Hyperscanning shows friends explore and strangers converge in conversation

The study asks how social connection (friends vs. strangers) shapes conversational strategies that balance finding common ground (convergence) and exploring new ground (divergence). Conversations can establish social connection quickly, but people are uncertain about what makes a good conversation. Prior work suggests alignment in language, behavior, physiology, and neural activity supports common ground and positive social outcomes. Conversely, novelty and exploration increase engagement and enjoyment. The authors hypothesize that friends, who share prior common ground, will benefit from exploring new ground (divergence), whereas strangers, lacking shared reference points, will primarily converge to establish common ground. They test whether these strategies manifest in neural and linguistic mental state space and in topic dynamics during live, semi-structured intimacy-building conversations, and how these dynamics relate to conversation quality.

The paper integrates two lines of prior research. First, common ground: conversational partners align linguistically (word choice, syntax), behaviorally (movement synchrony), physiologically, and neurally; such convergence is associated with competence, warmth, cooperation, liking, intimacy, social influence, and cohesion. Neural alignment indicates similar interpretations and shared mental states; misalignment occurs when interpretations diverge. Second, exploration and novelty: novelty increases engagement and reduces boredom; people value information intrinsically and derive pleasure from diverse, less repetitive, faster-paced, deeper, and novel-content conversations, which relate to greater enjoyment, well-being, stronger social ties, and relief from negative affect. Relationship context matters: friends leverage shared history, self-disclose more, and have broader, more relaxed, topic-diverse conversations, while strangers experience awkwardness with long silences and may approach conversations differently. Prior work shows friends can generate more topics and make broader semantic associations, though not necessarily communicate more efficiently. These literatures motivate testing how friendship status modulates the balance between convergence and exploration, and how these strategies predict conversational success.

Design: fMRI hyperscanning study with 60 dyads (30 friends, 29 strangers analyzed; 63 initially recruited; 4 excluded for incomplete data). Friends self-identified and reported interacting at least 4 days/week for ≥3 months; strangers were randomly paired. Participants were adults (18+), consented under Princeton IRB, and compensated. Conversation task: Dyads, in separate MRI scanners, engaged in semi-structured conversations based on the Fast Friends procedure with 20 prompts increasing in intimacy (8 low, 6 medium, 6 high). Each trial: 9 s prompt + 180 s turn-taking; participants pressed a button to pass turns; instructed to fill the full 3 minutes. Two within-dyad conditions: Generate (free responses; primary focus of analyses) and Read (scripted from prior pairs). Five runs (13.6 min each). Order of conditions randomized; investigators were not blinded to condition allocation. Data acquisition: Speech recorded with MR-compatible dual optical microphones and adaptive noise-cancellation and enhancement. fMRI collected on 3T Siemens Skyra/Prisma with identical parameters (TR=1500 ms, TE=28 ms, flip angle=80°, 3.0 mm isotropic functional; T1-weighted anatomical 1 mm isotropic). Preprocessing: fMRIPrep v20.2.0 for robust preprocessing: INU correction, skull-strip, normalization to ICBM 152 2009c template, tissue segmentation, fieldmap distortion correction via T1 inversion-based method and average fieldmap template, BBR coregistration, concatenated transforms with Lanczos interpolation, CompCor (tCompCor and aCompCor) components, framewise displacement. Additional confound regression to mitigate speech-related motion: six motion parameters, cosines, derivatives, squares; WM and CSF signals with derivatives and squares. Cleaned residuals used. Trials truncated by 7 TRs at start and end to avoid prompt on/off BOLD responses. Neural mental state decoding: Built predictive models to decode three mental state dimensions (3D Mind Model: rationality, social impact, valence) from whole-brain activation using LASSO-PCR, trained and validated on four independent fMRI datasets designed to vary along these dimensions. Regions contributing: rationality (right inferior frontal gyrus, MPFC), social impact (default mode network: PCC, angular gyrus, MPFC), valence (vMPFC, MPFC). Applied models to conversation data volume-wise to obtain time series of the three dimensions for each participant. Neural distance metric: Computed Mahalanobis distance between partners’ 3D mental state coordinates at each time point within trials; averaged within trials for certain analyses. Primary dependent variable: time-varying dyadic neural distance (smaller distance = greater alignment). Linguistic mental state decoding: Exploratory parallel analysis using NLP (affectR) to infer 3D mental state coordinates from the words used in each speech turn. Computed Mahalanobis distance between partners per turn within trials. Exclusions: two dyads lacked analyzable trials (>3 turns required); removed trials with >12 turns (1.5*IQR rule; 17% of trials) to ensure meaningful NLP analysis; included words-per-turn as a covariate in robustness checks. Internal consistency between neural and linguistic measures was assessed via distance correlation. Topic modeling (content space): Exploratory analysis using BERTopic to extract topic embeddings per turn across dyads. Measures: (1) number of topics generated per dyad, (2) number of topic switches (including returning to prior topics), (3) mean pairwise cosine distance between topics within a dyad, and (4) robustness check with Euclidean distance. Also modeled turn-by-turn cosine distance over time in multilevel models. Conversation quality measures: Post-conversation survey captured enjoyment, closeness, similarity, anxiety while speaking/listening, desire to interact again, and desire to become friends. Factor analysis identified four latent factors; negative affect (reverse-coded) showed low reliability and was not analyzed further. Statistical analysis: Given nesting (time points within trials within participants), used multilevel models (R nlme) with Gaussian link and random intercepts to test effects of relationship (friends vs. strangers) and time (within-trial time points; trial number) and their interactions on distances in neural, linguistic, and topic spaces. Trials mean-centered. Between-group comparisons for topic measures via two-sample t-tests (Welch when variances unequal). For strangers, additional exploratory multilevel regressions tested whether divergence over time (slopes) predicts closeness, enjoyment, and similarity. Predictive modeling: LASSO regression evaluated the relative predictive power (out-of-sample) of neural, linguistic, and topic divergence measures for conversation enjoyment, with model comparison via BIC.

• Conversation quality (friends vs. strangers): Factor analysis yielded four latent factors. Friends scored higher on closeness (t(57)=7.56, p<0.001, d=1.97, 95% CI [1.09, 1.87]), enjoyment (t(57)=5.67, p<0.001, d=1.48, 95% CI [0.61, 1.27]), similarity (t(57)=10.81, p<0.001, d=2.82, 95% CI [1.38, 2.00]), and reverse-coded negative affect (t(57)=3.23, p=0.002, d=0.84, 95% CI [0.32, 1.36]); negative affect had low reliability (Cronbach’s α=0.63; interrater r=0.29, p=0.03) and was not examined further.
• Neural mental state dynamics: Significant interaction of time × relationship: friends diverged while strangers converged in 3D neural mental state space (β=-0.001, SE=0.0003, 95% CI [-0.001, -0.00034], p=0.001). Friends began more aligned than strangers, then diverged to exceed strangers’ distance over time.
• Linguistic mental state dynamics: Exploratory analysis showed a similar trend (β=-0.02, SE=0.014, 95% CI [-0.05, 0.005], p=0.11), with a significant three-way interaction indicating that after about half the trials, friends diverged and strangers converged (β_Friend turns×trials=-0.007, SE=0.003, 95% CI [-0.012, -0.002], p=0.006; β_Stranger turns×trials=-0.005, SE=0.004, 95% CI [-0.013, 0.002], p=0.19; F_overall=4.67, p=0.009).
• Topic dynamics (content space): Friends explored topics more broadly and rapidly (turn-by-turn divergence: β=-0.009, SE=0.002, 95% CI [-0.013, -0.005], p<0.0001). Friends generated more topics (t(57)=3.00, p=0.004, d=0.78, 95% CI [0.81, 4.14]), switched topics more often (t(57)=2.43, p=0.019, d=0.64, 95% CI [1.06, 11.49]), and made larger topic jumps (mean cosine distance: t(57)=2.75, p=0.008, d=0.72, 95% CI [0.005, 0.031]).
• Divergence predicts better conversations among strangers: Greater divergence over time was associated with higher quality conversations for strangers. Closeness: more linguistic divergence predicted feeling closer (β=0.058, SE=0.015, 95% CI [0.03, 0.09], p<0.001). Enjoyment: more neural divergence predicted greater enjoyment (B=0.0013, SE=0.0003, 95% CI [0.0007, 0.0017], p<0.001). Similarity: more topic exploration predicted feeling more similar (B=0.007, SE=0.003, 95% CI [0.0008, 0.011], p=0.01); trends for linguistic (B=0.03, SE=0.018, 95% CI [-0.005, 0.06], p=0.10) and neural (B=0.0005, SE=0.0002, 95% CI [-0.00009, 0.001], p=0.10) divergence.
• Predictive modeling: LASSO regression highlighted neural divergence as the most robust out-of-sample predictor of enjoyment (β_neural_slope=0.14, 95% CI_boot [0, 0.33]; β_topic_slope=0.02, 95% CI_boot [0, 0.20]; β_ling_slope=0, 95% CI_boot [0, 0.14]; PMSE=0.009, PRMSE=0.008). A model with only the neural measure had the lowest BIC (57.7), indicating best fit-complexity trade-off.

The findings reveal two distinct conversational strategies that depend on preexisting social connection. Friends, starting with greater common ground, tend to explore new ground: their neural and linguistic mental states diverge over time and they traverse topic space more broadly, generating more topics, switching more frequently, and making larger semantic jumps. Strangers, starting with less common ground, tend to converge neurally and linguistically and to exploit topics longer. Importantly, among strangers, conversations that diverge more—resembling the exploratory style of friends—are associated with higher enjoyment, closeness, and perceived similarity. These results reconcile theories emphasizing common ground (alignment across modalities linked to positive social outcomes) with those emphasizing novelty (engagement and well-being), suggesting that the optimal strategy depends on relationship context and may unfold sequentially (initial convergence to reduce uncertainty followed by exploration). The superior predictive value of neural measures suggests that continuous, bidirectional tracking of partners’ mental states via fMRI captures drivers of conversational success beyond language production alone.

This study demonstrates that conversational trajectories differ by relationship status: friends explore (diverge) while strangers find common ground (converge). Using fMRI hyperscanning and NLP/topic modeling, the authors tracked dyads’ neural and linguistic mental states and topic movements in real time during semi-structured conversations. Friends’ divergence and strangers’ convergence were evident across neural, linguistic, and topic measures. For strangers, greater divergence predicted better conversations. These findings suggest that shifting from a default focus on common ground to exploring new ground can enhance conversational success, especially for new acquaintances. Future research should test generalization to other conversational goals (e.g., persuasion, teaching, learning), longer unstructured interactions outside the scanner with voluntary engagement, and alternative neuroimaging modalities (EEG, fNIRS) to assess the portability of neural predictors.

• Context and ecological validity: Conversations occurred in MRI scanners, were time-limited, and structured by prompts; participants were effectively compelled to converse, which may differ from natural, voluntary interactions. Generalization to real-world settings requires caution.
• Measurement modality: Findings rely on fMRI for neural decoding; while advantageous in spatial resolution, generalization to other modalities (EEG, fNIRS) remains to be established.
• Design features: Investigators were not blinded to condition allocation. The conversational goal focused on affiliation-building (Fast Friends), limiting generalization to other goals (e.g., persuasion, instruction).
• Reliability of some outcomes: The negative affect factor showed low reliability and was not analyzed further.
• Temporal scope: Evidence suggests strangers may begin exploring after initial convergence, especially in longer interactions; the present task duration may constrain observing full strategy sequences.
• Data constraints: Linguistic analyses excluded trials with very short turns and those with >12 turns; raw text data are unavailable due to privacy, which may limit some reproducibility aspects (processed outputs are provided).