betaFeedback

WorksEngineering

ANCHOR: A Port-Hamiltonian Model for Forecasting Team Dynamics in Human-in-the-Loop Systems

Abstract

Cyber-physical-human systems close a loop around people by sensing human behavior, predict it and act on it. Human teams are a natural target, because collaboration breaks down in ways that a timely prompt could prevent. Collaborative mixed reality makes this loop practical, since each headset senses who speaks, where its wearer looks and who stands together, and it can deliver a light prompt. In this loop, the headsets are sensors, the prompts are actuators and the team is the plant. Closing the loop needs a plant model that is small enough to learn from few teams, safe to iterate and readable by a designer. We propose ANCHOR, which models a team as a ball in an energy bowl. The bottom of the bowl is where the team rests, and it is set by the team's recent behavior, its own history and other teams. Formally, ANCHOR is a switched port-Hamiltonian system with 52 parameters, and each part stands for a testable claim from team psychology. Because of this structure, for any fitted parameters, a forecast moves away from rest only when task progress or a prompt pushes it, and prompts of bounded size that are spaced in time move it a bounded distance. On \nGroups{} four-person teams held out from training, a team that drifts from its usual behavior returns most of the way within half a minute. After 32 s, about 24% of the deviation remains, and ANCHOR predicted 30%. Forecasts are clearly better when the resting point blends its three sources than when it uses one. ANCHOR matches the accuracy of a black-box baseline with about 40,000 parameters. It anticipates conversation collapse with an AUROC of 0.85. In simulation under assumed prompt effects, it removes 2.2 times as much breakdown time per prompt as periodic prompting.

CCS Concepts

• Computer systems organization → Embedded and cyberphysical systems; • Human-centered computing → Collaborative and social computing theory, concepts and paradigms.

Introduction

Many goal-oriented tasks are carried out by teams, from operating rooms and control rooms to design reviews and training. How well a team performs depends not only on its members’ skills but also on how they interact, for example on whether everyone gets to speak and whether members attend to the same things [19, 35]. These patterns change during a session, and a team can drift into

a breakdown in which one member dominates or the team loses its shared focus. A timely prompt can correct such a drift, and coaching helps most when it arrives at the right moment [15]. A system that senses a team, anticipates its breakdowns and prompts it is therefore a cyber-physical-human system in which people are part of the loop [5, 31].

Collaborative mixed reality (MR) makes this loop practical. Each headset already senses who speaks, where its wearer looks and who stands together [29], and it can deliver a light prompt, so the headsets act as sensors and the prompts as actuators (Figure 1). We study this loop with four-person teams that sort images together in MR. What the loop needs is a plant model. A controller needs a forecast several steps ahead that meets three demands. The model must learn from tens of teams, because each team is expensive to record. It must stay well behaved when iterated and when prompts are added. Finally, a designer must be able to read and question it.

Diagram of ANCHOR as a port-Hamiltonian model for team dynamics

Figure 1. ANCHOR in one picture. Left: the team (the ball) is pulled toward a resting point. Its height is the energy, regulation sets how fast it returns, and task progress moves the bowl. Right: ANCHOR turns headset data into forecasts, warnings and cue timing.

The idea. Team psychology offers a simple picture. Small groups restore their balance after a disturbance [4], recurring social patterns act as attractors [3, 33], and groups pull deviating members back to their norms [10]. This is how a ball behaves in a bowl: it rolls toward the bottom, it loses energy on the way, and the bowl can be moved from outside. We take this picture literally and give the team an energy, which is high when the team is far from its usual way of working and which regulation removes over time.

Port-Hamiltonian (pH) systems give this picture a precise form [25, 34]. A pH model describes a system by the energy it stores, how that energy is dissipated, and the ports through which energy enters from outside, and it was first developed for physical systems such as circuits and robots. For teams, the energy is a modelling hypothesis, and so it is set as one. We call the model ANCHOR, after its central element: a resting point anchored on the team’s recent behavior, its own history and the norms of other teams. The model replaces the question “what comes next?” with three questions that a person can reason about:

  1. Where does this team rest?

  2. How does it return there?

  3. What moves it away?

Contributions.

  • A plant model of a human team. To our knowledge, ANCHOR is the first port-Hamiltonian model of human team interaction. Each of its 52 parameters has a plain meaning, and each part stands for a principle of team psychology (section 3).

  • A resting point with three anchors. The model learns how much each behaviour follows the moment, the team and other teams (sections 3.1 and 6.3).

  • Guarantees that hold by construction. Every forecast obeys an energy balance across phase switches, and it stays bounded under any spaced schedule of bounded cues, for every parameter value (§ 6).

  • A direct test on real teams. We measure how much of a real deviation survives over time, against the model and shuffled sessions, and we remove each part of the model to test its principle on held-out teams (§ 8.1 and 6.2).

  • A route to the closed loop. With far fewer parameters than a black box of equal accuracy, the model supports early warning and cue timing, which we evaluate in simulation under assumed cue effects (sections 6.4 and 6.5).

The System and Its Data

Sensing. We use the MURMR dataset [29], collected and described in recent work in the literature [6, 29]. 17 co-located teams of four, each member wearing a Meta Quest Pro, sorted 28 images into six emotion categories by discussion and consensus, with no time limit and no scripted turns. This image sorting task has been adopted in various group dynamics studies [6, 29]. Every 16 s we read a 32 s window that records how long each member spoke, how long each pair’s gaze hovered simultaneously on the same image, and how many seconds each pair stood within 0.46 m on the floor. Sessions have 21–74 windows (688 in total). The attention and proximity signals are sparse, covering about 2% and 4% of a window on average.

The state. The state of a team at window kk is a vector qk\mathbf{q}_k of 20 numbers between 0 and 1. Four hold the speaking time of each member, six the shared attention of each pair and six the proximity of each pair. The last four are team-level quantities measured directly from the window: mean speaking, attention density, proximity density and speaking balance. Balance is the normalized entropy of the speaking shares, so it is 1 when all four members speak equally and 0 when at most one speaks. Table 1 lists the blocks. We keep team quantities as separate coordinates because a team state is more than its parts, and section 6.2 shows that this matters.

BlockSizeEach entry
Speaking4Member’s speaking time
Attention6Pair’s simultaneous gaze on one image
Proximity6Pair’s time within 0.46 m
Team4Mean of each behavior; speaking balance

Table 1. The 20 numbers of the state qk\mathbf{q}_k for one window. Each entry is a share of the window, so it lies between 0 and 1.

Task progress. The MR application keeps task logs of who places which image in which category, and when. From them we compute the remaining work πk∈[0,1]\pi_k \in[0,1], the share of the 28 images not yet decided, where an image counts as decided once its label has stayed unchanged for 64 s. It uses only information available at window kk, so a live system can compute it.

The goal. From training teams we want to learn a map that takes the history of a new team up to window kk and forecasts its state q^k+h\hat{\mathbf{q}}_{k+h} for h=1,…,16h = 1,\ldots,16 (16–256 s ahead). The forecast must stay bounded when iterated and when cues are applied.

The ANCHOR Model

The model rests on six principles from team psychology.

P1 Teams have a resting point. Interaction settles into recurrent patterns, described as group equilibria [4], attractors [3, 33] or set points [13].

P2 Deviations are regulated back. After a disturbance, such as one member dominating, groups restore their pattern [4, 10].

P3 Team states emerge across levels. Team-level properties arise from members and pairs, but they are not functions of them [18, 20].

P4 Teams move through phases. Exploration and consolidation alternate or follow each other [12, 24].

P5 The task drives interaction. Coordination changes as decisions settle [24].

P6 Interventions must be timed. Coaching and cues work when delivered at the right moment [15].

Table 2 maps each principle onto one part of the model, which makes each one a hypothesis that held-out teams can confirm or reject. Readers from circuits or robotics may find the analogy of table 8 in section C helpful. The next subsections answer the three questions in turn: where the team rests (section 3.1), how it returns (§ 5.2) and what moves it (§ 5.3), and table 7 in section A lists every symbol.

Resting Point and Energy

The resting point q⋆\mathbf{q}^\star is the interaction pattern a team would settle into if nothing new happened. At any moment the team is somewhere else, because one member may be dominating or two members may have just huddled around an image. We write z=q−q⋆\mathbf{z} = \mathbf{q} - \mathbf{q}^\star for this deviation and measure it with an energy:

H=12zTKz.(1)H = \frac{1}{2}\mathbf{z}^{T}K\mathbf{z}. \tag*{(1)}

The matrix KK is the stiffness. It is positive definite, so HH is a true bowl whose only lowest point is q⋆\mathbf{q}^\star, and a larger KK makes the bowl steeper.

Each team has its own way of working, so ANCHOR computes the resting point at every forecast start as a blend of three anchors:

q⋆=wrec⊙qˉrec+wses⊙qˉses+wpop⊙qˉpop.(2)\mathbf{q}^\star= \mathbf{w}_{\mathrm{rec}} \odot\bar{\mathbf{q}}_{\mathrm{rec}} + \mathbf{w}_{\mathrm{ses}} \odot\bar{\mathbf{q}}_{\mathrm{ses}} + \mathbf{w}_{\mathrm{pop}} \odot\bar{\mathbf{q}}_{\mathrm{pop}}. \tag*{(2)}

@page 2/21 starts_mid=1 ends_mid=0 teams. The model replaces the question “what comes next?” with three questions that a person can reason about:

  1. Where does this team rest?

  2. How does it return there?

  3. What moves it away?

Contributions.

  • A plant model of a human team. To our knowledge, ANCHOR is the first port-Hamiltonian model of human team interaction. Each of its 52 parameters has a plain meaning, and each part stands for a principle of team psychology (section 3).

  • A resting point with three anchors. The model learns how much each behaviour follows the moment, the team and other teams (sections 3.1 and 6.3).

  • Guarantees that hold by construction. Every forecast obeys an energy balance across phase switches, and it stays bounded under any spaced schedule of bounded cues, for every parameter value (§ 6).

  • A direct test on real teams. We measure how much of a real deviation survives over time, against the model and shuffled sessions, and we remove each part of the model to test its principle on held-out teams (§ 8.1 and 6.2).

  • A route to the closed loop. With far fewer parameters than a black box of equal accuracy, the model supports early warning and cue timing, which we evaluate in simulation under assumed cue effects (sections 6.4 and 6.5).

The System and Its Data

Sensing. We use the MURMR dataset [29], collected and described in recent work in the literature [6, 29]. 17 co-located teams of four, each member wearing a Meta Quest Pro, sorted 28 images into six emotion categories by discussion and consensus, with no time limit and no scripted turns. This image sorting task has been adopted in various group dynamics studies [6, 29]. Every 16 s we read a 32 s window that records how long each member spoke, how long each pair’s gaze hovered simultaneously on the same image, and how many seconds each pair stood within 0.46 m on the floor. Sessions have 21–74 windows (688 in total). The attention and proximity signals are sparse, covering about 2% and 4% of a window on average.

The state. The state of a team at window kk is a vector qk\mathbf{q}_k of 20 numbers between 0 and 1. Four hold the speaking time of each member, six the shared attention of each pair and six the proximity of each pair. The last four are team-level quantities measured directly from the window: mean speaking, attention density, proximity density and speaking balance. Balance is the normalized entropy of the speaking shares, so it is 1 when all four members speak equally and 0 when at most one speaks. Table 1 lists the blocks. We keep team quantities as separate coordinates because a team state is more than its parts, and section 6.2 shows that this matters.

Task progress. The MR application keeps task logs of who places which image in which category, and when. From them we compute the remaining work πk∈[0,1]\pi_k \in[0,1], the share of the 28 images not yet decided, where an image counts as decided once its label has stayed unchanged for 64 s. It uses only information available at window kk, so a live system can compute it.

BlockSizeEach entry
Speaking4Member’s speaking time
Attention6Pair’s simultaneous gaze on one image
Proximity6Pair’s time within 0.46 m
Team4Mean of each behavior; speaking balance

Table 1 (PDF p. 2). The 20 numbers of the state qk\mathbf{q}_k for one window. Each entry is a share of the window, so it lies between 0 and 1.

The goal. From training teams we want to learn a map that takes the history of a new team up to window kk and forecasts its state q^k+h\hat{\mathbf{q}}_{k+h} for h=1,…,16h = 1,\ldots,16 (16–256 s ahead). The forecast must stay bounded when iterated and when cues are applied.

The ANCHOR Model

The model rests on six principles from team psychology.

P1 Teams have a resting point. Interaction settles into recurrent patterns, described as group equilibria [4], attractors [3, 33] or set points [13].

P2 Deviations are regulated back. After a disturbance, such as one member dominating, groups restore their pattern [4, 10].

P3 Team states emerge across levels. Team-level properties arise from members and pairs, but they are not functions of them [18, 20].

P4 Teams move through phases. Exploration and consolidation alternate or follow each other [12, 24].

P5 The task drives interaction. Coordination changes as decisions settle [24].

P6 Interventions must be timed. Coaching and cues work when delivered at the right moment [15].

Table 2 maps each principle onto one part of the model, which makes each one a hypothesis that held-out teams can confirm or reject. Readers from circuits or robotics may find the analogy of table 8 in section C helpful. The next subsections answer the three questions in turn: where the team rests (section 3.1), how it returns (§ 5.2) and what moves it (§ 5.3), and table 7 in section A lists every symbol.

Resting Point and Energy

The resting point q⋆\mathbf{q}^\star is the interaction pattern a team would settle into if nothing new happened. At any moment the team is somewhere else, because one member may be dominating or two members may have just huddled around an image. We write z=q−q⋆\mathbf{z} = \mathbf{q} - \mathbf{q}^\star for this deviation and measure it with an energy:

H=12zTKz.(3)H = \frac{1}{2}\mathbf{z}^{T}K\mathbf{z}. \tag*{(3)}

The matrix KK is the stiffness. It is positive definite, so HH is a true bowl whose only lowest point is q⋆\mathbf{q}^\star, and a larger KK makes the bowl steeper.

Each team has its own way of working, so ANCHOR computes the resting point at every forecast start as a blend of three anchors:

q⋆=wrec⊙qˉrec+wses⊙qˉses+wpop⊙qˉpop.(4)\mathbf{q}^\star= \mathbf{w}_{\mathrm{rec}} \odot\bar{\mathbf{q}}_{\mathrm{rec}} + \mathbf{w}_{\mathrm{ses}} \odot\bar{\mathbf{q}}_{\mathrm{ses}} + \mathbf{w}_{\mathrm{pop}} \odot\bar{\mathbf{q}}_{\mathrm{pop}}. \tag*{(4)}
QuestionPart of the modelWhat it means for a teamSensor source
Where does it rest?Resting point q⋆q^\star (P1)The pattern it returns to, anchored on its past and other teamsSociogram history
State across levels qq (P3)Team states emerge from members and pairs but are not reducible to themAudio, gaze, head pose
How does it return?Energy HH, regulation RσR_\sigma (P2)Distance from the pattern, and how fast deviations are regulated back—
Link gains in KK, RσR_\sigma, JσJ_\sigmaWho interacts with whom constrains coordinationTeam structure
What moves it away?Task input g ug\,u (P5)Task progress shifts the patternTask logs
Phase σ\sigma (P4)Exploration and consolidation regulate differentlyTask logs
Cue dd (P6)A well-timed prompt moves the teamDelivered by the system

Table 2. From team behaviour to ANCHOR, in plain terms. The principles P1–P6 and their sources are listed above, and table 4 gives the test of each part on held-out teams.

Here qˉrec\bar{q}_{\mathrm{rec}} is a weighted average of the last five windows (about 80 s), qˉses\bar{q}_{\mathrm{ses}} is the average of the session so far, and qˉpop\bar{q}_{\mathrm{pop}} is the average of the training teams or population. The symbol ⊙\odot multiplies element by element, and for each kind of behavior the three learned weights are non-negative and sum to one. Each anchor is a hypothesis: teams persist in what they just did (qˉrec\bar{q}_{\mathrm{rec}}), each team has its own style (qˉses\bar{q}_{\mathrm{ses}}), and teams doing the same task share norms (qˉpop\bar{q}_{\mathrm{pop}}). The learned weights show which hypothesis holds for which behavior.

A first look at real data. Before the formal evaluation, we test the bowl picture on a simple event, a member who suddenly speaks much more than usual. Figure 2 averages all 137 such bursts on held-out teams, with ANCHOR fitted without each team. Speaking falls from 46% at the burst to 27% 32 s later, against a forecast of 30%, so the ball pushed up the bowl rolls back at about the predicted rate. Over two minutes the forecast error is 14.1 points, against 17.0 for the session mean and 22.6 for persistence. Section 6 tests this across all behaviors.

Figure 2. Speaking bursts fade, and ANCHOR predicts how. Curves average all 137 bursts on held-out teams in which a member speaks at least 20 points above their session average. ANCHOR (blue), fitted without the team, follows the decay, whereas persistence assumes that the burst continues and the session mean assumes that it never happened.

Regulation and Exchange

The energy says how far the team is from rest, and the dynamics say how it moves. A PH system describes them by splitting the motion into parts with clear roles [34]:

q˙=(Jσ−Rσ)Kz+g u.(5)\dot{q} = \left( J_\sigma- R_\sigma\right)Kz + g\,u . \tag*{(5)}

The vector KzKz is the gradient of the energy, which points uphill in the bowl. The subscript σ\sigma is the team’s current phase, exploration or consolidation (P4), and g ug\,u is the push that task progress gives; section 3.3 says how both are computed.

Regulation RσR_\sigma is positive definite and removes energy. It is the term that brings a team back, and a deviation in one coordinate fades with a time constant of about 16 s/(RiiKii)116\ \mathrm{s}/(R_{ii}K_{ii})^{1}, so its entries read directly as recovery times. Exchange JσJ_\sigma is skew-symmetric (Jσ=−JσTJ_\sigma=-J_\sigma^\mathsf{T}), so it moves energy between coordinates without changing the total. It could describe talk turning into shared focus, and we keep it as a hypothesis to test.

The energy shows why these roles are clean. Differentiating eq. (1) along eq. (3) with q⋆q^\star fixed gives

H˙=(Kz)Tq˙=−(Kz)TRσ(Kz)+(Kz)Tg u,(6)\dot{H}=(Kz)^\mathsf{T}\dot{q}=-(Kz)^\mathsf{T}R_\sigma(Kz)+(Kz)^\mathsf{T}g\,u , \tag*{(6)}

since the exchange term (Kz)TJσ(Kz)(Kz)^\mathsf{T}J_\sigma(Kz) vanishes for a skew-symmetric JσJ_\sigma. Without task input the ball can only go down. This one line is the seed of the guarantees in section 4, and table 2 lists the principle that each part stands for.

From team structure to matrices. The state has 20 coordinates, so KK, RσR_\sigma and JσJ_\sigma are each 20×2020\times20, and learning every entry would take hundreds of parameters per matrix. It would also make the forecast depend on who happens to be called member “A”. We instead build them from the team’s structure, so that all coordinates playing the same role share one number (fig. 3a). These relations stay the same when members are renamed, so the model treats the four peers alike. We model the matrices as follows.

The stiffness matrix KK. Inside the stiffness, the idea is that related coordinates tend to move together. If A and B attend together, the other pairs that include A or B are pulled along (fig. 3a), and a member’s speaking is tied in the same way to the pairs that include them. We encode five such links: the attention and proximity of the same pair, attention pairs that share a member, proximity pairs that share a member, a speaker and the attention pairs containing them, and a speaker and the proximity pairs containing them. Each link makes it cost energy when two linked coordinates deviate differently, adding 12γℓ(za−zb)2\frac{1}{2}\gamma_\ell(z_a-z_b)^2 for each linked pair of coordinates aa and bb, with one learned gain γℓ≥0\gamma_\ell\ge0 per kind of link. Summed over one kind of link, these terms form the graph Laplacian LℓL_{\ell} of that link,1 and the full stiffness matrix KK is

What each term means for a team of four

Figure 3. What each term means for a team of four (A–D). Each kind of link in (a) has a single learned gain, which keeps the model small and independent of how members are labelled.

K=(diag⁡(eκS14,eκA16,eκP16)+∑ℓ=15γℓLℓ00diag⁡(eκT))+χLcons.(7)K = \left(\begin{array}{cc} \operatorname{diag}\left(e^{\kappa_S}1_4,e^{\kappa_A}1_6,e^{\kappa_P}1_6\right) + \sum_{\ell=1}^{5}\gamma_{\ell}L_{\ell} & 0 \\ 0 & \operatorname{diag}\left(e^{\kappa_T}\right) \end{array}\right) + \chi L_{\mathrm{cons}}. \tag*{(7)}

The top-left block acts on the 16 member and pair coordinates. Its first term gives each coordinate its own steepness, with one value shared by the four speaking coordinates (eκSe^{\kappa_S}), one by the six attention pairs (eκAe^{\kappa_A}) and one by the six proximity pairs (eκPe^{\kappa_P}); 1n1_n is a vector of nn ones, and the exponentials keep these values positive. Its second term adds the five links. The bottom-right block gives each of the four team coordinates its own steepness, the four entries of eκTe^{\kappa_T}. The last term ties the team coordinates to the member and pair coordinates with one gain χ≥0\chi\ge0. For speaking, attention and proximity it adds 12χ(t−m)2\frac{1}{2}\chi(t-m)^2, where tt is the team coordinate of that behavior and mm is the mean of its member or pair coordinates, so the team level stays close to what its members and pairs show. Speaking balance is an entropy rather than an average, so it is left out of this tie. Counting κS,κA,κP\kappa_S,\kappa_A,\kappa_P, the five γℓ\gamma_{\ell}, the four team steepnesses and χ\chi gives the 13 stiffness values of table 3. Each term is positive semidefinite and the diagonal is positive, so KK is positive definite for every value of these gains, which is what makes HH a true bowl without any constraint on training.

The regulation matrix RσR_{\sigma}. Regulation is simpler, since it has no links:

Rσ=diag⁡(eζσ,S14,eζσ,A16,eζσ,P16,eζσ,T).(8)R_{\sigma} = \operatorname{diag}\left(e^{\zeta_{\sigma,S}}1_4,e^{\zeta_{\sigma,A}}1_6,e^{\zeta_{\sigma,P}}1_6,e^{\zeta_{\sigma,T}}\right). \tag*{(8)}

Each coordinate gets its own recovery rate, with one value shared by the four speaking coordinates (eζσ,Se^{\zeta_{\sigma,S}}), one by the six attention pairs (eζσ,Ae^{\zeta_{\sigma,A}}) and one by the six proximity pairs (eζσ,Pe^{\zeta_{\sigma,P}}), while each of the four team coordinates has its own value, the four entries of eζσ,Te^{\zeta_{\sigma,T}}. The exponentials keep these values positive, so RσR_{\sigma} is positive definite. Each phase σ\sigma has its own set of seven values, which gives the 14 regulation values of table 3, and comparing the two sets directly asks whether a team recovers faster when consolidating than when exploring.

The exchange matrix JσJ_{\sigma}. Exchange links different behaviors rather than coordinates of the same kind:

Jσ=(jσ,SASSA+jσ,APSAP000).(9)J_{\sigma} = \left(\begin{array}{cc} j_{\sigma,SA}S_{SA}+j_{\sigma,AP}S_{AP} & 0 \\ 0 & 0 \end{array}\right). \tag*{(9)}

Here SSAS_{SA} has +1+1 in the row of each speaker and the column of each attention pair containing them, and −1-1 in the mirrored entry, so it can move energy between a member’s speaking and the shared attention of their pairs. SAPS_{AP} does the same between the attention and the proximity of each pair. The gains jσ,SAj_{\sigma,SA} and jσ,APj_{\sigma,AP} can take either sign, which sets the direction of the flow. The team block is zero, so exchange acts only on members and pairs. Because SSAS_{SA} and SAPS_{AP} are skew-symmetric, so is JσJ_{\sigma}, which is what prevents it from adding or removing energy. Each phase has its own two gains, which gives the 4 exchange values of table 3. These two links encode the hypothesis that talk draws members into shared focus and that standing together enables it.

Task, Phase, and Cues

Three inputs can move the team. Task progress enters through the port gg in eq. (3), where unu_n is the forecast change in remaining work hh steps ahead and g=Kbg=Kb sets how strongly task progress pushes each coordinate, with one learned value of bb per behavior and one per team coordinate (7 in all). A steady input moves the bottom of the bowl, and the team then follows at its own pace (fig. 3d). Phases change how the team regulates. We use two values of σ\sigma: exploration while at least half of the images are expected to be undecided, and consolidation otherwise. All phases share the same energy HH, because a shared energy keeps switching between stable modes stable [21]. A cue, such as a prompt that invites a quiet member to speak, is a kick that moves the deviation, z←z+dz \leftarrow z+d, and the model then predicts how long the kick lasts.

With the port, every part of the model is now in place. Table 3 counts its 52 parameters, each with a name and a meaning. A free symmetric KK alone would need 210 values3 while ANCHOR uses 13 parameters for KK. The full construction is in section C.

PartSymbols (values)Count
Stiffness KKκS,κA,κP\kappa_S,\kappa_A,\kappa_P (3), γ1…γ5\gamma_1\ldots\gamma_5 (5), κT\kappa_T (4), χ\chi (1)13
Regulation RσR_{\sigma}ζσ,S,ζσ,A,ζσ,P,ζσ,T\zeta_{\sigma,S},\zeta_{\sigma,A},\zeta_{\sigma,P},\zeta_{\sigma,T}(3+4)×2=14(3+4)\times2=14
Exchange JσJ_{\sigma}jσ,SA,jσ,APj_{\sigma,SA},j_{\sigma,AP}2×2=42\times2=4
Port g=Kbg=Kbbb3+4=73+4=7
Anchorswrec,wses,wpopw_{\mathrm{rec}},w_{\mathrm{ses}},w_{\mathrm{pop}}(3+4)×2=14(3+4)\times2=14
Total52

Table 3. Where the 52 free parameters come from, with the symbols of eqs. (5) to (7). Each value is shared by all coordinates of one behavior (3) or belongs to one team coordinate (4), regulation and exchange have one set per phase (2), and 2 of the 3 anchor weights are free because they sum to one.

Forecasting and Learning

A forecast predicts where the team will be in each of the next windows if nothing new happens apart from the task. ANCHOR computes it by rolling the ball forward one window at a time. It starts from the current deviation z0=qk−q⋆z_0=q_k-q^\star, with q⋆q^\star and all matrices held fixed for that forecast. At each step, regulation shrinks the deviation, exchange moves part of it between behaviors and the task pushes it, and the forecast is the resting point plus whatever deviation is left.

Each step uses the implicit midpoint rule, which evaluates the motion halfway through the step rather than at its start. This keeps the energy balance of eq. (6) exact, whereas a plain Euler step can add energy that the model does not have [16]. Because q⋆q^\star is fixed, eq. (5) reads z˙=Aσz+g u\dot{z}=A_\sigma z+g\,u with Aσ=(Jσ−Rσ)KA_\sigma=(J_\sigma-R_\sigma)K, which is linear in zz, so each step is a matrix multiplication:4

zh+1=Cσzh+Gσuˉh,q^k+h=q⋆+zh(10)z_{h+1}=C_\sigma z_h+G_\sigma\bar{u}_h,\qquad\hat{q}_{k+h}=q^\star+z_h \tag*{(10)}

Without task input, zh+1=Cσzhz_{h+1}=C_\sigma z_h, so CσC_\sigma gives the part of the current deviation that is left one window later. The second term adds the task’s push, since GσG_\sigma turns the task input uˉh\bar{u}_h during the window into a change of each coordinate. Explicitly, Cσ=(I−12Aσ)−1(I+12Aσ)C_\sigma=(I-\frac{1}{2}A_\sigma)^{-1}(I+\frac{1}{2}A_\sigma) and Gσ=(I−12Aσ)−1gG_\sigma=(I-\frac{1}{2}A_\sigma)^{-1}g, and uˉh=12(uh+uh+1)\bar{u}_h=\frac{1}{2}(u_h+u_{h+1}) is the task input halfway through the step. For one coordinate on its own, CσC_\sigma reduces to the factor (1−a2)/(1+a2)(1-\frac{a}{2})/(1+\frac{a}{2}) with a=RiiKiia=R_{ii}K_{ii}, which is close to e−ae^{-a} when aa is small. Far ahead, every forecast settles at the bottom of the bowl, so the resting point sets long-range accuracy and the dynamics set the path there.

Learning. The parameters are chosen so that forecasts match what the training teams actually did over the next few windows. We learn all parameters θ\theta by minimizing the error of forecasts up to a training horizon of htrh_{\mathrm{tr}} windows,

L(θ)=∑k∑h=1htr∥q^k+h(θ)−qk+h∥ν2(11)\mathcal{L}(\theta)=\sum_k\sum_{h=1}^{h_{\mathrm{tr}}}\left\|\hat{q}_{k+h}(\theta)-q_{k+h}\right\|_\nu^2 \tag*{(11)}

where ∥x∥ν2=∑cνcxc2\lVert x\rVert_\nu^2=\sum_c \nu_c x_c^2. The weights matter because the behaviors live on very different scales. Shared attention needs two members’ gaze on the same image at the same moment, so it fills only about 2% of a window on average, against about 10% for a member’s speaking, and its variance is about 14 times smaller (calculated from MURMR dataset [29]). With a plain squared error, a forecast that ignored attention would cost little, and the fit would serve speaking and proximity. Setting νc\nu_c to the inverse training variance of coordinate cc’s behavior divides each error by that behavior’s typical spread, so every behavior counts equally. The loss needs no energy or stability term, because the structure already guarantees these properties.

Guarantees for Closed-Loop Use

A forecaster that is iterated on its own output must stay bounded for a planner built on it to be trustworthy. ANCHOR stays bounded by construction, because of the form of its matrices rather than their fitted values. We state the two results here, with formal proofs in section D, and then show how the model runs in a live loop.

Proposition 1 (Energy balance). For every phase sequence and every input, each forecast step satisfies

Hh+1−Hh=−vTRσv+vTg uˉh(12)H_{h+1}-H_h=-v^\mathsf{T}R_\sigma v+v^\mathsf{T}g\,\bar{u}_h \tag*{(12)}

where v=12K(zh+zh+1)v=\frac{1}{2}K(z_h+z_{h+1}) is the energy gradient KzKz of eq. (6) at the midpoint of the step.

In words, regulation removes energy, exchange conserves it, and the task alone adds it. Without task input, the forecast therefore stays inside the bowl where it started, however far it runs and however often the phase switches. On the other hand, a cue adds deviation directly between steps, as treated next. The proof is the discrete form of eq. (6): the midpoint rule makes the step use one gradient vv, so the exchange term vanishes exactly, and only the current step’s matrices appear, which is why any phase sequence is allowed.

Because every regulation value is positive, each step without task input also removes a positive share of the energy in every direction. We measure the distance from rest with the energy norm ∥z∥K=zTKz=2H\lVert z\rVert_K=\sqrt{z^\mathsf{T}Kz}=\sqrt{2H}, built from the energy of eq. (1). In each step this norm shrinks by at least the factor ρ=max⁡σ∥Cσ∥K<1\rho=\max_\sigma\lVert C_\sigma\rVert_K<1, the largest fraction of this distance that can survive after one step (lemma 1 in section D).

Proposition 2 (Bounded cue schedules). Without task input, let a cue dhd_h enter step hh as zh+1=Cσ(zh+dh)z_{h+1}=C_\sigma(z_h+d_h), with dh=0d_h=0 between cues, and let cues satisfy ∥dh∥K≤δ\lVert d_h\rVert_K\leq\delta, so that no cue gives a team at rest (z=0z=0) more energy than 12δ2\frac{1}{2}\delta^2, and be at least τ\tau steps apart. Then for every phase sequence and cue schedule,

∥zh∥K≤ρh∥z0∥K+δ1−ρτ(13)\lVert z_h\rVert_K\leq\rho^h\lVert z_0\rVert_K+\frac{\delta}{1-\rho^\tau} \tag*{(13)}

In words, every step still shrinks whatever deviation is there by at least the factor ρ\rho, but a cue first adds a kick of size at most δ\delta, and the step then acts on the kicked deviation, so the deviation can grow at a cue. Between cues the dynamics remove deviation, and the bound says that this balance stays below δ/(1−ρτ)\delta/(1-\rho^\tau), so spaced cues cannot pile up into a runaway prediction. The proof sums this fading: the rr-th most recent cue contributes at most δρτr\delta\rho^{\tau r}, and the geometric series gives the bound on all cues together (full proof in section D). This is what a planner needs where any schedule it considers has a known worst case. We express such bounds in cues’ worth, multiples of δ\delta. The bounds scale with δ\delta, so they hold whatever the size of a real cue turns out to be.

When the resting point moves. Both results describe one forecast with a fixed resting point q⋆q^\star. A live system recomputes q⋆q^\star at every window, so cues that change behavior also move the bowl. Section E extends the analysis to this case. The same bound holds for all time if the resting point leans more on other teams (a larger wpopw_{\mathrm{pop}} in eq. (2)), which pins the bowl in place but makes forecasts Algorithm 1 One pass of the live loop, every 16 s.

Require: fitted parameters θ\theta, breakdown thresholds ϑj\vartheta_j, minimum cue spacing τ\tau

  1. Read the new window qkq_k and the remaining work πk\pi_k.

  2. Update the three anchors and the resting point q⋆q^\star by eq. (2).

  3. z0←qk−q⋆z_0 \leftarrow q_k-q^\star; roll eq. (10) forward to obtain q^k+1,…,q^k+16\hat{q}_{k+1},\ldots,\hat{q}_{k+16}.

  4. if a team coordinate jj of q^\hat{q} falls below ϑj\vartheta_j then

  5. Raise a warning.

  6. end if

  7. if no cue was issued in the last τ\tau windows and the forecast with a kick dd improves on the forecast without it by enough then

  8. Issue the cue dd.

  9. end if

less accurate, because the model then ignores more of the team’s own behavior (proposition 6 in section F). The energy norm lets a designer compute this trade-off between forecast accuracy and how far repeated cues can move the team before deployment.

The live loop. In a live system the team is the plant, the headsets are the sensors and the prompts are the actuators, and ANCHOR is the model inside the controller. Algorithm 1 shows how it runs, and Figure 10 in section C draws the same loop as a block diagram. All computation is closed form, so a forecast and a refit are cheap enough to run within the 16 s stride.

The closed loop

Figure 10. The closed loop. The team is the plant, and ANCHOR is the model inside the controller. In this paper the cue arrow is simulated, since the recorded sessions contain no cues.

Evaluation Setup

We evaluate ANCHOR as a model of team dynamics through five questions. First, do real teams behave as the model assumes, drifting back toward a resting point after a deviation (section 6.1)? Second, is each part of the model, and so each principle, needed to forecast teams it has not seen (section 6.2)? Third, do the fitted parameters give a readable account of how teams work (section 6.3)? Fourth, is the model as accurate as larger models without its structure (section 6.4)? Fifth, are its forecasts useful in the closed loop, for early warning and for timing cues (section 6.5)? The rest of this section explains our evaluation setup.

Teams the model has not seen. Every result is measured on teams left out of training. We use leave-one-team-out cross-validation. From every window k≥5k \ge5 of the held-out team, once five windows have been observed, each model forecasts the next 16 windows (256 s). ANCHOR is trained with htr=8h_{\mathrm{tr}}=8 windows (128 s), so the evaluation also tests forecasts beyond the training horizon.

Error and comparison. Error is the normalized mean squared error (nMSE), scaled by the training variance so that 1 means an error as large as the data’s natural variation. We report it separately for the 16 member and pair coordinates and for the four team coordinates. To compare two models, we compute each model’s error on every held-out team and test whether one is consistently lower across the 17 teams with a paired Wilcoxon signed-rank test, which makes no assumption about how the errors are distributed. We fitted several variants of ANCHOR, which differ in how the matrices are built and in which optional terms they include (section C), and we report the variant with the lowest combined member-and-pair and team error.

Baselines and controls. To judge accuracy, we compare ANCHOR with simple rules and with learned models that have no built-in structure. The rules are persistence, the recent average and the session mean. The learned models are an iterated VAR(2), a direct multi-horizon ridge regression and a small MLP. They also train on symmetry-augmented data, in which renaming the members gives 23 further valid copies, because ANCHOR has this symmetry built in. An attractor-only control predicts the learned resting point without dynamics, which isolates the dynamics add.

Results

The results answer the three questions of section 3 as follows, and table 4 summarizes them part by part.

  • Where does the team rest? At a point blended from the moment, the team and other teams, in proportions that differ by behavior. This is the model’s strongest element.

  • How does it return? By sliding back without swinging, with recovery times of about 20–60 s.

  • What moves it away? Task progress and phase, in the expected direction; a deployment will measure the effect of cues.

Do real teams behave like a ball in a bowl?

How we measure fading. The guarantees describe the model, and whether the picture also holds for real teams is an empirical question. To test it directly, we ask how much of a team’s deviation from its resting point survives over time. For every window kk of a held-out team, we compute zk=qk−qk⋆z_k=q_k-q_k^\star and the later deviation zk,h=qk+h−qk⋆z_{k,h}=q_{k+h}-q_k^\star from the same resting point, and we project one on the other:

mh=∑kzkTKzk,h∑kzkTKzk.(14)m_h=\frac{\sum_k z_k^{\mathsf{T}}Kz_{k,h}}{\sum_k z_k^{\mathsf{T}}Kz_k}. \tag*{(14)}

This mhm_h is the share of a deviation that remains after hh windows. A value of 1 means that the deviation has not changed, 0 means that it is gone, and a negative value would mean that the team overshoots to the other side. Window noise averages out because the projection is signed. As a control test, we shuffle the order of the windows in each session and repeat the calculation. The shuffled session has the same states but no order in time, so its mhm_h is close to zero, and any fading that real teams show behind it comes from how they move through time.

Real teams fade toward rest. Real deviations fade the way a ball settles in a bowl, and on average close to the forecast (Figure 4). After 32 s, the first step free of window overlap, 24% of a deviation remains against a forecast of 30%. Each further 16 s removes about 30% of what remains, and a constant fraction is what a ball settling in a bowl should show. Most of a deviation has faded after two minutes, which is why a training horizon of htr=8h_{\mathrm{tr}}=8 windows (128 s) suffices. Shuffled sessions, whose resting points are computed in the same way, drop to about −1%-1\% at once, so the gradual fading of real teams comes from how they move through time rather than from how the resting point is computed.

Figure 4. Real deviations fade, on average close to the forecast. Share of a deviation that remains hh windows later, mhm_h, on held-out teams. Left: observed (band: interquartile range over teams), the ANCHOR forecast, and sessions with shuffled windows; the inset zooms on 24–56 s, around the 32 s values compared in the text. The 16 s point shares half its window with the start (fig. 9 in section B), so comparisons start at 32 s. Right: the same observed quantity for each behavior separately, computed from that behavior’s coordinates alone (with KK replaced by the identity).

Reading the match with the forecast. The forecast comes from models fitted without each team and trained on the forecast error of eq. (11) rather than on the mhm_h of eq. (14). Because least-squares training of a linear relaxation tends to reproduce the average decay of its training data, the autoregressive control and the differences between behaviors, which follow, give the sharper tests.

  • Against an autoregressive control. Real teams also fade faster than a simple autoregressive process would. Because any autocorrelated process fades, we also compare with a stronger control, an AR(1) surrogate fitted to each coordinate of each session. Measured the same way, the AR(1) surrogate keeps 37% of a deviation after 32 s, whereas real teams keep 24%, and real teams keep less in all 17 teams (p=1.5e−5p = 1.5e-5). At team level the values are 29% and 17% (real teams lower in 13 of 17, p=0.008p = 0.008). Real deviations therefore fade faster than autoregression predicts, partly because overlapping windows inflate its fitted lag.

  • By behavior. The match is closest for the team-level state, where 17% remains at 32 s against a forecast of 17% (p=0.96p = 0.96). Speaking (24% against 35%, p=0.06p = 0.06) and proximity (29% against 39%, p=0.01p = 0.01) fade slightly faster at 32 s and match from 48 s on. Pair attention fades fastest, with 10% remaining against a forecast of 31% ($p < 0.001), so the model is conservative for attention up to 96 s, consistent with brief, sparse hours that make a signal look as if it fades at once.

Does each principle hold?

By a principle we mean one of the claims from team psychology that ANCHOR builds in (P1–P6 in section 3), each carried by one part of the model (table 2). A principle holds if removing its part makes forecasts of unseen teams worse. By this test, four parts are needed: the blended resting point, the dynamics, the team-level state and the shared link structure; the others are on par and carry the guarantees and input channels. Table 4 removes each element in turn and retrains, which is the central test of the paper because it checks every part on teams the model has not seen. To read the table, take the first row: when the resting point uses recent behavior only, error rises by 19% for members and pairs, ANCHOR is better for all 17 teams, and the difference is significant.

PartHypothesisAlternative testedMember and pair nMSETeam nMSEVerdict
Resting point q⋆q^\star (P1)Teams return to a point set by recent behaviour, their own history and other teamsq⋆q^\star from recent behaviour only+19.3% (17/17, 2e-5)+16.5% (14/17, 0.004)supported
q⋆q^\star from session history only+3.4% (14/17, 0.035)+9.7% (14/17, 0.020)supported
Regulation RσR_\sigma (P2)Deviations fade, so modelling the return helps at 16–48 sattractor only, no dynamics+11.3% (17/17, 2e-5)+8.5% (15/17, 5e-4)supported
Team state qq (P3)Team states are not reducible to members and pairscomputed from member forecast–+270.3% (16/17, 3e-5)supported
Shared link gains in KK, RσR_\sigma, JσJ_\sigmaCoordinates with the same role can share one valueunconstrained matrices+14.4% (16/17, 4e-4)+9.7% (17/17, 2e-5)supported
Phase σ\sigma (P4)Teams regulate differently in exploration and consolidationsingle phase+0.1% (9/17, 0.782)-0.4% (7/17, 0.611)on par
Task input gug_u (P5)Task progress moves the resting pointno progress port+1.6% (9/17, 0.487)+2.2% (9/17, 0.644)on par
Exchange JσJ_\sigmaInteraction passes between behaviours without loss–gains ∣Jσ∣≤0.05\lvert J_\sigma\rvert\le0.05; regulation dominates small
Sensed damping D(t)D(t)Dominance and conflict add dampingno sensed damping-0.0% (8/17, 0.459)-0.1% (7/17, 0.431)on par
Inertia MMTrends persist across windowsmomentum added+0.2% (11/17, 0.487)+0.2% (5/17, 0.579)on par
Cue dd (P6)Well-timed cues move a team–simulated in section 6.5simulated

Table 4. The hypothesis behind each part of the model and what held-out teams say about it. Each test replaces one part with the alternative shown and retrains the full model, which here also contains the sensed damping. Cells give the change in held-out nMSE, so a positive change means the part helps, followed by the number of teams for which ANCHOR is better and the paired Wilcoxon pp. Errors are over 16–256 s, except for P2 (16–48 s, where dynamics act) and P3 (speaking balance). Shaded: supported (p<0.05p < 0.05). On par: the reduced model reaches the same accuracy, and these parts are kept because they carry a guarantee or the channel for task progress and cues.

The table shows that the resting point must blend its anchors and that dynamics add to it. Team states must be modelled at the team level. A given member is silent 44% of the time, so a forecast of expected speaking times looks evenly shared even when one member dominates, and balance derived from the member forecast has 3.7 times the error of forecasting it directly (2.86 against 0.77), which is what multilevel team theory predicts [20]. The results for a resting point from recent behavior only, the dynamics, speaking balance and the shared structure also hold after correcting for the 17 tests of the table (section F.1).

The other elements answer the third question simply. Phases, the task port, sensed damping and inertia are on par with the reduced model, and the exchange gains stay within ∣Jσ∣≤0.05|J_\sigma| \le0.05, so at this resolution teams slide back to rest rather than swing. The phase changes once in 15 of 17 sessions, typically 5 minutes in, so 11% of forecasts meet a switch within 48 s (section F.3). We keep the phases and the port because they carry the guarantees and the channel for the task and for cues, and the selected model leaves out the sensed damping.

What does the fitted model say about teams?

Each behavior rests on a different anchor: shared attention on the team’s own history, proximity on habit and norms, and speaking balance on the moment. The fitted model therefore reads like a description of how teams work (fig. 6). Who attends together is a trait of the team (session weight 0.81), while how close members stand mixes the team’s own habit with other teams (weights 0.41 and 0.50). How evenly they talk belongs to the moment: the resting balance follows the moment (0.43) and other teams (0.56), with almost no weight on the team’s own history (0.02), and it recovers fastest, in 12–22 s. Figure 5 shows this on one held-out team: over the session the resting balance moves between 0.24 and 0.54, following

Figure 5. The bowl moves. Speaking balance of held-out team 8 (the longest session), with the resting point q⋆q^\star recomputed every window as the blend of its three anchors (fitted weights in parentheses). Bottom: this coordinate’s bowl, 12Kii(x−q⋆)2\frac{1}{2}K_{ii}(x-q^\star)^2, at five marked moments that include the lowest and highest resting point, and where the team actually was (dot). The bottom slides with the conversation while the team moves up and down its walls.

Figure 6. Reading a team off the fitted model. Left: what each behavior’s resting point follows, from the learned anchor weights. Right: recovery time constants in each phase. Table 15 in section F.3 lists all values.

the moment but pulled back toward other teams (0.44). Recovery is fast overall, with time constants in consolidation of 17.0 s for shared attention, 25.9 s for speaking and 27.6 s for proximity, and it is slower in exploration. As images are decided, the fitted port moves the resting point toward more shared attention and more co-presence, and resting balance rises by about 0.25 on its 0–1 scale from start to finish, a descriptive reading of the fitted port.

How accurate is ANCHOR?

ANCHOR performs on par with augmented ridge, with 52 parameters against about 40,000 (table 5). Its 16–256 s error is the lowest at both levels, with ridge at +1.5% and +1.4% higher error ($p = 0.306 and 0.487), so the two are on par. Persistence, which assumes that the team stays as it is now, has about half again ANCHOR’s error over 16–256 s (+49.2% for members and pairs, +43.5% for the team). The other learned baselines, including a GRU sequence model, have significantly higher error over 16–48 s. Error falls in clear steps as parts are added (member and pair level, 16–48 s), from +23.1% higher error than ANCHOR for the recent average, with nothing fitted, to +11.3% for the resting point alone (21 values) and +1.7% when each behavior decays toward it at its own rate (seven more). ANCHOR’s remaining 24 values, for stiffness, links, phases and exchange, close this last gap (p=0.006p = 0.006), and at team level the two tie (+0.0%). With the same anchors, 434 free values have +1.8% and +4.4% higher error.

Figure 7 shows the error by horizon. Over 16–48 s, ANCHOR also beats the attractor-only control, which has +11.3% and +8.5% higher error (p=2e−5p = 2e-5 and 5e−45e-4), whereas the two meet far ahead because the forecast is then simply where the team relaxes to. The dynamics therefore supply the path from the present to the resting point, which is the horizon at which a system would act. ANCHOR stays significantly better than assuming an average team up to 96 s (pairs) and 80 s (team) after correcting for the 16 horizons tested (section F.1). Because the structure is small, ANCHOR learns from few teams: on mean error it stays below augmented ridge at every training-set size, significantly up to eight teams at member and pair level and four at team level (Figure 14 in section F.2). With very few teams, the resting point alone gives a robust forecast.

Figure 7. Forecast error (nMSE, lower is better) by horizon on held-out teams, against the iterated and simple baselines (augmented ridge is in Table 5), for the 16 member and pair coordinates (left) and the four team coordinates (right). Grey: attractor-only control; shaded: the first 16–48 s. Values above 1 at long horizons still beat assuming an average team (section F.2).

ModelParamsMember and pair 16–48 sMember and pair 16–256 sTeam 16–48 sTeam 16–256 sVerdict
Attractor only21+11.3+1.5+8.5+0.1better
Persistence0+38.8+49.2+32.1+43.5better
Recent average0+23.1+22.6+16.2+22.1better
Session mean0+28.0+13.1+46.4+36.4better
VAR(2)820+15.3+10.4+16.6+11.2better
VAR(2) + aug.820+6.4+8.3+7.6+9.0better
Direct ridge43,840+28.0+15.3+23.2+14.2better
Ridge + aug.43,840−1.1+1.5−1.1+1.4on par
MLP14,944+39.3+14.7+44.9+25.9better
MLP + aug.14,944+33.9+10.0+32.3+15.2better
GRU27,168+38.7+15.7+36.8+15.7better
GRU + aug.27,168+28.9+7.3+26.1+8.5better
Anchored scalar28+1.7+0.2+0.0+0.4better
Anchored free434+1.8+0.7+4.4+1.9better
Anchored free + aug.434+1.5+0.2+0.1+0.8better

Table 5. Forecast error of each model relative to ANCHOR on held-out teams: how much higher its error is, averaged over teams (%; positive: higher error than ANCHOR). Bold: significant (paired Wilcoxon over the 17 teams, p<0.05p < 0.05). Verdict for ANCHOR: better if significantly better in at least one column and worse in none; on par if no column is significant. The anchored models use ANCHOR’s fitted anchors, which favors them. Win counts, pp-values and absolute errors are in tables 13 and 14 in section F.2.

Figure 14. Error versus the number of training teams, as the median over held-out teams and random training subsets. With very few teams the resting point alone is the safest forecast. ANCHOR, looser ties and ANCHOR, free matrices are the permutation-equivariant and unconstrained variants of section C.

Can forecasts give early warning and cue timing?

Guarantees on the fitted model. On the fitted model, a cue’s predicted effect fades within about a minute, and repeated cues stay within a few cues’ worth. The fitted contraction factor is ρ=0.814\rho= 0.814, so the predicted effect of one cue halves within 54 s even in the slowest direction, team proximity density during exploration, and faster for speaking (Figure 2). With one cue every 32 s, the bound of proposition 2 is 3.0 cues’ worth. A single cue therefore gives a short nudge and does not change where the team rests, unless the team’s own later behaviour moves the anchors. In 60-window (16 min) rollouts for all held-out teams, with phase switching and task input, the energy balance held at 100% of steps, and every forecast coordinate stayed between −0.79 and 0.79, so the forecasts stayed bounded even 16 min ahead; Figure 15 in section F.3 shows one such forecast. In the live loop, where the resting point is recomputed, the computed bound is about 6.8 cues’ worth over 16 min and grows slowly with time (section E). This yields a design insight that shared attention has almost no population weight (0.012), so sustained cues on attention gradually become part of what the model treats as normal.

Figure 15. The energy balance on one held-out forecast. Left: a 24-window forecast (thick) from the dotted origin, with the observed values (faint), for three coordinates on their 0–1 scale. Right: the forecast’s energy HkH_k, measured from the resting point at the origin, with task input (solid) and for the same forecast without task input (dash-dotted), and the bound of proposition 1, the initial energy plus the energy supplied by task progress (dashed). Without input the energy only falls; with input every rise comes from the task and stays below the bound.

Early warning. ANCHOR anticipates conversation collapse well, mostly because it knows where each team rests. A breakdown is a team coordinate falling below the 25th percentile of the training teams: conversation collapse for speaking balance (below about 0.2, close to one effective speaker), loss of shared focus for attention density, and low proximity for proximity density. A warning is a forecast that a breakdown absent now will begin within the next 64 s. At this threshold, ANCHOR warns of conversation collapse with AUROC 0.85, catching 74% of onsets at a 20% false-alarm rate, and of loss of shared focus with AUROC 0.76. At the 15th and 35th percentiles collapse AUROC is about 0.75, and ANCHOR stays within 0.02 of the best model (Table 18 in section F.4). The attractor-only control ties it on both, so most of this signal comes from knowing where the team rests; plausibly, teams whose resting balance sits near the threshold are the ones that collapse. For low proximity, ANCHOR reaches 0.66 and augmented ridge 0.68. Table 6 gives all three. For chronically uneven teams, an alarm level calibrated per team would reduce extra alarms (Figure 16 in section F.4). At about one false alarm per 80 s of healthy time, which suits a facilitator’s display.

ModelConversation collapse AUCConversation collapse Det.Conversation collapse WarnLoss of shared focus AUCLoss of shared focus Det.Loss of shared focus WarnLow proximity AUCLow proximity Det.Low proximity Warn
ANCHOR (ours)0.8574%310.7657%350.6634%29
Attractor only0.8573%310.7656%350.6228%29
Recent average0.8368%320.7658%370.6326%28
VAR(2) + aug.0.8064%300.6943%360.6539%28
Ridge + aug.0.8060%310.7142%350.6843%32
MLP + aug.0.7450%310.7251%370.5625%33

Table 6. Warning of a breakdown that starts within the next 64 s and is absent now, on held-out teams. AUC: AUROC (0.5 = chance); Det.: share of onsets detected when alarms fire at 20% of origins without a breakdown; Warn: mean warning time (s). Shaded: ANCHOR; bold: best in each column.

Figure 16. Early warning on the longest session (team 8). Grey: observed speaking balance; blue: ANCHOR forecast made 32 s earlier, which follows the slow drift of the balance rather than its window-to-window swings; dashed: collapse threshold; shaded: windows in collapse; triangles: alarms at the pooled 20% false-alarm level. Most collapses are preceded by an alarm, but this chronically uneven team also triggers many alarms that are not followed by a collapse.

ThresholdCollapseFocusLow prox.Predictive bestvs. periodic
15th0.75 / 0.760.77 / 0.770.65 / 0.695/123.9×
25th (main)0.85 / 0.850.76 / 0.760.66 / 0.688/122.2×
35th0.75 / 0.770.77 / 0.790.66 / 0.698/121.4×

Table 18. Sensitivity to the breakdown threshold (training percentile). Early warning: AUROC of ANCHOR / best other model for conversation collapse, loss of shared focus and low proximity. Intervention: settings (of 12) in which ANCHOR-predictive is the most cue-efficient policy, and its breakdown time removed per cue relative to periodic cueing at β=0.5\beta= 0.5 s.d. The remaining settings are won by reactive cueing, and at the 35th percentile also by the forecast policy.

Timing cues, in simulation. If cues work, timing them by forecast gets more out of each one than cueing on a schedule. The recorded sessions contain no cues, so we simulate cues whose effect we assume and sweep, on two plants: ANCHOR itself and an independently learned VAR that the controller has never seen. Each episode starts from the first six windows of a held-out team and runs for 40 windows on the plant, which steps forward from its own forecast

plus an error drawn from its one-step errors on real data. A cue adds an assumed β\beta training standard deviations to speaking balance or to shared attention in the window it is issued, and all policies keep at least two windows between cues (protocol in section F.5). We compare four policies. Periodic cueing issues a cue every eight windows, whatever the team does. Reactive cueing issues a cue once the breakdown is visible in the current window. Forecast cueing issues a cue when the lowest forecast over the next 48 s falls below a level calibrated on training teams. ANCHOR-predictive compares the forecast with and without a cue, and it issues a cue when the predicted drop in the coming shortfall exceeds such a level. At an assumed effect of 0.5 training standard deviations, it removes 1.51 points of breakdown time per cue, which is 2.2 times as much as periodic cueing (0.70) and 1.3 times as much as reactive cueing (1.17), pooled over plants (Figure 8). It is the most efficient policy in 6 of 6 settings on its own plant and in 2 of 6 on the VAR plant, where reactive cueing is competitive. Selectivity trades total effect for efficiency: for balance cues on ANCHOR’s plant, from 13.3% of time in breakdown, periodic cueing reaches 11.4% with 4.7 cues per 10 min and ANCHOR-predictive reaches 12.1% with 1.8 (Table 19 in section F.5). These results show how well a policy times its cues for a given cue effect, and a deployment will measure that effect.

Figure 8. Simulated cues under an assumed effect β\beta. Breakdown time removed per cue (percentage points of the episode; higher is better), on ANCHOR itself (top) and on an independently learned VAR plant (bottom). ANCHOR-predictive (thick blue) is on top in 6 of 6 settings on its own plant and in 2 of 6 on the VAR plant.

CuePlantNonePeriodicReactiveForecastANCHOR-predictive
BalanceANCHOR13.311.4 (4.7)11.4 (3.8)10.8 (5.1)12.1 (1.8)
AttentionANCHOR11.55.7 (4.7)5.6 (2.3)5.3 (1.9)5.6 (1.7)
BalanceVAR12.611.0 (4.7)10.4 (3.4)10.4 (4.8)11.5 (1.7)
AttentionVAR12.57.6 (4.7)7.0 (2.8)9.0 (1.9)9.7 (1.5)

Table 19. Time in breakdown and cue budget at cue effect β=0.5\beta= 0.5 s.d. Each cell gives the share of the episode spent in the breakdown (%, lower is better), with the number of cues per 10 min in parentheses. Shaded: ANCHOR-predictive, the most selective policy.

Related Work

Human-in-the-loop CPS. Schirner et al. [31] frame systems that sense people and act with them as human-in-the-loop CPS, in which the human is part of the loop rather than an external user. Prior systems adapt to one person, for example by timing collision warnings to a driver [9] or by planning around a driver’s latent trust [32], and recent work keeps decisions fair across the several people a system serves [37, 39] and spreads privacy risk evenly across the people whose data train it [38]. We study a team as the plant, where the state lies in the relations among people and the actuator is a prompt that they may or may not follow.

Sensing and predicting teams. Wearable sensing showed that turn-taking and proximity relate to team outcomes [26, 36], and smartphones can infer who converses with whom [22]. In collaborative MR, MURMR [29] supplies the sensing pipeline and data we use, S2S [6] relates sensed signals to perceived collaboration, and MoCoMR [28] simulates individual behavior. TeamLLM [30] predicts the next window with language models, but it degrades sharply when iterated. Classical social-influence models let opinions relax toward a weighted consensus anchored by individual prejudices [7, 11, 27], and dyadic models let partners relax toward a set point [13]. ANCHOR shares their spirit, and it extends the idea to multimodal interaction at three levels, with learned anchors and formal guarantees.

Physics-informed learning and passivity. Hamiltonian and dissipative networks use energy structure as a prior for physical systems [14, 40], and port-Hamiltonian networks add dissipation and external forcing [8]. Other work builds stability into deep dynamics by learning a Lyapunov function with the model [23], or verifies neural Lyapunov and barrier certificates formally [1]. Passivity and dissipativity are also established tools for designing CPS [2], including switched systems under quantization [41], and we carry the same reasoning to a human plant. Teams have no known energy law, so here each part of the structure is tested on held-out data (Table 4). A shared energy is the classical route to stability under arbitrary switching [21], and the pH form builds it in [34].

Discussions and Conclusion

Why port-Hamiltonian? At this resolution the data select a simple member of the family, in which regulation dominates exchange, and a scalar relaxation toward the same resting point comes close to ANCHOR (section 6.4). Such a relaxation, with decay between 0 and 1, is itself a pH system with Jσ=0J_{\sigma}=0 and diagonal KK and RσR_{\sigma}, so the data choose within the family. What they chose is the family itself. Its guarantees hold for every parameter value, whereas a free linear map, which was stable in all 170 retrainings on 2–16 teams (radius at most 0.92), is checked after each fit. It names each part after a claim about teams, which turns model design into a set of experiments. Richer effects such as exchange and inertia also fit inside the same energy balance, so they can be tested without losing the guarantees.

What the picture says. Over the minutes-long course of a session, team interaction is well described by relaxation toward an anchored resting point. This gives ideas from team psychology a computable form, in which equilibria, self-regulation and emergent states become numbers that can be estimated and tested. We offer the energy picture as a description of what teams do on average, not as a claim that they literally are pH systems.

Scope. The study covers one task, one room size and four-person peer teams, so the fitted anchors and gains describe this task, and the contribution is the structure, which carries to new tasks by refitting. The configuration was selected on the evaluation folds, and several neighboring configurations are within 1% of it.

Conclusion. ANCHOR turns the ball-in-a-bowl picture into a port-Hamiltonian plant with 52 readable parameters and guarantees that hold by construction. For design, a system should act on the team’s trajectory and spend cues on conversation balance, where breakdowns are short and recurrent, while attention calls for sustained support. The next step is to close the loop with real cues in a deployment study.

Acknowledgments

This work was supported by the National Science Foundation under grant number 2339266.

References

Appendices

The appendices give supporting detail. Section A lists the notation, section B describes data preprocessing and section C gives the complete model. Section D contains full proofs, section E treats the case in which the resting point is recomputed, section F reports additional results on significance, accuracy, the fitted model, early warning and simulated cues.

Notation

Table 7 lists every symbol of the main text.

SymbolMeaning
kk, hhwindow at which a forecast starts; steps ahead (16 s each)
qkq_k, q^k+h\hat{q}_{k+h}team state (20 numbers); its forecast
q⋆q^\starresting point (equilibrium, or attractor)
qˉrec\bar{q}_{\mathrm{rec}}, qˉses\bar{q}_{\mathrm{ses}}, qˉpop\bar{q}_{\mathrm{pop}}recent, session and population anchors
wrecw_{\mathrm{rec}}, wsesw_{\mathrm{ses}}, wpopw_{\mathrm{pop}}learned anchor weights
z=q−q⋆z=q-q^\stardeviation from the resting point
HH, KKenergy; stiffness (steepness of the bowl)
κ\kappa, γℓ\gamma_\ell, LℓL_\elllog-steepness, one per behaviour and one per team coordinate; gain and Laplacian of link ℓ\ell
χ\chi, LconsL_{\mathrm{cons}}gain and matrix tying the team coordinates to the member and pair coordinates
1n\mathbf{1}_nvector of nn ones
RσR_\sigma, JσJ_\sigmaregulation (removes energy); exchange (moves energy)
ζσ\zeta_\sigmalog-recovery rate, one per behaviour and one per team coordinate
jσ,SAj_{\sigma,SA}, jσ,APj_{\sigma,AP}; SSAS_{SA}, SAPS_{AP}exchange gains; their skew-symmetric link matrices
σ\sigmatask phase (exploration or consolidation)
MM, D(t)D(t)inertia; sensed damping (tested variants, not selected)
πk\pi_k, uhu_hremaining work; forecast change in it
g=Kbg=Kbport gain, with one learned value of bb per behaviour and one per team coordinate
AσA_\sigma, CσC_\sigma, GσG_\sigma, uˉh\bar{u}_h(Jσ−Rσ)K(J_\sigma-R_\sigma)K; one-step maps of the forecast; input at the step midpoint
θ\theta, ν\nu, htrh_{\mathrm{tr}}all parameters; inverse training variances in the loss; training horizon
dhd_h, δ\delta, τ\taucue; bound on its size; minimum cue spacing
∥z∥K\lVert z\rVert_K, ρ\rhoenergy norm 2H\sqrt{2H}; shrink factor per step
mhm_hshare of a deviation remaining after hh windows
β\betaassumed cue effect in simulation (training standard deviations)

Table 7. All symbols used in the main text. Each is also introduced where it first appears.

Data and Preprocessing

Windows. We use the windowed values of the MURMR dataset [29]: 32 s windows with a 16 s stride, consecutive within each session, indexed by k=0,1,…k=0,1,\ldots. Every coordinate is normalized to [0,1][0,1]. Figure 9 shows how consecutive windows overlap.

Windows showing overlapping consecutive windows

Figure 9. Windows are 32 s long and start every 16 s. Window k+1k+1 shares its first 16 s with window kk (grey), so a comparison one step ahead partly compares the same recording with itself. Window k+2k+2 starts where window kk ends, so 32 s is the first step free of overlap.

The speaking time of member ii is divided by 32 s, and since the conversation graph is broadcast, every outgoing edge of ii carries this value.

Attention and proximity. The attention of pair ijij is the total time both members’ gaze hovers simultaneously on the same image, ∑(min⁡(ei,ej)−max⁡(si,sj))\sum(\min(e_i,e_j)-\max(s_i,s_j)) over overlapping hover intervals on one image, ignoring overlaps below 0.013 s, divided by 32 s. Hovers are events on virtual objects and are not verified fixations. The proximity of pair ijij is the number of seconds in the window with distance at most 0.4572 on the logged XX–ZZ floor plane, divided by 33, the number of whole-second stamps in a window. Logged units are treated as meters. The split of the proximity anchor between session and population shifts with preprocessing choices more than the other weights, so we read it as a mix of the two.

Team coordinates. From each window we compute qt=(sˉ,dˉ,rˉ,η)q^t=(\bar{s},\bar{d},\bar{r},\eta): the mean speaking time sˉ\bar{s} over the four members, the mean joint-attention weight dˉ\bar{d} and mean proximity weight rˉ\bar{r} over the six pairs, and the speaking balance η=−∑iωilog⁡ωi/log⁡4\eta=-\sum_i\omega_i\log\omega_i/\log4 with speaking shares ωi=si/∑jsj\omega_i=s_i/\sum_j s_j (η=0\eta=0 when nobody speaks).

Anchors. With nrec=5n_{\mathrm{rec}}=5 windows and decay λrec=0.7\lambda_{\mathrm{rec}}=0.7, the recent anchor is the exponential average

qˉrec,k=∑j=0nrec−1λrecjqk−j∑j=0nrec−1λrecj;\bar{q}_{\mathrm{rec},k}=\frac{\sum_{j=0}^{n_{\mathrm{rec}}-1}\lambda_{\mathrm{rec}}^j q_{k-j}}{\sum_{j=0}^{n_{\mathrm{rec}}-1}\lambda_{\mathrm{rec}}^j};

the session anchor is qˉses,k=1k+1∑j≤kqj\bar{q}_{\mathrm{ses},k}=\frac{1}{k+1}\sum_{j\leq k}q_j; the population anchor qˉpop\bar{q}_{\mathrm{pop}} is the mean over all windows of the training teams.

Sensed damping (tested, not selected). Speaking dominance ϕi=∣sirec∣/∑jsjrec−14\phi_i=\lvert s_i^{\mathrm{rec}}\rvert/\sum_j s_j^{\mathrm{rec}}-\frac{1}{4} uses the recent anchor srecs^{\mathrm{rec}} of the speaking times. Overrides come from the object-interaction logs: an override i→ji\to j is recorded when member ii places an image whose last placement was by j≠ij\ne i with a different label. The count oijo_{ij} is the exponential average of these events over nrecn_{\mathrm{rec}} windows, and pair ijij receives the conflict oij+ojio_{ij}+o_{ji}.

Progress signals. We compared three task-progress signals. (i) The label-based error is the fraction of images whose current label differs from the group’s final label, which uses end-of-session information and is therefore not causal. (ii) Coverage is the fraction of images never placed. (iii) Remaining work, used throughout the paper, is the fraction of images that have not been placed with a label that then stayed unchanged for four windows (64 s). All three relate to shared attention in the same way, with within-team Spearman correlations between attention density and the signal of -0.44 (label-based), -0.46 (coverage) and -0.47 (remaining work), so the causal signal loses nothing. At a forecast origin kk we extrapolate the signal as π^k+h=clip⁡(πk−vh,0,1)\hat{\pi}_{k+h}=\operatorname{clip}(\pi_k-vh,0,1) with rate v=max⁡(vpop,exponential average of πj−1−πj)v=\max(v_{\mathrm{pop}},\text{exponential average of }\pi_{j-1}-\pi_j), where vpopv_{\mathrm{pop}} is the mean decrease per window in the training teams while π>0\pi>0. The recent level πˉk\bar{\pi}_k is the exponential average of π\pi over the last nrecn_{\mathrm{rec}} windows.

The 1 s interaction table. MURMR also releases per-second pairwise features, whose speaking indicators we used only for the timescale diagnostic of section F.3. Each member’s per-second speaking indicator is identical across the three pairs that contain the member. The per-second timelines end earlier than the windowed ones, so the two resolutions are not sample-aligned. Aggregated into 32 s windows and aligned at the best offset for each team, the indicator correlates with the windowed conversation weights at r=0.58r = 0.58–0.920.92 per team.

Complete Model Specification

The selected model is fully specified by eqs. (7) to (9), the port g=Kbg = Kb of § 5.3 and Table 3. Here we add what the main text leaves out. We write the coordinates in four blocks: S (speakers, 4), A (attention pairs, 6), P (proximity pairs, 6) and T (team coordinates, 4).

A circuit analogy. Table 8 shows the same port-Hamiltonian structure in an electrical circuit and in a team, for readers who know the former.

TermCircuitTeam
Energy HHCharge stored in capacitorsDistance from the usual pattern
Regulation RσR_\sigmaResistors, which dissipateSelf-regulation back to the pattern
Exchange JσJ_\sigmaLossless couplingTalk turning into shared focus
Port g ug\,uVoltage sourceTask progress shifting the pattern
Kick ddCurrent pulseA prompt to the team
Phase σ\sigmaSwitch positionExploration or consolidation

Table 8. The same structure in a circuit and in a team. The team column is a modelling hypothesis, not a physical law. The symbols are defined in sections 3.2 and 3.3 and listed in table 7 in section A.

Tie between levels. Lcons=∑B∈{S,A,P}1nBvBvBTL_{\mathrm{cons}} = \sum_{B\in\{S,A,P\}} \frac{1}{n_B}v_Bv_B^{\mathsf T} with vB=eTB−ecv_B = e_{T_B} - e_c, where ece_c is the unit vector of coordinate cc, TBT_B is the team coordinate of behaviour BB and nBn_B is its number of coordinates. Then 12χzTLconsz\frac{1}{2}\chi z^{\mathsf T}L_{\mathrm{cons}}z is the sum of the terms 12χ(t−m)2\frac{1}{2}\chi(t-m)^2 of § 5.2. The fitted value is χ=0.08\chi= 0.08.

Sensed damping (tested, not selected). With sensed damping, regulation becomes Rσ+D(t)R_\sigma+ D(t) with

D(t)=diag⁡(μdomϕ,μov,μov,0)D(t) = \operatorname{diag}(\mu_{\mathrm{dom}}\phi,\mu_{\mathrm{ov}},\mu_{\mathrm{ov}},0)

where ϕ∈R4\phi\in\mathbb{R}^{4} stacks the speaking dominances, o∈R6o \in\mathbb{R}^{6} the pair conflicts oij+ojio_{ij}+o_{ji} (section B), and μdom,μov≥0\mu_{\mathrm{dom}},\mu_{\mathrm{ov}} \ge0 are learned. The selected model has D=0D = 0.

Anchor weights. The weights of eq. (2) are softmaxes of free logits, one triple per behaviour and one per team coordinate.

Permutation-equivariant family. Any two coordinates of the member and pair block have one of 15 symmetric relation types—SS (same, different); AA, PP and AP (same pair, pairs sharing a member, disjoint pairs); SA and SP (member in pair, member not in pair)—and the 7 cross-type relations (SA, SP, AP) admit an orientation, giving 7 skew-symmetric matrices. The member and pair block of KK is B2+ϵIB^2+\epsilon I and R=F2R=F^2 (ϵ=10−3\epsilon= 10^{-3}), with B,FB,F linear combinations of the symmetric basis and JJ a linear combination of the skew basis. Same-type skew couplings (for example member ii to member jj) would require an orientation between interchangeable members and are therefore excluded by symmetry.

Unconstrained family. The member and pair block of KK is BBT+ϵIBB^{\mathsf T}+\epsilon I, R=FFTR=FF^{\mathsf T} with lower-triangular B,FB,F, and J=W−WTJ=W-W^{\mathsf T} with free WW.

Second-order variant (inertia). This variant adds a momentum pp with diagonal inertia MM, so the state is (z,p)(z,p) and the energy Hessian is blkdiag⁡(K,M−1)\operatorname{blkdiag}(K,M^{-1}), with J=[0I−IJσPP]J = \begin{bmatrix} 0 & I \\ -I & J_{\sigma}^{PP} \end{bmatrix}, R=blkdiag⁡(0,RσPP+D(t))R = \operatorname{blkdiag}(0,R_{\sigma}^{PP}+D(t)) and g=(0,Kb)g=(0,Kb), with MM diagonal per modality and team coordinate; the initial momentum is p0=MΔqˉp_0=M\Delta\bar{q}, where Δqˉ\Delta\bar{q} is the exponential average of recent window-to-window differences.

Observer (+obs). Over the last six windows, the observer alternates the frozen step with the update q^←q^−+ℓ∘mk∘(qk−q^−)\hat{q}\leftarrow\hat{q}^{-}+\ell\circ m_k\circ(q_k-\hat{q}^{-}), where the gain ℓ∈(0,1)\ell\in(0,1) is learned per modality and mkm_k marks the observed modalities.

Live loop. Figure 10 draws the loop of algorithm 1 as a block diagram.

Learning and selection. The loss is

L(θ)=120N∑origins k∑h=1htr∑c=120νc(q^k+h,c−qk+h,c)2+λreg∥θ∥2,\mathcal{L}(\theta)=\frac{1}{20N}\sum_{\mathrm{origins}\ k}\sum_{h=1}^{h_{\mathrm{tr}}}\sum_{c=1}^{20}\nu_c\left(\hat{q}_{k+h,c}-q_{k+h,c}\right)^2+\lambda_{\mathrm{reg}}\lVert\theta\rVert^2,

with θ\theta all parameters, NN the number of origin-step pairs with an observed target, νc\nu_c the inverse training variance of the coordinate’s modality (or team coordinate), λreg=10−6\lambda_{\mathrm{reg}}=10^{-6} (10−410^{-4} for the unconstrained family), minimized by L-BFGS-B (at most 300 iterations) with exact gradients. The configuration is selected by the sum of pooled member and pair- and team-level nMSE over 16–256 s. Table 9 lists all hyperparameters.

HyperparameterValue
Recent anchor nrecn_{\mathrm{rec}}, decay λrec\lambda_{\mathrm{rec}}5 windows, 0.7
Training horizon htrh_{\mathrm{tr}} / evaluation horizon8 / 16 windows
Phase threshold on π^\hat{\pi}0.5
Stability window for π\pi4 windows (64 s)
OptimizerL-BFGS-B, ≤300\leq300 it., exact gradients
Regularization λreg\lambda_{\mathrm{reg}}10−610^{-6} (10−410^{-4} unconstrained)
VAR(2) ridge / direct ridge1 / 10
MLP32 tanh units, weight decay 10−310^{-3}, 400 it.
Symmetry augmentation23 relabellings (MLP: 7)
Breakdown threshold ϑj\vartheta_j25th percentile of training teams
Alarm window / false-alarm rate4 windows / 20%
Simulation40 windows, 30 seeds, cue spacing τ=2\tau= 2

Table 9. Hyperparameters.

Baselines. VAR(2) is qk+1=W[qk;qk−1;1]q_{k+1}=W[q_k;q_{k-1};1], fitted by ridge regression and iterated. Direct ridge uses one linear map per horizon whose inputs are the last five windows (100 values), the session mean (20), the extrapolated progress inputs (16) and a constant, which gives 137 features and about 4×1044\times10^4 parameters. The MLP uses the same standardized inputs and one hidden layer. Augmentation adds the 23 non-identity relabellings of each training team, since relabelling the members permutes speakers and pairs and leaves team coordinates unchanged.

Proofs

We give proofs for the first-order model. They use only that the energy is quadratic with a positive definite Hessian, so they also hold for the second-order variant with KK replaced by blkdiag⁡(K,M−1)\operatorname{blkdiag}(K,M^{-1}), except that its regulation is singular and only ∥Cσ∥K≤1\|C_\sigma\|_K\leq1 is guaranteed. For a matrix XX, we write ∥X∥K=max⁡z≠0∥Xz∥K/∥z∥K=∥K1/2XK−1/2∥2\|X\|_K=\max_{z\ne0}\|Xz\|_K/\|z\|_K=\|K^{1/2}XK^{-1/2}\|_2.

Lemma 1 (Contraction). Let K≻0K\succ0, J=−JTJ=-J^{\mathsf T} and R⪰0R\succeq0, and let A=(J−R)KA=(J-R)K. Then I−12AI-\frac{1}{2}A is invertible and C=(I−12A)−1(I+12A)C=(I-\frac{1}{2}A)^{-1}(I+\frac{1}{2}A) satisfies ∥C∥K≤1\|C\|_K\leq1. If R≻0R\succ0, then ∥C∥K<1\|C\|_K<1.

Proof. With ξ=K1/2z\xi=K^{1/2}z, the map becomes C~=(I−12A~)−1(I+12A~)\widetilde C=(I-\frac{1}{2}\widetilde A)^{-1}(I+\frac{1}{2}\widetilde A) with A~=K1/2(J−R)K1/2=J~−R~\widetilde A=K^{1/2}(J-R)K^{1/2}=\widetilde J-\widetilde R, where J~\widetilde J is skew-symmetric and R~⪰0\widetilde R\succeq0. If (I−12A~)v=0(I-\frac{1}{2}\widetilde A)v=0 for v≠0v\ne0, then ∥v∥2=12vTA~v=−12vTR~v≤0\|v\|^2=\frac{1}{2}v^{\mathsf T}\widetilde A v=-\frac{1}{2}v^{\mathsf T}\widetilde Rv\leq0, a contradiction, so the matrix is invertible. For any vv, let ψ=(I−12A~)−1v\psi=(I-\frac{1}{2}\widetilde A)^{-1}v. Then

∥C~v∥2−∥v∥2=∥(I+12A~)ψ∥2−∥(I−12A~)ψ∥2=ψT(A~+A~T)ψ=−2ψTR~ψ≤0.\|\widetilde C v\|^2-\|v\|^2 =\|(I+\tfrac{1}{2}\widetilde A)\psi\|^2-\|(I-\tfrac{1}{2}\widetilde A)\psi\|^2 =\psi^{\mathsf T}(\widetilde A+\widetilde A^{\mathsf T})\psi =-2\psi^{\mathsf T}\widetilde R\psi\leq0.

with strict inequality for v≠0v\ne0 when R~≻0\widetilde R\succ0. By compactness of the unit sphere, ∥C~∥2<1\|\widetilde C\|_2<1 in that case. □\square

In the first-order model every diagonal entry of RσR_\sigma is an exponential and any sensed damping is non-negative, so Rσ≻0R_\sigma\succ0 and ρ=max⁡∥Cσ∥K<1\rho=\max\|C_\sigma\|_K<1. The reported value ρ=0.814\rho=0.814 is this maximum over all origins and phases of the fitted model.

Proof of proposition 1. With Hh=12zhTKzhH_h=\frac{1}{2}z_h^{\mathsf T}Kz_h and KK symmetric,

Hh+1−Hh=12(zh+1−zh)TK(zh+1+zh)=vT(zh+1−zh)H_{h+1}-H_h=\frac{1}{2}(z_{h+1}-z_h)^{\mathsf T}K(z_{h+1}+z_h)=v^{\mathsf T}(z_{h+1}-z_h)

with v=12K(zh+zh+1)v=\frac{1}{2}K(z_h+z_{h+1}), which is the energy gradient at the midpoint. The implicit midpoint rule gives zh+1−zh=(Jσ−Rσ)v+guˉhz_{h+1}-z_h=(J_\sigma-R_\sigma)v+g\bar u_h. Hence

Hh+1−Hh=vTJσv−vTRσv+vTguˉh.H_{h+1}-H_h=v^{\mathsf T}J_\sigma v-v^{\mathsf T}R_\sigma v+v^{\mathsf T}g\bar u_h.

The first term is zero because JσJ_\sigma is skew-symmetric, and the second is non-positive because Rσ⪰0R_\sigma\succeq0. The argument uses only the matrices of the current step, so it holds for every phase sequence. □\square

Proof of proposition 2. Set dh=0d_h=0 at steps without a cue, so that zh+1=Cσh(zh+dh)z_{h+1}=C_{\sigma_h}(z_h+d_h). Unrolling gives

zh=Φ(h,0)z0+∑hi<hΦ(h,hi)dhi,Φ(h,j)=Cσh−1⋯Cσj,z_h=\Phi(h,0)z_0+\sum_{h_i<h}\Phi(h,h_i)d_{h_i},\qquad\Phi(h,j)=C_{\sigma_{h-1}}\cdots C_{\sigma_j},

and ∥Φ(h,j)∥K≤ρh−j\|\Phi(h,j)\|_K\leq\rho^{h-j}. Order the cues before hh from the most recent, r=0,1,…r=0,1,\ldots. Because they are at least τ\tau steps apart, the rr-th satisfies h−hi≥rτh-h_i\geq r\tau. Therefore ∥zh∥K≤ρh∥z0∥K+δ∑r≥0ρrτ=ρh∥z0∥K+δ/(1−ρτ)\|z_h\|_K\leq\rho^h\|z_0\|_K+\delta\sum_{r\geq0}\rho^{r\tau}=\rho^h\|z_0\|_K+\delta/(1-\rho^\tau). □\square

Lemma 2 (Shift of the resting point). For a fixed phase and a constant input uu, the first-order model has the unique equilibrium z∗=Tuz^*=Tu with T=−Aσ−1gT=-A_\sigma^{-1}g, which is also the fixed point of eq. (8). This Γ\Gamma is the “Shift” of table 15, averaged over origins and phases.

Proof. AσA_\sigma is invertible because K≻0K\succ0 and the symmetric part of Jσ−RσJ_\sigma-R_\sigma is negative definite. The continuous equilibrium solves Aσz∗+gu=0A_\sigma z^*+gu=0. The discrete fixed point solves (I−12Aσ)z∗=(I+12Aσ)z∗+gu(I-\frac{1}{2}A_\sigma)z^*=(I+\frac{1}{2}A_\sigma)z^*+gu, which is the same equation. □\square

Proposition 3 (Invariance to relabelling the members). Let PP be the 20×2020\times20 permutation induced by renaming the four members, which permutes speakers and pairs and leaves team coordinates fixed. Then PKPT=KPKP^{\mathsf T}=K, PJσPT=JσPJ_\sigma P^{\mathsf T}=J_\sigma and the baseline regulation commutes with PP, and sensed damping, when used, permutes with the data. If the population anchor is symmetric, Pqˉpop=qˉpopP\bar q_{\mathrm{pop}}=\bar q_{\mathrm{pop}}, then renaming the input data renames the forecast in the same way.

Proof. The per-behaviour diagonal terms are constant on each block. Each link matrix is defined by a relation between coordinates, such as “speaker ii belongs to pair ijij”, that renaming preserves, so these matrices commute with PP. Sensed damping, recent anchors and session anchors are computed per member and per pair, so they permute with the data. The attractor weights are shared within each kind of behaviour, so every quantity in the forecast is either invariant or permutes with PP, and the solution of a permuted linear system is the permuted solution. □\square

The reported results use the per-coordinate population mean, which is not exactly symmetric, for example because member labels are tied to headsets. Replacing it with its symmetric version changes held-out error by +0.1%+0.1\% at the member and pair level (p=0.963p=0.963) and by −0.0%-0.0\% at the team level (p=0.890p=0.890).

When the Resting Point Is Recomputed

Propositions 1 and 2 hold for one forecast. A live system recomputes qk∗q_k^* at every window, so the energies of successive forecasts are measured from different points and do not add up. We state what still holds for the first-order model. Real teams. All coordinates of qq lie in [0,1][0,1]. Let DD be the largest energy distance between two points of [0,1]20[0,1]^{20}, ρˉ=sup⁡∥Cσ∥K\bar{\rho}=\sup\|C_{\sigma}\|_{K} and γˉ=sup⁡∥Gσ∥K\bar{\gamma}=\sup\|G_{\sigma}\|_{K}. Let WrecW_{\mathrm{rec}} and WsesW_{\mathrm{ses}} be the diagonal matrices of anchor weights, and let S=∑j=04λrecjS=\sum_{j=0}^{4}\lambda_{\mathrm{rec}}^{j}.

Proposition 4 (Forecasts of a real team). For any observations in [0,1]20[0,1]^{20}, the resting point q⋆q^{\star} stays in [0,1]20[0,1]^{20}. For every window kk, phase sequence, input with ∣uˉh∣≤1|\bar{u}_{h}|\leq1 and cue schedule as in proposition 2,

∥q^k+h−qk⋆∥K≤ρˉhD+δ1−ρˉρˉτ+γˉ1−ρˉ.\lVert\hat{q}_{k+h}-q_{k}^{\star}\rVert_{K}\leq\bar{\rho}^{h}D+\frac{\delta}{1-\bar{\rho}}\bar{\rho}^{\tau}+\frac{\bar{\gamma}}{1-\bar{\rho}}.

Once five windows have been observed, the resting point moves by at most

∥qk+1⋆−qk⋆∥K≤D(∥Wrec∥KS+∥Wses∥Kk+2).\lVert q_{k+1}^{\star}-q_{k}^{\star}\rVert_{K}\leq D\left(\frac{\lVert W_{\mathrm{rec}}\rVert_{K}}{S}+\frac{\lVert W_{\mathrm{ses}}\rVert_{K}}{k+2}\right).

Proof. The anchors are averages of points in the box, and eq. (2) is a convex combination in each coordinate, so qk⋆q_{k}^{\star} is in the box and ∥z0∥K≤D\lVert z_{0}\rVert_{K}\leq D. Within a forecast qk⋆q_{k}^{\star} is fixed, so unrolling zh+1=Cσ(zh+dh)+Gσuˉhz_{h+1}=C_{\sigma}(z_{h}+d_{h})+G_{\sigma}\bar{u}_{h} as in proposition 2 gives the first bound. For the second, the session anchor changes by (qk+1−qˉses,k)/(k+2)(q_{k+1}-\bar{q}_{\mathrm{ses},k})/(k+2), whose norm is at most D/(k+2)D/(k+2). The recent anchor changes by (qk+1−v)/S(q_{k+1}-v)/S, where v=∑j=03(λrecj−λrecj+1)qk−j+λrec4qk−4v=\sum_{j=0}^{3}(\lambda_{\mathrm{rec}}^{j}-\lambda_{\mathrm{rec}}^{j+1})q_{k-j}+\lambda_{\mathrm{rec}}^{4}q_{k-4} is a convex combination of observations, so its norm is at most D/SD/S. The population anchor is constant. □

For the fitted model, D=4.77D=4.77 and γˉ/(1−ρˉ)=1.01\bar{\gamma}/(1-\bar{\rho})=1.01. These bounds need no assumption about how teams respond to cues, and they are conservative, because they allow the resting point to travel anywhere in the box.

The model feeding on itself. To ask how far repeated cues could move the resting point, let the model’s own states build the anchors. Write ehe_{h} for the state minus qˉpop\bar{q}_{\mathrm{pop}}, let Wpop=I−Wrec−WsesW_{\mathrm{pop}}=I-W_{\mathrm{rec}}-W_{\mathrm{ses}}, and take no task input:

eh+1=ah+Cσ(eh+dh−ah),ah=Wreceˉrec,h+Wseseˉses,h,(15)e_{h+1}=a_{h}+C_{\sigma}(e_{h}+d_{h}-a_{h}),\qquad a_{h}=W_{\mathrm{rec}}\bar{e}_{\mathrm{rec},h}+W_{\mathrm{ses}}\bar{e}_{\mathrm{ses},h}, \tag*{(15)}

where eˉrec,h\bar{e}_{\mathrm{rec},h} and eˉses,h\bar{e}_{\mathrm{ses},h} are the anchors of ee. Only the population anchor pulls toward a fixed point, since the other two follow the state, so that the bowl moves with the ball.

Proposition 5 (Cues in the self-fed loop). Fix the phase sequence and let e0=0e_{0}=0. (i) eq. (13) is linear in the history, so eT=∑hΨ(T,h)dhe_{T}=\sum_{h}\Psi(T,h)d_{h}, and for every schedule with ∥dh∥K≤δ\lVert d_{h}\rVert_{K}\leq\delta and spacing at least τ\tau, ∥eT∥K≤δBτ(T)\lVert e_{T}\rVert_{K}\leq\delta B_{\tau}(T) with Bτ(T)=max⁡schedules∑h∥Ψ(T,h)∥KB_{\tau}(T)=\max_{\mathrm{schedules}}\sum_{h}\lVert\Psi(T,h)\rVert_{K}. (ii) If Wpop≻0W_{\mathrm{pop}}\succ0 and a constant cue dˉ\bar{d} is applied in a fixed phase, the only constant solution is e∗=Wpop−1(I−Cσ)−1Cσdˉe^{\ast}=W_{\mathrm{pop}}^{-1}(I-C_{\sigma})^{-1}C_{\sigma}\bar{d}. (iii) If Wpop=0W_{\mathrm{pop}}=0, every constant vector solves eq. (13) without cues, so e=0e=0 is not asymptotically stable.

Proof. (i) The anchors are linear in past states with coefficients that depend only on the step, so superposition gives the sum and the triangle inequality gives the bound. Dynamic programming computes the maximum over schedules exactly. (ii) At a constant solution both anchors equal e∗e^{\ast}, so e∗−a=Wpope∗e^{\ast}-a=W_{\mathrm{pop}}e^{\ast}, and eq. (13) reads (I−Cσ)Wpope∗=Cσdˉ(I-C_{\sigma})W_{\mathrm{pop}}e^{\ast}=C_{\sigma}\bar{d}, where I−CσI-C_{\sigma} is invertible by lemma 1. (iii) If the whole history equals cc, then ah=ca_{h}=c and eh+1=ce_{h+1}=c. □

A simple small-gain argument for a time-uniform bound would require ∥Cσ∥K+∥(I−C)Wrec∥K+∥(I−C)Wses∥K<1\lVert C_{\sigma}\rVert_{K}+\lVert(I-C)W_{\mathrm{rec}}\rVert_{K}+\lVert(I-C)W_{\mathrm{ses}}\rVert_{K}<1, and for the model fitted on all teams this sum is 1.61. We therefore compute Bτ(T)B_{\tau}(T) directly, in both phases, and for every time of one phase switch. With one cue per 32 s, B2(T)B_{2}(T) is 5.2 over 8 min, 6.8 over 16 min and 8.2 over 32 min, against 3.0 for a fixed resting point. With one cue per 16 s it is 12.9 over 16 min, against 5.4. A simulation in exploration with cues of one kind of behaviour at a time gives 2.0 cues’ worth for speaking, 2.5 for proximity and 5.9 for shared attention after 16 min.

The fitted population weight explains the difference. It is 0.37 for speaking and 0.50 for proximity, but only 0.012 for shared attention, which is therefore close to case (iii). The constant solution in (ii) would lie about 138 cues’ worth away in the worst direction at one cue per 32 s, but the approach is extremely slow, because the session anchor averages all past windows: the simulated deviation is 12.6 after 2000 windows, about nine hours. Simulation confirms the approach: with weights (0.3, 0.3, 0.4) the simulated state is within 1% of it after 4000 windows.

Proposition 6 (Time-uniform bound). In the self-fed loop of eq. (13) with ∥dh∥K≤δ\lVert d_{h}\rVert_{K}\leq\delta, let κ=sup⁡σ∥Cσ∥K+∥(I−Cσ)Wrec∥K+∥(I−Cσ)Wses∥K\kappa=\sup_{\sigma}\lVert C_{\sigma}\rVert_{K}+\lVert(I-C_{\sigma})W_{\mathrm{rec}}\rVert_{K}+\lVert(I-C_{\sigma})W_{\mathrm{ses}}\rVert_{K}. If κ<1\kappa<1, then ∥eh∥K≤ρδ/(1−κ)\lVert e_{h}\rVert_{K}\leq\rho\delta/(1-\kappa) for all hh.

Proof. From eq. (13), eh+1=Cσ(eh+dh)+(I−Cσ)(Wreceˉrec,h+Wseseˉses,h)e_{h+1}=C_{\sigma}(e_{h}+d_{h})+(I-C_{\sigma})(W_{\mathrm{rec}}\bar{e}_{\mathrm{rec},h}+W_{\mathrm{ses}}\bar{e}_{\mathrm{ses},h}). The anchors are convex combinations of past states, so if all past states have norm at most MM, then ∥eh+1∥≤κM+ρδ\lVert e_{h+1}\rVert\leq\kappa M+\rho\delta. Taking M=ρδ/(1−κ)M=\rho\delta/(1-\kappa) closes the induction from e0=0e_{0}=0. □

Over the leave-one-team-out fits kk reaches 1.76, so the condition calls for a larger population weight. A large population weight in effect fixes the bowl in place and trades some accuracy for it. Blending the anchor weights toward the population with share λ\lambda, without refitting, gives κ=0.96\kappa=0.96 at λ=0.8\lambda=0.8 and κ=0.89\kappa=0.89 at λ=0.9\lambda=0.9, with bounds of 20.6 and 7.7 cues’ worth for all time. Member and pair nMSE then rises from 1.082 to 1.201 and 1.230, and team nMSE from 1.002 to 1.115 and 1.143. Refitting the other parameters under the floor is a natural way to reduce this cost.

Proposition 5 describes the model feeding on itself, which is how a planner sees the loop, and the computed bound covers one phase switch, as in 15 of the 17 sessions.

Additional Results

Significance and multiple tests

Every comparison of two models uses the error of each on the same 17 held-out teams. A paired Wilcoxon signed-rank test gives a pp-value, the probability of a difference at least this consistent if the two models were equally good, and we call a result significant when p<0.05p<0.05. Each test then has a 5% chance of a false positive, so a table of many tests is likely to contain one by chance; with 17 independent tests the chance of at least one is about 58%. We therefore also apply the Holm correction [17], which sorts the mm pp-values and compares the smallest with 0.05/m0.05/m, the next with 0.05/(m−1)0.05/(m-1), and so on, stopping at the first that fails. This keeps the chance of any false positive among all mm tests below 5%. In table 4 (m=17m=17), the results for a resting point from recent behaviour only,

VariantParamsMember and pair 16–48 sMember and pair 16–256 sTeam 16–256 s
Selected model520.7581.0821.002
1st-order, diag450.7621.0881.011
1st-order, diag, switched540.7601.0831.002
1st-order, equiv690.7691.0931.013
1st-order, equiv, switched950.7711.0860.998
1st-order, full4240.8501.1681.075
1st-order, full, switched6840.8521.2391.079
2nd-order, diag520.7591.0861.009
2nd-order, diag, switched610.7611.0881.016
2nd-order, equiv760.7631.0881.011
2nd-order, equiv, switched1020.7681.0921.015
2nd-order, full4310.8221.1281.033
2nd-order, full, switched6910.8641.1751.062
attractor: recent only400.8871.2931.152
attractor: session only400.7681.0991.063
no task-progress port470.7711.1231.057
no friction features520.7581.0821.002
no scale coupling530.7591.0821.001
attractor only, no dynamics210.8481.0991.009

Table 11. All model families and ablations (nMSE on held-out teams, lower is better). Params: free parameters, counted as in table 3. The selected model is within 1% of several neighbours. The rows from “attractor: recent only” on remove one part from the switched topology-tied model with sensed damping (54 parameters), as in table 4; diag, equiv and full denote the topology-tied, looser-tie and free-matrix families, friction is the sensed damping, and scale coupling is the tie χ\chi between levels.

the dynamics, speaking balance and the shared structure pass, and a resting point from session history only is significant before the correction (p=0.035p = 0.035 and 0.020). For the horizon comparison of Figure 11 (m=16m = 16), the vertical lines mark how far ANCHOR stays better after the correction.

Figure 11. How far ahead ANCHOR is useful. Skill is the error reduction relative to assuming an average training team (higher is better; solid: member and pair level, dashed: team level). Vertical lines mark the last horizon up to which ANCHOR beats that reference with p<0.05p < 0.05 at every horizon (solid: pairs, dashed: team, dotted: both levels with Holm correction for the 16 tests).

Forecast accuracy

This part gives the detail behind section 6.4: per-cell comparisons, absolute errors, errors by horizon, team and quantity, robustness to sensor corruption, data efficiency and the anchored baselines.

How far ahead. Figure 11 compares ANCHOR with assuming an average training team, by horizon. Long-range errors above 1, as in Figure 7, still carry skill. The error is scaled by the variation of the training teams, and a held-out team’s later windows differ from that variation: at 128 s, assuming an average team gives 1.32 against 1.10 for ANCHOR.

Sensor corruption. We corrupt input windows, never targets, with Gaussian noise or with dropout of each modality of a window with probability pp. When models are trained with the same corruption, ANCHOR or its observer has lower error than the learned baselines tested (Figure 12 and Table 10), and the attractor-only control is within about 0.01 of it, slightly lower at the two strongest noise levels; augmented ridge was not part of this test. With noise of standard deviation 0.2 its nMSE rises from 1.01 to 1.14, against 1.38 for VAR. Without retraining, dropout remains benign, but strong noise hurts ANCHOR far more than VAR (2.97 against 1.29), because its resting point averages the corrupted windows, so noise must be matched in training.

CorruptionShift +obsShift ANCHORShift VARShift MLPMatched +obsMatched ANCHORMatched VARMatched MLP
Noise σ=0.05\sigma= 0.051.141.131.131.211.051.051.311.18
Noise σ=0.1\sigma= 0.11.471.471.171.291.111.111.381.20
Noise σ=0.2\sigma= 0.22.992.971.291.461.141.141.381.23
Dropout p=0.1p = 0.11.011.011.121.161.011.011.121.16
Dropout p=0.3p = 0.31.031.031.131.171.041.031.131.17
Dropout p=0.5p = 0.51.041.041.141.171.061.041.151.18

Table 10. Robustness (member and pair nMSE, 16–256 s; lower is better). Shift: trained on clean data, tested on corrupted inputs. Matched: trained and tested on corrupted inputs. +obs: ANCHOR with a model-based observer. Clean data: ANCHOR 1.01, VAR 1.12, MLP 1.16.

Figure 12. Member and pair error under additive sensor noise (left) and per-modality window dropout (right) when training and test inputs are corrupted alike.

Per team and per quantity. Figure 13 compares ANCHOR with the strongest baselines team by team, and Table 12 breaks the team-level error down by quantity; attention density is the hardest quantity for every model.

ModelSpeaking levelAttention densityProximity densitySpeaking balance
ANCHOR (ours)0.861.191.130.82
Attractor only0.881.201.140.82
Persistence1.221.611.581.25
Recent average1.071.381.351.02
VAR(2) + aug.0.971.611.250.86
MLP + aug.1.111.611.261.04

Table 12. Team-level error by quantity (nMSE, 16–256 s; lower is better). Attention density is the hardest quantity for every model.

Figure 13. Per-team error of ANCHOR against the strongest baselines (one point per held-out team and baseline). Points in the shaded area below the diagonal are teams for which ANCHOR is better.

ModelMember and pair 16–48 sMember and pair 16–256 sTeam 16–48 sTeam 16–256 s
Attractor only+11.3% (17/17, 2e-5)+1.5% (12/17, 0.057)+8.5% (15/17, 5e-4)+0.1% (9/17, 0.747)
Persistence+38.8% (17/17, 2e-5)+49.2% (17/17, 2e-5)+32.1% (16/17, 3e-5)+43.5% (17/17, 2e-5)
Recent average+23.1% (17/17, 2e-5)+22.6% (16/17, 3e-5)+16.2% (16/17, 3e-4)+22.1% (14/17, 0.001)
Session mean+28.0% (17/17, 2e-5)+13.1% (12/17, 0.006)+46.4% (16/17, 5e-5)+36.4% (14/17, 0.002)
VAR(2)+15.3% (16/17, 3e-5)+10.4% (12/17, 0.080)+16.6% (14/17, 5e-4)+11.2% (11/17, 0.071)
VAR(2) + aug.+6.4% (12/17, 0.017)+8.3% (12/17, 0.120)+7.6% (12/17, 0.031)+9.0% (9/17, 0.284)
Direct ridge+28.0% (17/17, 2e-5)+15.3% (17/17, 2e-5)+23.2% (15/17, 4e-4)+14.2% (13/17, 0.015)
Ridge + aug.-1.1% (8/17, 0.431)+1.5% (10/17, 0.306)-1.1% (8/17, 0.678)+1.4% (11/17, 0.487)
MLP+39.3% (17/17, 2e-5)+14.7% (13/17, 0.004)+44.9% (17/17, 2e-5)+25.9% (12/17, 0.031)
MLP + aug.+33.9% (17/17, 2e-5)+10.0% (12/17, 0.051)+32.3% (16/17, 2e-4)+15.2% (10/17, 0.353)
GRU+38.7% (17/17, 2e-5)+15.7% (15/17, 1e-4)+36.8% (15/17, 2e-4)+15.7% (12/17, 0.120)
GRU + aug.+28.9% (17/17, 2e-5)+7.3% (13/17, 0.017)+26.1% (15/17, 1e-4)+8.5% (10/17, 0.487)
Anchored scalar+1.7% (14/17, 0.006)+0.2% (9/17, 0.487)+0.0% (9/17, 0.611)+0.4% (10/17, 0.818)
Anchored free+1.8% (12/17, 0.011)+0.7% (10/17, 0.404)+4.4% (13/17, 0.013)+1.9% (10/17, 0.378)
Anchored free + aug.+1.5% (11/17, 0.020)+0.2% (10/17, 0.517)+0.1% (9/17, 0.782)+0.8% (9/17, 0.611)

Table 13. Forecast error of each model relative to ANCHOR on held-out teams, in detail: change in error averaged over teams (%; positive: higher error than ANCHOR), number of teams for which ANCHOR is better, and paired Wilcoxon pp. Table 5 summarizes these values.

ModelParamsMember and pair 16–48 sMember and pair 16–256 sTeam 16–48 sTeam 16–256 sTop pair
ANCHOR (ours)520.7581.0820.7051.00230%
ANCHOR + observer560.7591.0820.7071.00430%
Attractor only210.8481.0990.7751.00930%
Persistence01.0491.6270.9131.41631%
Recent average00.9271.3300.8071.20631%
Session mean00.9711.2291.0301.35730%
VAR(2)8200.8991.2580.8501.18418%
VAR(2) + aug.8200.8261.2420.7801.17015%
Direct ridge43,8400.9931.2810.9021.19516%
Ridge + aug.43,8400.7541.1100.7031.04427%
MLP14,9441.0841.2911.0731.34017%
MLP + aug.14,9441.0571.2581.0041.25318%
Anchored scalar280.7711.0890.7021.014–
Anchored free4340.7731.0960.7351.030–
Anchored free + aug.4340.7701.0890.7061.021–
GRU27,1681.0661.2830.9921.213–
GRU + aug.27,1680.9941.2020.9261.158–

Table 14. Forecast error on held-out teams (nMSE; lower is better) over the first three windows (16–48 s) and over all 16 (16–256 s). Top pair: how often the most engaged pair is identified 128 s ahead. Shaded: ANCHOR; bold: best in each column. Params: number of values fitted from the data. The anchored models use ANCHOR’s fitted anchors, which favours them. These errors pool all forecasts before averaging, whereas tables 5 and 13 average per team first, as in table 4; ratios of the values here therefore differ slightly from the percentages there.

Top pair. ANCHOR predicts how much a pair will interact better than which pair will lead, where persistence is level or slightly stronger (31% against 30%).

How much data the model needs. Figure 14 trains every model on 2–16 teams. On mean error ANCHOR stays below augmented ridge at every size, significantly up to eight training teams at member and pair level and up to four at team level; with more teams the two are on par, and at 12–16 teams the median member and pair error favours ridge. ANCHOR reaches the attractor-only control from about eight teams at member and pair level and about 12 at team level, without a significant gain there. The variant with free matrices needs the most data.

What the structure adds. All anchored baselines take ANCHOR’s fitted anchors, so they share its resting point. A scalar relaxation fits one decay and one port gain per behaviour and per team coordinate by least squares. A free linear map fits zh+1=Azh+buhz_{h+1}=A z_h+b u_h with ridge regression. The GRU has 32 hidden units, reads the last five windows, and adds the session mean and the task input to a linear output layer. Over 170 retrainings on 2, 4, 8 and 16 teams, ρ\rho reached at most 0.998 for ANCHOR by construction, and the free map had a spectral radius of at most 0.92 empirically. At small sizes the scalar relaxation had the lowest member and pair error, which shows that the fewest parameters win when data are scarcest.

The fitted model and real teams

This part gives the fitted values behind section 6.3 and further checks of how real teams move: raw energy, the energy balance on one forecast, the timing of the phase, richer switching and the scale of seconds.

Raw energy on real teams. The fading of fig. 4 appears across many windows, and noise hides it in any one of them. After the 122 windows with the highest energy, the observed energy measured from the same resting point is 0.65 times its starting value 48 s later, against 0.70 for shuffled sessions and 0.06 for the forecast, and at most lags real and shuffled sessions have similar energy.

Coordinatewrecw_{\mathrm{rec}}wsesw_{\mathrm{ses}}wpopw_{\mathrm{pop}}ShiftCons.Expl.
Speaking (member)0.440.190.37-0.0582637
Attention (pair)0.180.810.01-0.0731728
Proximity (pair)0.100.410.50-0.0662855
Team speaking level0.340.130.53-0.0423832
Team attention density0.350.580.07-0.0602243
Team proximity density0.190.460.35-0.1111578
Team speaking balance0.430.020.56-0.2541222

Table 15. Fitted model on all teams. Anchor weights; shift of the resting point per unit of remaining work (negative: it rises as images are decided); and recovery time constants in seconds in consolidation and exploration.

Energy is quadratic, so window-to-window noise always adds to it, whereas the forecast is the noise-free expected path. The signed projection of eq. (14), which we chose after seeing this result, is the measure that matches an expected path. Noise in the starting window still lowers even mhm_h by the same factor, because it inflates the denominator of eq. (14) but cancels in the numerator, so the levels of mhm_h may be slightly understated while the ratio between successive steps, about 30% removed per 16 s, is unaffected. Energy is also distinct from task progress (rank correlation -0.08).

The energy balance on one forecast. Figure 15 follows one 24-window forecast for a held-out team to show what proposition 1 says. The energy is measured from the resting point at the forecast origin. Without task input, the same forecast loses energy at every step and is back at rest within a few windows, which is the half of the proposition that makes the model safe to iterate: a forecast cannot drift away from rest on its own. With task input, the energy first falls in the same way and then rises and levels off, because the forecast change in remaining work supplies energy through the port and moves the forecast toward a shifted resting point (lemma 2). The dashed bound, the initial energy plus the energy supplied by the task, is loose here, because it leaves out the energy that regulation removes; what it guarantees is that every rise in energy comes from the task.

When the phase changes. Remaining work crosses the phase threshold once in 15 of 17 sessions, and for the first time at a median of 304 s (interquartile range 288–320 s), in sessions of 5.6–19.7 min. Of all forecasts, 11% meet a switch within 48 s and 26% within 256 s. The forecast phase matches the phase later observed at 82% of horizons. The phase therefore works much like an early and late part of the session, and the slower recovery in exploration also reflects time in session. The task signal also partly reaches the forecast through the anchors: the median within-team correlation with remaining work is −0.44 for the attention density of a window and −0.83 for its resting point, and a session average also trends with time. In a scalar relaxation, the task term has about one eighth of the magnitude of the deviation for speaking, one fifth for proximity and balance, and one half for attention. These findings are consistent with phase and port being on par in accuracy.

Richer switching and nonlinear steps. With ANCHOR’s anchors, we compared the scalar relaxation with 22 switched variants, each scored at four settings. They switched decay, and optionally shifted the resting point, by task phase, by behavioural clusters or by MURMR’s interaction states [29], or they changed decay with the sign or size of the deviation. All the switched variants remain pH systems. After Holm correction, the scalar relaxation matched all of them, and the largest gain, 2.9% (task phase with a shifted resting point, team level, 16–256 s), was below significance. A generous bound confirms this: when three hidden states of team behaviour are inferred from the whole held-out session and known at every future step, with the resting point shifting by state, error falls by 3.6–6.0% across levels and horizons, although this bound sees the future. MURMR’s states, read with its released encoder, are stable for most pairs within a session, because several encoder inputs are pair-level summaries, so here they describe the pair more than a moment in time. A free nonlinear correction with no stability constraint grew without bound (error +51% at 16–48 s), which illustrates the value of built-in stability.

Is there exchange or inertia at the scale of seconds? The model finds no evidence of lossless routing or inertia at 16 s, but such structure might still exist at finer scales. To test this, we analysed the per-second interaction table at bins of 1–8 s (Table 16) and compared real teams with surrogate teams assembled from unrelated members. For every signal and bin size, real teams show no more time-irreversibility, rotation or oscillation than the surrogates, and proximity and attention are, if anything, smoother and more reversible. The dissipative picture therefore holds down to one second for these two signals. Conversation could not be tested conclusively. In the per-second speaking signal, simultaneous speech occurs as often as if members spoke independently (median ratio of observed to independent overlap 1.08), whereas genuine turn-taking would make overlap rarer than chance. Cross-talk between co-located headsets and one-second bins that span speaker changes are likely causes, so testing conversational exchange requires segment-level diarization free of cross-talk.

SignalBin (s)Irreversibility TeamIrreversibility Surr.Irreversibility ppRotational share TeamRotational share Surr.Rotational share ppMin. autocorrelation TeamMin. autocorrelation Surr.
Speaking10.2790.2930.7150.0100.0140.782-0.063-0.081
Speaking20.2870.3090.7670.0340.0340.326-0.066-0.092
Speaking40.2700.2880.7270.0930.0870.220-0.069-0.101
Speaking80.2980.2950.6240.1850.1850.473-0.052-0.084
Proximity10.1270.1350.4800.0210.0160.063+0.030-0.038
Proximity20.1340.1460.5640.0360.0350.379+0.033-0.037
Proximity40.1310.1400.5660.0610.0720.875+0.039-0.039
Proximity80.1520.1700.6810.1330.1560.872+0.067-0.005
Attention10.4380.4840.8700.0880.0880.627-0.065-0.081
Attention20.4180.4680.9060.1020.1420.956-0.077-0.099
Attention40.3850.4230.8860.2440.2240.264-0.074-0.106
Attention80.3740.4330.9770.3270.3380.443-0.050-0.094

Table 16. Second-scale diagnostic on the 1 s interaction table: time-irreversibility of the lagged cross-covariance (signature of lossless exchange JJ), share of rotational modes of a fitted linear model (signature of JJ or inertia), and the most negative autocorrelation within 60 s (signature of inertia), for real teams and for surrogate teams assembled from members or pairs of different teams. The proximity signal here is the continuous pair distance. pp: one-sided Mann–Whitney test that real teams show more of the signature.

Early warning

Early warning, in full. Warning times in table 6 largely reflect the 64 s alarm window, since every model warns about as early.

Table 17: Early-warning AUROC (0.5 = chance, 1 = perfect) for a breakdown at a single future window, by horizon: all origins / onsets only (breakdown absent at the origin).

Model32 s all32 s onset64 s all64 s onset96 s all96 s onset
Conversation collapse
ANCHOR (ours)0.800.830.780.800.760.78
Attractor only0.810.830.780.790.760.78
Persistence0.760.730.750.720.720.69
Recent average0.800.820.780.780.760.78
VAR(2) + aug.0.760.740.780.780.790.80
Ridge + aug.0.780.770.750.740.750.75
MLP + aug.0.750.740.740.740.740.74
Loss of shared focus
ANCHOR (ours)0.810.750.740.700.670.62
Attractor only0.800.740.740.700.680.61
Persistence0.820.730.740.700.730.66
Recent average0.810.740.760.720.720.64
VAR(2) + aug.0.720.670.630.590.590.55
Ridge + aug.0.730.680.720.680.680.59
MLP + aug.0.750.690.710.650.660.58
Low proximity
ANCHOR (ours)0.710.670.610.630.580.61
Attractor only0.710.640.610.600.570.58
Persistence0.740.680.620.650.570.59
Recent average0.710.640.610.600.570.59
VAR(2) + aug.0.690.660.540.550.440.42
Ridge + aug.0.720.660.680.700.650.66
MLP + aug.0.670.600.600.580.590.58

Table 17.

Early warning by horizon. Table 17 complements table 6 for a single future window. For conversation collapse the leading models are close, and the augmented VAR is strongest at 96 s. Augmented ridge leads for low proximity from 64 s on, and the augmented VAR is weakest for loss of shared focus.

Sensitivity to the breakdown threshold. Table 18 repeats early warning and the simulated interventions with breakdowns defined at the 15th and 35th training percentile. At every threshold, ANCHOR stays within 0.02 of the best model for conversation collapse and loss of shared focus and ahead of the learned baselines, and augmented ridge leads for low proximity. Collapse AUROC is 0.85 at the main threshold and about 0.75 at the other two. The advantage of ANCHOR-predictive over periodic cueing shrinks as breakdowns become more frequent, because a fixed schedule then hits a breakdown more often by chance.

Simulated cues

Simulation protocol. Each simulated episode starts from the first six windows of a held-out team and runs for 40 windows (about 11 min). At every window the plant, either ANCHOR fitted on all teams or an independently learned VAR(2) on relabel-augmented data, predicts the next window, and an error drawn at random from the plant’s one-step errors on real data is added, so the simulated team varies as real teams do. A cue adds β∈{0.25,0.5,1}\beta\in\{0.25, 0.5, 1\} training standard deviations to its targets in the window it is issued: speaking balance for balance cues, and attention density with all six pair attentions for attention cues. The forecast and ANCHOR-predictive policies use ANCHOR fitted without the simulated team, and their trigger levels are set so that, on the training teams, they would fire about as often as periodic cueing; in simulation their rates differ

from periodic cueing (table 19), which is why we report the effect per cue. Each team is simulated with 30 random seeds, and every policy sees the same random draws. We report the share of the episode in breakdown and the breakdown time removed per cue relative to no cueing.

Interventions for every effect size. Table 20 gives the full simulation results behind Figure 8. Larger cue effects raise every policy’s efficiency. ANCHOR-predictive is the most efficient policy in 8 of the 12 settings, and reactive cueing in the other four: attention cues on the VAR plant at every effect size and balance cues on the VAR plant at β=1\beta= 1.

CuePlantβ\betaNonePeriodicReactiveForecastANCHOR-predictive
BalanceANCHOR0.2513.312.4 / 0.1812.3 / 0.2412.0 / 0.2312.8 / 0.34
BalanceANCHOR0.5013.311.4 / 0.3811.4 / 0.4710.8 / 0.4712.1 / 0.62
BalanceANCHOR1.0013.39.9 / 0.689.8 / 0.948.9 / 0.939.8 / 1.06
AttentionANCHOR0.2511.58.2 / 0.657.5 / 1.377.2 / 1.627.2 / 1.64
AttentionANCHOR0.5011.55.7 / 1.155.6 / 2.415.3 / 2.985.6 / 3.22
AttentionANCHOR1.0011.53.1 / 1.684.0 / 3.813.9 / 5.084.2 / 5.28
BalanceVAR0.2512.611.8 / 0.1611.4 / 0.3011.1 / 0.2712.1 / 0.33
BalanceVAR0.5012.611.0 / 0.3110.4 / 0.5810.0 / 0.5011.5 / 0.60
BalanceVAR1.0012.69.7 / 0.568.9 / 1.068.4 / 0.939.2 / 1.04
AttentionVAR0.2512.59.4 / 0.628.5 / 1.189.7 / 1.169.8 / 1.15
AttentionVAR0.5012.57.6 / 0.977.0 / 1.819.0 / 1.739.7 / 1.69
AttentionVAR1.0012.55.2 / 1.455.8 / 2.498.6 / 2.419.6 / 2.26

Table 20. Simulated interventions for every assumed cue effect β\beta (training s.d.): share of the episode in the breakdown (%, lower is better) / breakdown time removed per cue (percentage points, higher is better). ANCHOR-predictive is the most efficient policy in 8 of the 12 settings, and reactive cueing in the others.

References

  1. [1]Alessandro Abate, Daniele Ahmed, Alec Edwards, Mirco Giacobbe, and Andrea Peruffo. 2021. FOSSIL: A Software Tool for the Formal Synthesis of Lyapunov Functions and Barrier Certificates Using Neural Networks. In Proceedings of the 24th International Conference on Hybrid Systems: Computation and Control (HSCC). ACM, Nashville, TN, USA, 1–11. doi:10.1145/3447928.3456646DOI
  2. [2]Panos J. Antsaklis, Bill Goodwine, Vijay Gupta, Michael J. McCourt, Yue Wang, Po Wu, Meng Xia, Han Yu, and Feng Zhu. 2013. Control of Cyberphysical Systems Using Passivity and Dissipativity Based Methods. European Journal of Control 19, 5 (2013), 379–388.DOI
  3. [3]Holly Arrow, Joseph E. McGrath, and Jennifer L. Berdahl. 2000. Small Groups as Complex Systems: Formation, Coordination, Development. and Adaptation. Sage.
  4. [4]Robert F. Bales. 1953. The Equilibrium Problem in Small Groups. In Working Papers in the Theory of Action, Talcott Parsons, Robert F. Bales, and Edward A. Shils (Eds.). Free Press, Glencoe, IL, 111–161.
  5. [5]Klaus Bengler, Wenner Damm, Andreas Luedtke, Reiger Jochem, Benedikt Austel, Bianca Biebl, Martin Fränzle, Willem Hagemann, Moritz Held, David Hess, Klas Ihme, Severin Kacianka, Alyssa J. Kerscher, Laine Forrest, Sebastian Lehnhof, Alexander Pretschner, Astrid Rakow, Daniel Sonntag, Janos Sztipanovits, Maike Schwammberger, Mark Schweda, Anirudh Unni, and Eric Veith. 2024. A References Architecture for Human Cyber Physical Systems, Part II: Fundamental Design Principles for Human-CPS Interaction. ACM Transactions on Cyber-Physical Systems 8, 1 (2024), 1–27. doi:10.1145/3622880
  6. [6]Yasra Chandio, Diana Romero, Salma Elmalaki, and Fatima M. Anwar. 2026. What Sensors See, What People Feel: An Exploratory Study of Subjective Collaboration Perception in Mixed Reality. In Proceedings of the IEEE Conference on Virtual Reality and 3D User Interfaces (IEEE VR). 162–172. doi:10.1109/VR67842.2026.00039DOI
  7. [7]Morris H. DeGroot. 1974. Reaching a Consensus. J. Amer. Statist. Assoc. 69, 345 (1974), 118–121.DOI
  8. [8]Shaan Desai, Marios Mattheakis, David Sondak, Pavlos Protopapas, and Stephen Roberts. 2021. Port-Hamiltonian Neural Networks for Learning Explicit Time-Dependent Dynamical Systems. Physical Review E 104, 3 (2021), 043412. doi:10.1103/PhysRevE.104.043412DOI
  9. [9]Salma Elmalaki, Huey-Ru Tsai, and Mani Srivastava. 2018. Sentio: Driver-in-the-Loop Reward Collision Warning Using Multisample Reinforcement Learning. In Proceedings of the 16th ACM Conference on Embedded Networked Sensor Systems (SenSys). ACM, Shenzhen, China, 28–40. doi:10.1145/3274783.3274843DOI
  10. [10]Daniel C. Feldman. 1984. The Development and Enforcement of Group Norms. Academy of Management Review 9, 1 (1984), 47–53.DOI
  11. [11]Noah E. Friedkin and Eugene C. Johnsen. 1990. Social Influence and Opinions. Journal of Mathematical Sociology 15, 3–4 (1990), 193–206.DOI
  12. [12]Connie J. G. Gersick. 1988. Time and Transition in Work Teams: Toward a New Model of Group Development. Academy of Management Journal 31, 1 (1988), 9–41.DOI
  13. [13]John M. Gottman, James D. Murray, Catherine C. Swanson, Rebecca Tyson, and Kristin R. Swanson. 2002. The Mathematics of Marriage: Dynamic Nonlinear Models. MIT Press.
  14. [14]Samuel Greydanus, Misko Dzamba, and Jason Yosinski. 2019. Hamiltonian Neural Networks. In Advances in Neural Information Processing Systems (NeurIPS), Vol. 32.arxiv.org/abs/1906.01563
  15. [15]J. Richard Hackman and Ruth Wageman. 2005. A Theory of Team Coaching. Academy of Management Review 30, 2 (2005), 269–287.DOI
  16. [16]Ernst Hairer, Christian Lubich, and Gerhard Wanner. 2006. Geometric Numerical Integration (2 ed.). Springer.
  17. [17]Steve W. J. K. Żolowski. 1979. A Simple Sequentially Rejective Multiple Test Procedure. Scandinavian Journal of Statistics 6, 2 (1979), 65–70.DOI
  18. [18]Steve W. J. Kozlowski, Georgia T. Chao, James A. Grand, Michael T. Braun, and Goran Kuljanin. 2013. Advancing Multilevel Research Design: Capturing the Dynamics of Emergence. Organizational Research Methods 16, 4 (2013), 581–615.
  19. [19]Steve W. J. Kozlowski and Daniel R. Ilgen. 2006. Enhancing the Effectiveness of Work Groups and Teams. Psychological Science in the Public Interest 7, 3 (2006), 77–124.
  20. [20]Steve W. J. Kozlowski and Katherine J. Klein. 2000. A Multilevel Approach to Theory and Research in Organizations: Contextual, Temporal, and Emergent Processes. In Multilevel Theory, Research, and Methods in Organizations, Katherine J. Klein and Steve W. J. Kozlowski (Eds.). Jossey-Bass, 3–90.
  21. [21]Daniel Liberzon. 2003. Switching in Systems and Control. Birkhäuser.DOI
  22. [22]Chengwen Luo and Mun Choon Chan. 2013. SocialWear: Collaborative Inference of Human Conversation Networks Using Smartphones. In Proceedings of the 11th ACM Conference on Embedded Networked Sensor Systems (SenSys). ACM, Rome, Italy, 1–14. doi:10.1145/2517352.2517353DOI
  23. [23]Gaurav Manek and J. Zico Kolter. 2019. Learning Stable Deep Dynamics Models. In Advances in Neural Information Processing Systems, Vol. 32. 11128–11136.arxiv.org/abs/2001.06116
  24. [24]Michelle A. Marks, John E. Mathieu, and Stephen J. Zaccaro. 2001. A Temporally Based Framework and Taxonomy of Team Processes. Academy of Management Review 26, 3 (2001), 356–376.DOI
  25. [25]Romeo Ortega, Arjan J. van der Schaft, Iven Mareels, and Bernhard Maschke. 2001. Putting Energy Back in Control. IEEE Control Systems Magazine 21, 2 (2001), 18–33.DOI
  26. [26]Alex Pentland. 2012. The New Science of Building Great Teams. Harvard Business Review 90, 4 (2012), 60–69.
  27. [27]Anton V. Proskurnikov and Roberto Tempo. 2017. A Tutorial on Modeling and Analysis of Dynamic Social Networks. Part I. Annual Reviews in Control 43 (2017), 65–79.DOI
  28. [28]Diana Romero, Fatima M. Anwar, and Salma Elmalaki. 2025. MoCoMR: A Collaborative MR Simulator with Individual Behavior Modeling. In Proceedings of the 3rd International Workshop on Human-Centered Sensing, Modeling, and Intelligent Systems (HumanSys). ACM, 114–119. doi:10.1145/3722570.3726894DOI
  29. [29]Diana Romero, Yasra Chandio, Fatima M. Anwar, and Salma Elmalaki. 2026. MURMR: A Multimodal Sensing Pipeline for Automated Group Behavior Analysis in Mixed Reality. IEEE Transactions on Visualization and Computer Graphics (ISMAR) (2026), 1–11. doi:10.1109/TVCG.2026.3747536DOI
  30. [30]Diana Romero, Xin Gao, Daniel Khalil, and Salma Elmalaki. 2026. TeamLLM: Exploring the Capabilities of LLMs for Multimodal Group Interaction Prediction. In 2026 IEEE 3rd International Workshop on Foundation Models for Cyber-Physical Systems and Internet of Things (FMSys). 49–54. doi:10.1109/FMSys72110.2026.00017DOI
  31. [31]Gunar Schirner, Duner Erdogmus, Kaushik Chowdhury, and Taskin Padir. 2013. The Future of Human-in-the-Loop Cyber-Physical Systems. Computer 46, 1 (2013), 36–45. doi:10.1109/MC.2012.428DOI
  32. [32]Shi Sheng, Erfan Pakrmanian, Kyunghae Han, Ziran Wang, John Lenneman, David Parker, and Lu Feng. 2022. Planning for Automated Vehicles with Human Trust. ACM Transactions on Cyber-Physical Systems 6, 4 (2022), 1–21. doi:10.1145/3561059
  33. [33]Robin R. Vallacher and Andrzej Nowak. 1997. The Emergence of Dynamical Social Psychology. Psychological Inquiry 8, 2 (1997), 73–99.DOI
  34. [34]Arjan J. van der Schaft and Dimitri Jeltsema. 2014. Port-Hamiltonian Systems Theory: An Introductory Overview. Foundations and Trends in Systems and Control 1, 2–3 (2014), 173–378.
  35. [35]Anita Williams Woolley, Christopher F. Chabris, Alex Pentland, Nada Hashmi, and Thomas W. Malone. 2010. Evidence for a Collective Intelligence Factor in the Performance of Human Groups. Science 330, 6004 (2010), 686–688.
  36. [36]Yanxia Zhang, Jeffrey Olenick, Chu-Hsiang Chang, Steve W. J. Kozlowski, and Hayley Hung. 2018. TeamSense: Assessing Personal Affect and Group Cohesion in Small Teams through Dyadic Interaction and Behavior Analysis with Wearable Sensors. Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies 2, 3 (2018), 1–22. doi:10.1145/3264960DOI
  37. [37]Tiany u Zhao and Salma Elmalaki. 2024. FinA: Fairness of Adverse Effects in Decision-Making of Human-Cyber-Physical-System. In Proceedings of the ACM/IEEE 15th International Conference on Cyber-Physical Systems (ICCPS). IEEE, Hong Kong, 202–211. doi:10.1109/ICCPS61052.2024.00025
  38. [38]Tianyu Zhao, Mahmoud Swerra, and Salma Elmalaki. 2026. FinP: Fairness-in-Privacy in Federated Learning by Addressing Disparities in Privacy Risk. Proceedings on Privacy Enhancing Technologies 2026, 4 (2026), 587–607. doi:10.56553/popets-2026-0136DOI
  39. [39]Tianyu Zhao, Mojtaba Taherisadr, and Salma Elmalaki. 2024. FAIRO: Fairness-aware Sequential Decision Making for Human-in-the-Loop CPS. In Proceedings of the ACM/IEEE 15th International Conference on Cyber-Physical Systems (ICCPS). IEEE, Hong Kong, 87–98. doi:10.1109/ICCPS61052.2024.00015DOI
  40. [40]Yaofeng Desmond Zhong, Biswavid Dey, and Amit Chakraborty. 2020. Dissipative SymODEN: Encoding Hamiltonian Dynamics with Dissipation and Control into Deep Learning. In ICLR Workshop on Integration of Deep Neural Models and Differential Equations.arxiv.org/abs/2002.08860
  41. [41]Feng Zhu, Han Yu, Michael J. McCourt, and Panos J. Antsaklis. 2012. Passivity and Stability of Switched Systems under Quantization. In Proceedings of the 15th ACM International Conference on Hybrid Systems: Computation and Control (HSCC). ACM, 237–244. doi:10.1145/2185632.2185668DOI

Paper details

vlnc:W_9s4au45e
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Engineering

Not recorded for this paper

View statement
Looking for related papers…
Contents