PaperlayerView exact revision on arXiv

Source-linked AI summary

Generative Interactions: Weaving Multiparty Human Motion with Bilevel Latent Dynamics

Ojas Shirekar, Yash Surange, Agustinas Jučas, Chirag Raman

arXiv:2609.37708v1cs.AIcs.LG

TL;DR

Social motion models often prioritize plausible trajectories while leaving interaction state implicit, despite the need to transfer across groups, tasks, and observation regimes. BRAID uses hierarchical group- and person-level latents with arbitrary observed context sets. Across tasks, it retains distributional fit and temporal alignment as evidence is removed, while producing varied completions and interpretable latent structure.

  • Problem

    Social motion methods often leave the interplay of group dynamics and individual traits implicit, while fragmented interaction data complicate transfer across datasets, contexts, and observation regimes.

    Source: §1 · p. 1

  • Method

    BRAID is a hierarchical sequential latent-variable model with group latents for shared interaction dynamics and person latents conditioned on group state, trained through cross-dataset meta-transfer learning.

    Source: §1 · p. 2

  • Results

    Across forecasting, partner inpainting, and dyadic prediction, BRAID retains distributional fit and temporal alignment as evidence is removed and produces varied completions; its unsupervised latents organize interactions by activity and recover participants across groups.

    Source: §6 · p. 9

  • Takeaways & Limitations

    BRAID's group and person latents provide inspectable social-state representations for describing interaction structure and individual behaviour.

    Source: §5.1 · p. 7, §6 · p. 9

    The evidence covers the evaluated social-motion datasets and depends on SMPL-X preprocessing quality.

    Source: §6 · p. 10

Abstract

from arXiv · show

Human social behaviour is not a collection of independent motions, but a jointly organised process in which group dynamics and individual variation continuously shape one another. Yet existing social motion models often prioritise plausible trajectories while leaving interaction state implicit, limiting their ability to transfer across groups, tasks, and partial-observation regimes. To address this gap, we introduce Bilevel Representations for Agent Interaction Dynamics (BRAID), a hierarchical sequential latent-variable model for generative multi-person interaction. BRAID explicitly formulates social motion generation as a meta-transfer learning problem: shared interaction priors are learned across datasets and adapted through arbitrary context sets of observed people and joints. The model represents each scene through a group-level latent state that captures shared interaction dynamics and person-level latent states that capture individual behaviour conditioned on the evolving group context. This modelling choice enables coherent generation under full, sparse, or partial observations while exposing compact social-state vectors that can serve as an interface for downstream embodied-agent systems. We evaluate BRAID under a unified SMPL-based representation on social forecasting, tracking and in-filling, and response generation, using metrics that assess not only reconstruction accuracy but also realism, diversity, temporal alignment, and interpersonal coordination. We further analyse the hierarchical latent space, showing that it captures separable group- and individual-level structure.

1 INTRODUCTION

Social interaction requires modeling both group coordination and individual variation, yet data scarcity and limited transfer make reusable social-state representations difficult to learn. BRAID addresses this with hierarchical latents and cross-dataset meta-transfer learning for generation under varied observations.

1 INTRODUCTION

  • The motivation is to represent social behaviour as coordinated motion shaped by group dynamics and individual variation, including non-verbal signals and multiple valid responses.Responses may vary with individual traits, group roles, cultural contexts, and situational norms.Source: §1 · p. 1
  • Fragmented interaction datasets cover only a fraction of real-world situations, making transfer across datasets, contexts, and observation regimes a central challenge.The paper frames the problem as both data scarcity and the need to generalize learned structure.Source: §1 · p. 1
  • Existing social motion methods often leave shared dynamics and individual traits implicit, while reusable social-state representations remain rarely addressed alongside generation under partial observations.Plausible trajectories are often the primary endpoint, although compact state representations could inform policies, planners, and memory systems.Source: §1 · p. 2
  • BRAID uses a hierarchical sequential latent-variable model in which group latents capture shared interaction dynamics and person latents encode individual behaviour conditioned on group state.The coupled temporal evolution supports coherent generation, and arbitrary context sets accommodate full, sparse, or partial observations.Source: §1 · p. 2
  • BRAID learns shared priors through cross-dataset meta-transfer learning across conversational, dance, and boxing data, exposing compact group- and person-level social-state vectors.The paper also identifies evaluation across datasets and context regimes, including partially observed interaction partners, and analysis of group- and person-level latent structure as contributions.Source: §1 · p. 2

2 RELATED WORK

Prior work spans single-person motion generation, social motion prediction, and Neural Process or hierarchical latent models. BRAID combines explicit group- and person-level latents to represent shared interaction structure and individual variation.

2 RELATED WORK

  • Single-person motion generators provide strong individual-behaviour priors, but social settings make pauses, gestures, and steps meaningful through coordination with other participants (Tevet et al., 2022b;a; Guo et al., 2022; Diomataris et al., 2024).BRAID models the interaction rather than an isolated body.Source: §2 · p. 2
  • Social motion methods cover multi-person prediction, inpainting, response generation, and specialized domains, while most retain plausible cues or trajectories as their primary endpoint (Mughal et al., 2024).Social Processes meta-learns conversational group dynamics but does not address full-body social motion generation (Raman et al., 2023).Source: §2 · p. 2
  • The comparison table covers interaction capabilities, defining latent space as an explicit compact representation and polyadic as support for P ≥3 participants.The supplied caption gives these comparison dimensions but no method-specific table values.Source: Table 1 · p. 3
  • BRAID exposes group and person-level latents so shared interaction dynamics can condition individual behaviour while remaining available as reusable social-state representations.Prior approaches may lack group-level state, leave it implicit, or embed it in denoising processes that do not naturally expose compact representations (Tanke et al., 2025).Source: §2 · p. 3
  • Neural Processes condition a latent variable on a context set, and BRAID extends this line with a temporal hierarchy of group and person latents (Garnelo et al., 2018b;a; Kim et al., 2019; Singh et al., 2019; Raman et al., 2023).The group latent captures shared interaction structure; person latents encode individual behaviour conditioned on it.Source: §2 · p. 3
  • BRAID applies hierarchical VAE ideas to social motion, with higher-level latents for emergent interaction properties and person-level latents for individual variation conditioned on shared context.Its staged KL annealing and free-bits terms follow established practice, with warm-up ordered to stabilize group structure before person-level variation.Source: §2 · p. 3

3 PROBLEM STATEMENT

The problem statement frames social motion as sequential prediction from observed context sets to target elements. BRAID extends a context-conditioned Neural Process with a temporal latent hierarchy for shared group structure and individual behaviour.

3 PROBLEM STATEMENT

  • At each time step, the temporal Neural Process setting uses a possibly empty observed context set, allowing the model to condition on available observations.The Figure 2 caption labels C as observed context and D ⊇ C as the full target set.Source: §3 · p. 3, Figure 2 · p. 4
  • Held-out elements in D \ C are query locations and target values to be predicted.The target set is the prediction domain beyond the observed context.Source: §3 · p. 3
  • A standard Neural Process conditions a single latent on the current context but does not model dependencies between successive processes; BRAID instead introduces L latent levels.The hierarchy is conditioned on context, latent history, and lower levels in the hierarchy.Source: §3 · p. 3
  • In the two-level social-motion model, the group latent represents shared interaction dynamics and person latents capture individual behaviour conditioned on group context.The person latents form the structured collection of person-level latents in the scene representation.Source: §3 · p. 4

4 METHOD

BRAID models multi-person motion with a two-level temporal latent hierarchy: group states encode shared interaction, while person states encode individual behavior conditioned on group context. A joint representation and context interface support generation when people or joints are observed, missing, or targeted for completion.

4 METHOD

  • The two-level hierarchy assigns shared interaction state to group latents and individual behavior conditioned on that state to person latents.The group state can capture phase, tempo, and proxemics.Source: §4 · p. 4

4.1 MOTION REPRESENTATION

  • Scenes encode SMPL-X shape, 6D joint rotations, root-relative joints, canonical root motion, and pairwise transforms that preserve spatial awareness.The predicted canonical transforms include frame-to-frame displacement, canonical-to-root transform, and pairwise partner transform.Source: §4.1 · p. 4
  • Missing joints or people are omitted from the context, so forecasting, tracking or in-filling, and response generation differ only in which indices are observed.The model receives partial observations at different time steps and generates the complete scene for each person in the triad.Source: §4.1 · p. 5, Figure 3 · p. 5

4.2 LATENT ENCODING AND GENERATIVE MODEL

  • Given context, BRAID samples a group latent, samples person latents conditioned on it, and decodes target motion quantities.A pretrained, frozen temporal motion VAE decodes body pose; separate heads predict canonical and partner transforms.Source: §4.2 · p. 5

4.3 CONTEXT ENCODING

  • Context encoding combines observed-joint tokens, cross-attention to target slots, person embeddings, temporal and cross-person attention, and a resampler for the group representation.The deterministic context representations include group and per-person summaries.Source: §4.3 · p. 5

4.4 INFERENCE AND LEARNING

  • BRAID approximates the intractable social-state posterior with an amortised variational distribution that mirrors the generative model’s temporal and hierarchical structure.For βg = βp = 1, the objective is a lower bound on ln pθ(Y | X, C); reconstruction losses supervise rotations, root-relative positions, and canonical transforms.Source: §4.4 · pp. 6–7
  • Precision-weighted posterior fusion combines a data-driven proposal with the prior, allowing their influence to vary with precision and interpolating between evidence and prior.The passage identifies sparse or absent context as a setting where this formulation is particularly beneficial.Source: §4.4 · p. 6

5 EXPERIMENTS

The experiments assess whether BRAID’s shared group- and person-level latent representation retains social information and supports generation across tasks and observation regimes. Results show structured latent organization and several gains under partial observation, alongside metric- and setting-specific limitations.

5 EXPERIMENTS

  • Experiments use Haggling, DnD, DD100, DuoBox, and Embody3D, and assess reference errors, diversity, distributional fit, artefacts, synchrony, and trajectory similarity.The evaluation prioritizes interaction-level structure, including synchrony and cross-person Soft-DTW, because reference errors alone do not distinguish socially incoherent motion from valid alternatives.Source: §5 · p. 7

5.1 RESULTS

  • Group latents organize activity structure, while person latents retain participant identity across changing groups; the full hierarchy also improves cross-person alignment over single-level variants.The same-person distance median is 6.36 versus 16.22 for different-person pairs, and nearest-neighbor retrieval identifies the same person in 99/239 queries (41.4%).Source: §5.1 · pp. 7–8
  • Under partial observation, BRAID lowers forecasting FD at every group size while retaining diversity and temporal structure, but three-person FS rises and MPJPE increases for two- and three-person groups.For two people, FD changes from 0.239 to 0.126; for three, from 0.221 to 0.141. Three-person FS rises from 0.071 to 0.235, and recorded motion has zero FS.Source: §5.1 · p. 8
  • In partner inpainting, BRAID improves DIV, FS, FD, and SDTW over RoHM on both datasets, while RoHM retains an advantage in cross-person SDTW.On DD100, DIV rises from 0.080 to 7.49, FS falls from 0.311 to 0.013, and SDTW falls from 7.017 to 4.623; recorded dancers overlap in 29% of frames, so lower interpenetration does not imply fewer artefacts.Source: Table 3 · p. 8, §5.1 · p. 9
  • In DuoBox prediction with a missing partner, BRAID produces more diverse samples with lower FD than RoHM, but RoHM has lower dyadic MPJPE and BRAID’s dyadic interpenetration exceeds the recorded rate.For dyadic prediction, DIV rises from 0.026 to 2.37 and FD falls from 0.325 to 0.115; dyadic IP is 0.009 versus the recorded 0.0008.Source: §5.1 · p. 9, Table 4 · p. 9
  • A DD100-only model retains plausible conversational formation from 40 frames of held-out Panoptic context but drifts toward dance-like rotations; adding conversational data lowers cross-person SDTW from 1.424 to 0.511 and FD from 1.94 to 0.27.Source: §5.1 · p. 9

6 DISCUSSION AND CONCLUSION

BRAID treats multi-person motion as conditional generation over arbitrary context sets and exposes inspectable group- and person-level social states. The discussion identifies possible agent uses and practical limits, while the appendix derives a hierarchy-aware evidence lower bound.

6 DISCUSSION AND CONCLUSION

  • Across forecasting, partner inpainting, and dyadic prediction, BRAID retains distributional fit and temporal alignment as evidence is removed while producing varied completions.Its latent states also organise interactions by activity and dance style and recover participants across groups.Source: §6 · p. 9
  • Zg could serve as a compact belief state for agents observing or joining groups, while Zind could support memory of people across encounters and adaptation to their interaction styles.The two levels could also be supplied to foundation models, vision-language-action models, and robot controllers; exchanging them suggests controllable social-scenario synthesis.Source: §6 · pp. 9–10
  • Evidence is limited to evaluated social-motion datasets, depends on SMPL-X preprocessing quality, and does not show complete disentanglement of the latent structure.Diversity is low in largely stationary five-person DnD scenes, and pose and timing errors can accumulate in long open-loop rollouts; new observations can correct predictions as they arrive.Source: §6 · p. 10

A DERIVATION OF THE EVIDENCE LOWER BOUND

  • The bound scores each person’s target-motion reconstruction and compares target-informed posteriors with context-conditioned priors at group and person levels.The group KL is counted once per time step, not once per person.Source: A DERIVATION OF THE EVIDENCE LOWER BOUND · pp. 16–17
  • With both KL weights at least one, the objective remains a lower bound; arbitrary weights below one do not generally preserve that guarantee.Conditional KL terms are evaluated analytically for diagonal Gaussian factors, while remaining expectations use reparameterised posterior samples; auxiliary reconstruction losses are outside the bound derivation.Source: A DERIVATION OF THE EVIDENCE LOWER BOUND · p. 17

B CROSS-GROUP PERSON RE-IDENTIFICATION

Cross-group analyses find persistent participant-specific structure in BRAID’s person latents without identity supervision during training. Distance comparisons and nearest-neighbour retrieval both indicate that same-person representations remain similar across groups.

B CROSS-GROUP PERSON RE-IDENTIFICATION

  • The analysis uses 268 person-group vectors from 87 groups, 1,600 sequences, and 91 individuals, with identity labels reserved for evaluation.Distances are Euclidean in the full latent space; PCA is used only for visualisation.Source: B CROSS-GROUP PERSON RE-IDENTIFICATION · p. 17
  • The median same-person distance is 6.36, versus 16.22 for different people (ratio 0.392).The median same-person pair lies at the 15.7th percentile of different-person distances; Figure 8 describes the percentile comparison and its uniform reference.Source: B CROSS-GROUP PERSON RE-IDENTIFICATION · p. 17, Figure 8 · p. 18
  • Same-person pairs have a median percentile of 16, placing a typical person closer to themselves across groups than 84% of stranger pairs.The comparison includes 399 same-person pairs; 38 relatively distant pairs above the 80th percentile show the structure is not equally strong for every pair.Source: B CROSS-GROUP PERSON RE-IDENTIFICATION · p. 17
  • 41.4% top-1 accuracy (99 of 239 queries) retrieves the correct identity across groups without metric learning or re-ranking.The retrieval excludes each query’s group, though different groups may still share partners or activities.Source: B CROSS-GROUP PERSON RE-IDENTIFICATION · p. 18

C TRAINING DETAILS

Training varies observed people and joints across clips so one model can address several motion tasks under arbitrary context sets. The implementation couples a hierarchical latent model to a frozen temporal motion VAE for body decoding.

C TRAINING DETAILS

  • Training uses Haggling, DnD, DD100, DuoBox, and Embody3D data, with examples containing two to five participants and randomly varied context patterns for forecasting, in-filling, and completion.The target is the full 200-frame window, and masks are regenerated at every update.Source: C TRAINING DETAILS · p. 18

D MODEL IMPLEMENTATION

  • The Embody3D and DnD models behind the N = 2–5 results use the base BRAID objective without the pretrained motion VAE or continuation stages because of resource limits.The appendix analyses and ablation tables in E–G and I also use BRAID without the motion VAE; the authors say they will update those results with the full pipeline.Source: D MODEL IMPLEMENTATION · p. 19

D.1 TEMPORAL MOTION VAE

  • A pretrained temporal motion VAE compresses motion by 4× into one code per four frames, then remains frozen after pretraining.Its encoder and decoder use complete motion and true canonical trajectories during pretraining.Source: §D.1 · pp. 19–20

D.2 COUPLING THE HIERARCHY TO THE MOTION VAE

  • Group latents update every five frames while person latents are sampled each frame conditioned on the current group sample; an adapter maps latents, histories, and canonical motion to codes decoded by the frozen VAE.Generation samples latents from context-conditioned priors, accesses neither target motion nor target codes, and anchors world-frame roots to observed roots.Source: §D.2 · p. 20

D.3 TRAINING STAGES

The final model-size comparison favours the 16M model on trajectory, cross-person alignment, and distributional metrics across Panoptic and DD100. Training continuations progressively couple rollout, trajectory, and VAE-code supervision.

D.3 TRAINING STAGES

  • Stage 1 trains the hierarchical model from scratch around the frozen VAE; Stage 2 adds dense posterior and mixed posterior-prior rollouts and removes the trajectory-loss stop-gradient to person latents.In the mixed branch, only posterior-sampled frames are scored, while the posterior retains access to the full target.Source: §D.3 · pp. 20–21
  • Stage 3 adds frozen-encoder code supervision, and Stage 4 adds frozen VAE mean codes to person-posterior evidence; the DD100 and DuoBox results use the Stage 4 checkpoint.Stage 4 leaves the group posterior, priors, and generation path unchanged.Source: §D.3 · p. 21

E MODEL SIZE COMPARISONS

  • The model-size sweep holds data, learning-rate schedule, and epoch budget fixed, so it evaluates selection under a common budget rather than establishing a scaling limit.The comparison does not test whether larger models benefit from different optimisation settings or more data.Source: E MODEL SIZE COMPARISONS · p. 22
  • The 16M model attains the lowest SDTW, cross-person SDTW, and FD on both Panoptic and DD100.Larger variants improve determinism error and sample spread, but have higher cross-person SDTW and FD in this comparison.Source: Table 6 · p. 22, Table 7 · p. 22, E MODEL SIZE COMPARISONS · p. 22

F CROSS-DATASET AND CROSS-DOMAIN GENERALISATION

The cross-domain evaluation tests whether shared training transfers conversational group structure to held-out groups when conversational motion is unseen during training. Joint training improves activity-specific forecasting, while the dance-only model retains broad pose and formation structure but may generate dance-like motions.

F CROSS-DATASET AND CROSS-DOMAIN GENERALISATION

  • Both models observe 40 conversational frames and generate 60 on held-out Panoptic groups, but conversation is unseen in training only for the DD100-only model.Adaptation uses context without test-time parameter updates; evaluation covers 200 batches of 32 sequences.Source: F CROSS-DATASET AND CROSS-DOMAIN… · p. 22
  • Joint training improves cross-person SDTW from 1.424 to 0.511 and FD from 1.94 to 0.27 against the DD100-only model on held-out Panoptic groups.Table 8 reports the comparison for 40 context frames and 60 generated frames.Source: Table 8 · p. 22, F CROSS-DATASET AND CROSS-DOMAIN… · p. 22
  • The jointly trained model maintains conversational formation and activity-appropriate gestures, whereas the dance-only model produces spinning motions characteristic of its training domain.The dance-only model still generates plausible poses and a sensible group formation.Source: F CROSS-DATASET AND CROSS-DOMAIN… · p. 22, Figure 9 · p. 23
  • The dance-only model has higher DIV and lower FS, so sample spread and foot stability need interpretation alongside activity-appropriate alignment and distributional fit.The authors do not establish equivalent performance on activities absent from training.Source: F CROSS-DATASET AND CROSS-DOMAIN… · p. 23

G CONTRIBUTION OF THE BILEVEL HIERARCHY

The hierarchy ablations associate separate person latents with interpersonal alignment and social context with both alignment and individual motion prediction. Qualitative probes show person-level variation under a fixed group state and changed interaction realizations when the group state changes.

G CONTRIBUTION OF THE BILEVEL HIERARCHY

  • The ablations compare the full group-and-person hierarchy with group-only and person-only variants while keeping the remaining architecture fixed.All variants, including the full model, are trained without the motion VAE.Source: G CONTRIBUTION OF THE BILEVEL HIERARCHY · p. 23
  • On Panoptic, removing person latents raises cross-person SDTW from 0.511 to 1.119, and removing social context raises it to 1.305; on DD100, the values are 0.785, 0.920, and 1.534.On DD100, removing social context also raises SDTW from 2.39 to 6.35 and MPJPE from 0.72 to 1.30.Source: Table 9 · p. 23, G CONTRIBUTION OF THE BILEVEL HIERARCHY · p. 24
  • The person-only variant has lower Panoptic FD (0.18 versus 0.27) and lower determinism error on both datasets, but these are not uniform advantages over the full hierarchy.FD measures per-person frame-feature distributions, while determinism captures one aspect of recurrence; neither alone establishes interpersonal alignment.Source: G CONTRIBUTION OF THE BILEVEL HIERARCHY · p. 24
  • With context and group latent fixed, varying person latents produces different poses, orientations, and relative arrangements in the DuoBox example.The person latents express individual realizations of an interaction, rather than motion details that leave group geometry invariant.Source: G CONTRIBUTION OF THE BILEVEL HIERARCHY · p. 24
  • Switching the group-latent trajectory between interactions involving the same people changes their postures and group arrangement through conditionally sampled person latents.Because person latents also change, this experiment does not isolate a direct group-latent effect with all other latents fixed.Source: G CONTRIBUTION OF THE BILEVEL HIERARCHY · p. 24

H EVALUATION METRICS AND INTERPRETATION

The evaluation combines measures of interaction structure, distributional agreement, variation, reference recovery, and physical artefacts, since no single score establishes plausible and socially appropriate continuation. An additional missing-joint forecasting comparison finds BRAID strongest on trajectory structure, interpersonal alignment, and distributional fit across both datasets, with complementary strengths among baselines.

H EVALUATION METRICS AND INTERPRETATION

  • The evaluation combines interaction structure, distributional agreement, sample variation, and physical diagnostics, with no single score establishing plausible and socially appropriate continuation.Random-sample and nearest-neighbor baselines replay training motion without modelling interaction and are excluded from best/second-best rankings in the main tables.Source: H EVALUATION METRICS AND INTERPRETATION · p. 24
  • CRQA measures recurrence between participants’ motion states using a dataset-specific distance radius calibrated to a 2% recurrence rate and then held fixed.RR is the fraction of recurrent state pairs; DET is the fraction of recurrent points in diagonal line structures, and reported deviations from ground truth are better when lower.Source: H EVALUATION METRICS AND INTERPRETATION · pp. 24–25
  • Figure 10 contrasts different person-latent samples under the same context and group latent with changed states after switching the group-latent source.Person latents are sampled conditionally on the active group state.Source: Figure 10 · p. 25, H EVALUATION METRICS AND INTERPRETATION · p. 25
  • SDTW allows temporal alignment, while cross-person SDTW tests whether generated interactions preserve relative temporal structure between participants; lower values are better.Off-diagonal recurrences can capture leading and following, rather than requiring identical movement at the same instant.Source: H EVALUATION METRICS AND INTERPRETATION · p. 25
  • FD compares generated and reference motion distributions, while DIV measures pairwise sample distance; neither alone establishes coverage of valid, socially aligned interactions.For hierarchy ablations, FD uses per-person frame features, and diversity should be interpreted alongside temporal structure and activity.Source: H EVALUATION METRICS AND INTERPRETATION · p. 25

I ADDITIONAL NEURAL-PROCESS BASELINE COMPARISONS

  • BRAID achieves the lowest SDTW, cross-person SDTW, and FD on both datasets in the additional baseline comparison.SP-style has lower determinism and velocity errors and lower foot skating; SNP-style has greater Panoptic diversity and slightly lower DD100 position error, while BRAID has the highest reported DD100 diversity.Source: Table 10 · p. 26, Table 11 · p. 26, I ADDITIONAL NEURAL-PROCESS BASELINE… · p. 26

H EVALUATION METRICS AND INTERPRETATION

  • MPJPE and MPJVE measure recovery of a particular recorded continuation, while foot skating and interpenetration check specific physical artefacts rather than overall plausibility.Recorded motion can itself contain non-zero skating or overlap, so values below recorded rates need not indicate better motion.Source: H EVALUATION METRICS AND INTERPRETATION · p. 26

I ADDITIONAL NEURAL-PROCESS BASELINE COMPARISONS

  • The additional comparison uses 70 observed frames, 20% randomly missing context joints, and a 130-frame prediction horizon, and should be read separately from the main tables.It compares BRAID with SP-style, SNP-style, and RoHM; the SP-style implementation accepts missing joints, while SNP-style uses a single group latent.Source: I ADDITIONAL NEURAL-PROCESS BASELINE… · p. 26

Related papers

Loading 2609.37708v1…