of their core claims: the long-horizon adaptation can be steered with internal representation. I can see that steering one response affects later dialogue, but this is not surprising. I’m not sure how the results show if steering controls other specific things the paper emphasized, such as the asymmetric model pull.

  • Many terms are not defined or re-defined without justification. For example, they seem to be using a very broad definition of “entrainment”, but did not discuss how this relates to the established, narrower definitions of entrainment in NLP and linguistics, such as one user adapting their language/behavior (unconsciously) to match the other. Other undefined terms include “attractor basins” and subjective vs. objective stance/task scores.

Questions for authors

None

1 comment

Dear Reviewer srv5,

Thank you for your honest feedback regarding interpretability of our manuscript. We have worked on restructuring the narrative and more selectively introducing terminology in order to emphasize the main contributions and results. We appreciate the opportunity to improve the comprehensiveness of the paper sufficiently for your reconsideration.

W1: regarding confusing to read, interpret, not sufficiently explained

We thank the reviewer for this honest feedback. After reviewing the overall structure, as well as specific terminology and phrasing used throughout the paper, we acknowledge that there were improvements that could be made in various areas of the paper in terms of clarity and digestibility. We have taken care to eliminate any unnecessary terminology that serves little purpose in terms of the overall contributions, and dedicated more area towards defining key terms and ideas that remain relevant.

We have made changes to explicitly define relevant terms such as “attractor states/basins, self vs. mixed-play” which our work builds off of and is an integral background to our claims, although the sources of these terms were originally cited in the introduction and background sections of the original submission. We have gone through the full manuscript and pruned out differently-worded variations of the same concept to reduce confusion whenever possible. For other terminology that may be unfamiliar (specifically in regards to your confusion regarding subjective vs. objective stance/task scores), we have added a dedicated terminology section in the appendix (Appendix B.1), with the scoring procedures and prompts in Appendix C.

Adjacently to the terminology issue, we realize that the clarity of our motivations needed improvement. In addition to the terminology changes, we also further explain the motivations behind key terminology choices: specifically, borrowing entrainment as a mechanism fundamentally linked with external interaction as opposed to intrinsic characteristics, and applying it to the niche of interacting conversational agents. You can now find this explanation in the Introduction, with a pointer to Appendix B.1.

In addition to cleaning up some results and figures, we agree that some of the claims/contributions made were phrased ineffectively or in contradictory terms to our intentions. We have revised these claims throughout the manuscript, including the abstract, introduction, background, results, and conclusion, to provide a more clear and coherent delivery of our contribution that aligns with our results. In particular, the third contribution and Section 5.3 now state what the steering results do and do not show: a single intervention reliably shifts the targeted variable and its effect is specific to the learned direction (new control-direction experiment, Appendix G.6), but its effect on partner-relative convergence is small and, in all but one setting–position cell, not distinguishable from zero (Table 1). The revised submission omits the claim that steering controls asymmetric model pull.

W2: regarding unjustifiable definition/usage of terms Response: As mentioned in the introduction and background sections of the original submission, this paper builds off of previous work, especially Ko, Geiping’s “Attractor States Emerge in Multi-Turn LLM Conversations”. The terminology used such as “attractor basins”, as well as the subjective vs. objective stance methodology is a direct reference to these works, where they are introduced. We acknowledge that both the direct terminology and connection could have been made clearer in the manuscript, and have made the appropriate changes in both the introduction and background sections of the main text as well as the Appendix (section B.1).

In regards to the usage of “entrainment”, we believe a significant part of our contribution is directly proposing a new definition of the word, specifically tied to the unique interactional behaviors of LLM agents described in this paper. As thus, it does not mechanistically relate to specific linguistic denotations of the word. Since “entrainment” can be used to describe a variety of phenomena across scientific fields that have no direct relation with each other, we consider it appropriate in this case for the contributional purposes of our paper. This clarification is additionally added to the introduction section of our revised paper.

Summary: We are thankful for the reviewer’s honest feedback. We would like to emphasize that we prioritized making improvements to the delivery of the content to address the confusion issues. However, we would like to clarify that there were justifications for many of the terms and they were selected deliberately. The connections and definitions have been made more explicit and we hope this improves the comprehensibility of our paper.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Reviewer Appreciation (1 of 2)

Dear Reviewers,

Thank you all for your reviews and constructive feedback. Each reviewer had unique concerns, generally with regards to the manuscript's clarity, statistical aggregation, methodology documentation, and interpretation of results. We have revised the paper significantly to address these concerns and to provide a more coherent and substantiated reading. As major changes have been made throughout the paper, we humbly request a broad reevaluation. Regarding individual-specific feedback, we have outlined all addressed changes via direct responses with corresponding section locations for each concern for your convenience. Table and appendix numbers below refer to the revised PDF; submitted-version numbers are given in parentheses where they differ.

  1. Clarified scope, terminology, and contributions. We have restructured the abstract and introduction, and revised the results and conclusion around our most key contribution: proposing agentic conversational entrainment, defined as partner-conditioned changes in the behavioral trajectories of interacting LLM agents. Overall, we have more explicitly defined terminology that directly originates from cited references for convenience and clarity (new Appendix B.1). In terms of limitations and scope, we add the distinction between targeted steering and observed changes in trajectory geometry, and we revise our claims to remove assertions to complete identification of the mechanisms behind partner influence or task performance.

  2. Corrected outlier-sensitive pull metrics and corresponding interpretations. We agree that arithmetic means of normalized partnerward pull can be, and were observed to be, dominated by outlier references. Figures 2B and 4 previously used inconsistent aggregation procedures, producing different magnitudes from the same underlying measurements. Both figures now use a common procedure based on dyad- and behavioral-dimension-level medians, followed by averaging across partners and family dimensions. In addition to terminology clarifications, the appendix now provides additional robust summaries and sensitivity analyses, including pooled numerator-to-denominator ratios and exclusion of observations with the closest reference states (Appendix B.2, F.4). These analyses support positive partnerward displacement for emotion, personality, semantics, and style across the evaluated model–setting combinations. Based on these analyses, we have withdrawn stance and task-adherence results and the interpretation that Gemma and Qwen diverge from their partners in affective space and revised the surrounding discussion accordingly. This has no significant impact on the broader observation that interaction results in systematic partner-relative redirection.

  3. Reconciled measurement coverage and expanded scoring documentation. The missing negotiation-semantic and interview-emotion entries reflected an outdated appendix. We have regenerated the relevant figures and tables (Figure 6; Tables 11–17 (8–13)) so that every point in Figure 4 has corresponding documented results. Appendix C now identifies the scorer or scoring method for each variable (Table 3), defines subjective and objective stance/task scores, and includes the scoring prompts (Appendix C.1). We explicitly acknowledge that these behavioral scores have not been validated against human ratings. Negotiation task-adherence decoding is noted as weaker evidence; we do not treat its previously reported predictive performance as adequate evidence for strong interpretations of decoded task trajectories. This does not affect the steering directions, which are differences of activation means rather than probe weights, or the observation that interventions are immediately effective.

  4. Expanded steering documentation, uncertainty reporting, and controls. We have expanded Appendix D to specify intervention layers (Tables 4–5), normalization, high/low construction, evaluation coverage, and the accounting behind the 32,400 response-level evaluations. The steering tables now include bootstrap confidence intervals (Table 1; Tables 20–21 (16–17)), and Table 1 reports every intervention position rather than selected maxima. We also added direction controls covering all four models and one target per setting, for 16 model–variable–setting cells (Section 5.3; Appendix G.6). Learned directions produced larger immediate target-score separation than all tested norm- and layer-matched random directions. Because random perturbations disrupted Gemma generation, we additionally evaluated shuffled-label directions with their learned-direction component removed. These controls preserved generation and again produced smaller targeted effects. These results strengthen the evidence for direction-specific immediate control within the tested subset; they do not independently establish long-horizon control or the mechanism of natural adaptation.

(Part 1 of 2)

AuthorsOpenReview ↗

Reviewer Appreciation (2 of 2)

  1. Refined claims about causation, context and beneficial control. Because the main steering evaluation uses different target behavior families across settings, we refine the wording of our claims, especially in the results section, to specifically address variation across the tested setting–behavior combinations and intervention positions. We distinguish this variation from an isolated causal effect of conversational setting, and we add a common-variable experiment that steers valence in all four settings (Appendix G.5). We also clarify that behavioral-score changes, likelihood diagnostics, and repetition measurements do not necessarily equate with desired task outcomes. With these revisions, we believe that the key contributions of this paper have been made more robust and apparent.

These revisions serve to preserve and better emphasize our paper's central contribution: using representation analysis and intervention to characterize interaction-specific conversational adaptation, which we refer to as agentic conversational entrainment, while exploring patterns in behavioral persistence, partner-relative displacement, targeted activation intervention, and downstream interaction effects as observed in different interaction settings. We appreciate the reviewers' guidance in making these distinctions clearer and the supporting evidence more transparent.

(Part 2 of 2)

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.