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.
Figures & tables
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.
Block
Size
Each entry
Speaking
4
Member’s speaking time
Attention
6
Pair’s simultaneous gaze on one image
Proximity
6
Pair’s time within 0.46 m
Team
4
Mean of each behavior; speaking balance
Table 1. The 20 numbers of the state qk for one window. Each entry is a share of the window, so it lies between 0 and 1.
Question
Part of the model
What it means for a team
Sensor source
Where does it rest?
Resting point q⋆ (P1)
The pattern it returns to, anchored on its past and other teams
Sociogram history
State across levels q (P3)
Team states emerge from members and pairs but are not reducible to them
Audio, gaze, head pose
How does it return?
Energy H , regulation Rσ (P2)
Distance from the pattern, and how fast deviations are regulated back
—
Link gains in K , Rσ , Jσ
Who interacts with whom constrains coordination
Team structure
What moves it away?
Task input gu (P5)
Task progress shifts the pattern
Task logs
Phase σ (P4)
Exploration and consolidation regulate differently
Task logs
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.
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.
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.
Part
Symbols (values)
Count
Stiffness K
κS,κA,κP (3), γ1..γ5 (5), κT (4), χ (1)
13
Regulation Rσ
ζσ,S,ζσ,A,ζσ,P,ζσ,T
(3+4)×2=14
Exchange Jσ
jσ,SA,jσ,AP
2×2=4
Port g=Kb
b
3+4=7
Anchors
wrec,wses,wpop
(3+4)×2=14
Total
52
Table 3. Where the 52 free parameters come from, with the symbols of eqs. 5 , 6 and 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.
Figure 4. Real deviations fade, on average close to the forecast. Share of a deviation that remains h windows later, mh , 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 appendix B ), so comparisons start at 32 s. Right: the same observed quantity for each behavior separately, computed from that behavior’s coordinates alone (with K replaced by the identity).
Part
Hypothesis
Alternative tested
Member and pair nMSE
Team nMSE
Verdict
Resting point q⋆ (P1)
Teams return to a point set by recent behaviour, their own history and other teams
q⋆ from recent behaviour only
+19.3% (17/17, 2e-5)
+16.5% (14/17, 0.004)
supported
q⋆ from session history only
+3.4% (14/17, 0.035)
+9.7% (14/17, 0.020)
supported
Regulation Rσ (P2)
Deviations fade, so modelling the return helps at 16–48 s
attractor only, no dynamics
+11.3% (17/17, 2e-5)
+8.5% (15/17, 5e-4)
supported
Team state q (P3)
Team states are not reducible to members and pairs
computed from member forecast
–
+270.3% (16/17, 3e-5)
supported
Shared link gains in K,Rσ,Jσ
Coordinates with the same role can share one value
unconstrained matrices
+14.4% (16/17, 4e-4)
+9.7% (17/17, 2e-5)
supported
Phase σ (P4)
Teams regulate differently in exploration and consolidation
single phase
+0.1% (9/17, 0.782)
-0.4% (7/17, 0.611)
on par
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 p . Errors are over 16–256 s, except for P2 (16–48 s, where dynamics act) and P3 (speaking balance). Shaded: supported ( p<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.
Figure 5. The bowl moves. Speaking balance of held-out team 8 (the longest session), with the resting point q⋆ recomputed every window as the blend of its three anchors (fitted weights in parentheses). Bottom: this coordinate’s bowl, 21Kii(x−q⋆)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.
Member and pair
Team
Model
Params
16–48 s
16–256 s
16–48 s
16–256 s
Verdict
Attractor only
21
+11.3
+1.5
+8.5
+0.1
better
Persistence
0
+38.8
+49.2
+32.1
+43.5
better
Recent average
0
+23.1
+22.6
+16.2
+22.1
better
Session mean
0
+28.0
+13.1
+46.4
+36.4
better
VAR(2)
820
+15.3
+10.4
+16.6
+11.2
better
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<0.05 ). 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, p -values and absolute errors are in tables 13 and 14 in section F.2 .
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 ).
Conversation collapse
Loss of shared focus
Low proximity
Model
AUC
Det.
Warn
AUC
Det.
Warn
AUC
Det.
Warn
ANCHOR (ours)
0.85
74%
31
0.76
57%
35
0.66
34%
29
Attractor only
0.85
73%
31
0.76
56%
35
0.62
28%
31
Recent average
0.83
68%
32
0.76
58%
37
0.63
26%
32
VAR(2) + aug.
0.80
64%
30
0.69
43%
36
0.65
39%
28
Ridge + aug.
0.80
60%
31
0.71
42%
35
0.68
43%
32
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 8. Simulated cues under an assumed effect β . 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.
Appendix figures & tables22 assets
Supplementary material from the paper’s appendix.
Appendix
Symbol
Meaning
k , h
window at which a forecast starts; steps ahead (16 s each)
qk , q^k+h
team state (20 numbers); its forecast
q⋆
resting point (equilibrium, or attractor)
qˉrec,qˉses,qˉpop
recent, session and population anchors
wrec,wses,wpop
learned anchor weights
z=q−q⋆
deviation from the resting point
Appendix
Table 7. All symbols used in the main text. Each is also introduced where it first appears.
Figure 9. Windows are 32 s long and start every 16 s. Window k+1 shares its first 16 s with window k (grey), so a comparison one step ahead partly compares the same recording with itself. Window k+2 starts where window k ends, so 32 s is the first step free of overlap.
Term
Circuit
Team
Energy H
Charge stored in capacitors
Distance from the usual pattern
Regulation Rσ
Resistors, which dissipate
Self-regulation back to the pattern
Exchange Jσ
Lossless coupling
Talk turning into shared focus
Port gu
Voltage source
Task progress shifting the pattern
Kick d
Current pulse
A prompt to the team
Phase σ
Switch position
Exploration or consolidation
Appendix
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 appendix A .
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.
Recent anchor nrec , decay λrec
5 windows, 0.7
Training horizon htr / evaluation horizon
8 / 16 windows
Phase threshold on π^
0.5
Stability window for π
4 windows (64 s)
Optimizer
L-BFGS-B, ≤ 300 it., exact gradients
Regularization λreg
10−6 ( 10−4 unconstrained)
VAR(2) ridge / direct ridge
1 / 10
Appendix
Table 9. Hyperparameters.
Shift
Matched
Corruption
+obs
ANCHOR
VAR
MLP
+obs
ANCHOR
VAR
MLP
Noise σ=0.05
1.14
1.13
1.13
1.21
1.05
1.05
1.31
1.18
Noise σ=0.1
1.47
1.47
1.17
1.29
1.11
1.11
1.38
1.20
Noise σ=0.2
2.99
2.97
1.29
1.46
1.14
1.14
1.38
1.23
Dropout p=0.1
1.01
1.01
1.12
1.16
1.01
1.01
1.12
1.16
Dropout p=0.3
1.03
1.03
1.13
1.17
1.04
1.03
1.13
1.17
Appendix
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.
Member and pair
Team
Variant
Params
16–48 s
16–256 s
16–256 s
Selected model
52
0.758
1.082
1.002
1st-order, diag
45
0.762
1.088
1.011
1st-order, diag, switched
54
0.760
1.083
1.002
1st-order, equiv
69
0.769
1.093
1.013
1st-order, equiv, switched
95
0.771
1.086
0.998
Appendix
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 between levels.
Model
Speaking
Attention
Proximity
Speaking
level
density
density
balance
ANCHOR (ours)
0.86
1.19
1.13
0.82
Attractor only
0.88
1.20
1.14
0.82
Persistence
1.22
1.61
1.58
1.25
Recent average
1.07
1.38
1.35
1.02
VAR(2) + aug.
0.97
1.61
1.25
0.86
Appendix
Table 12. Team-level error by quantity (nMSE, 16–256 s; lower is better). Attention density is the hardest quantity for every model.
Member and pair
Team
Model
16–48 s
16–256 s
16–48 s
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)
Appendix
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 p . Table 5 summarizes these values.
Member and pair
Team
Top
Model
Params
16–48 s
16–256 s
16–48 s
16–256 s
pair
ANCHOR (ours)
52
0.758
1.082
0.705
1.002
30%
ANCHOR + observer
56
0.759
1.082
0.707
1.004
30%
Attractor only
21
0.848
1.099
0.775
1.009
30%
Persistence
0
1.049
1.627
0.913
1.416
31%
Recent average
0
0.927
1.330
0.807
1.206
31%
Appendix
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.
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<0.05 at every horizon (solid: pairs, dashed: team, dotted: both levels with Holm correction for the 16 tests).
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.
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.
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 appendix C .
Coordinate
wrec
wses
wpop
Shift
Cons.
Expl.
Speaking (member)
0.44
0.19
0.37
-0.058
26
37
Attention (pair)
0.18
0.81
0.01
-0.073
17
28
Proximity (pair)
0.10
0.41
0.50
-0.066
28
55
Team speaking level
0.34
0.13
0.53
-0.042
38
32
Team attention density
0.35
0.58
0.07
-0.060
22
43
Team proximity density
0.19
0.46
0.35
-0.111
15
78
Appendix
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.
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 Hk , 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 4.1 , 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.
Irreversibility
Rotational share
Min. autocorrelation
Signal
Bin (s)
Team
Surr.
p
Team
Surr.
p
Team
Surr.
Speaking
1
0.279
0.293
0.715
0.010
0.014
0.782
-0.063
-0.081
Speaking
2
0.287
0.309
0.767
0.034
0.034
0.326
-0.066
-0.092
Speaking
4
0.270
0.288
0.727
0.093
0.087
0.220
-0.069
-0.101
Speaking
8
0.298
0.295
0.624
0.185
0.185
0.473
-0.052
-0.084
Proximity
1
0.127
0.135
0.480
0.021
0.016
0.063
+0.030
-0.038
Appendix
Table 16. Second-scale diagnostic on the 1 s interaction table: time-irreversibility of the lagged cross-covariance (signature of lossless exchange J ), share of rotational modes of a fitted linear model (signature of J 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. p : one-sided Mann–Whitney test that real teams show more of the signature.
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.
32 s
64 s
96 s
Model
all
onset
all
onset
all
onset
Conversation collapse
ANCHOR (ours)
0.80
0.83
0.78
0.80
0.76
0.78
Attractor only
0.81
0.83
0.78
0.79
0.76
0.78
Persistence
0.76
0.73
0.75
0.72
0.72
0.69
Recent average
0.80
0.82
0.78
0.78
0.76
0.78
Appendix
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).
Threshold
Collapse
Focus
Low prox.
Predictive best
vs. periodic
15th
0.75 / 0.76
0.77 / 0.77
0.65 / 0.69
5/12
3.9 ×
25th (main)
0.85 / 0.85
0.76 / 0.76
0.66 / 0.68
8/12
2.2 ×
35th
0.75 / 0.77
0.77 / 0.79
0.66 / 0.69
8/12
1.4 ×
Appendix
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 s.d. The remaining settings are won by reactive cueing, and at the 35th percentile also by the forecast policy.
Cue
Plant
None
Periodic
Reactive
Forecast
ANCHOR-predictive
Balance
ANCHOR
13.3
11.4 (4.7)
11.4 (3.8)
10.8 (5.1)
12.1 (1.8)
Attention
ANCHOR
11.5
5.7 (4.7)
5.6 (2.3)
5.3 (1.9)
5.6 (1.7)
Balance
VAR
12.6
11.0 (4.7)
10.4 (3.4)
10.0 (4.8)
11.5 (1.7)
Attention
VAR
12.5
7.6 (4.7)
7.0 (2.8)
9.0 (1.9)
9.7 (1.5)
Appendix
Table 19. Time in breakdown and cue budget at cue effect β=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.
Cue
Plant
β
None
Periodic
Reactive
Forecast
ANCHOR-predictive
Balance
ANCHOR
0.25
13.3
12.4 / 0.18
12.3 / 0.24
12.0 / 0.23
12.8 / 0.34
Balance
ANCHOR
0.50
13.3
11.4 / 0.38
11.4 / 0.47
10.8 / 0.47
12.1 / 0.62
Balance
ANCHOR
1.00
13.3
9.9 / 0.68
9.8 / 0.94
8.9 / 0.93
9.8 / 1.06
Attention
ANCHOR
0.25
11.5
8.2 / 0.65
7.5 / 1.37
7.2 / 1.62
7.2 / 1.64
Attention
ANCHOR
0.50
11.5
5.7 / 1.15
5.6 / 2.41
5.3 / 2.98
5.6 / 3.22
Attention
ANCHOR
1.00
11.5
3.1 / 1.68
4.0 / 3.81
3.9 / 5.08
4.2 / 5.28
Appendix
Table 20. Simulated interventions for every assumed cue effect β (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.
Whether pairing people with AI helps or hurts is usually reported as a single average effect. Using a real-money prediction market (Polymarket) as an objective, externally resolved benchmark, this pilot shows that the value of human-AI collaboration depends on a specific, measurable form of human capital. Analyzed at the level of the individual forecaster, hybrid performance is trimodal: most people either deferred to the model (matching it) or used it to rubber-stamp a prior guess (performing worse than the model alone), while a minority engaged in genuine complementary reasoning and reached accuracy matching or even exceeding (i.e., lower error than) the market itself. Collaborative traits (perspective-taking, intellectual humility, and curiosity) rather than raw cognitive ability or model benchmarks, distinguished who reached that mode. The results are preliminary but statistically robust, and motivate a pre-registered replication now in preparation.
Vivienne Ming
The Human Trust; Possibility Science; UCL Global Business School for Health
Recent advances in LLM agents have enabled complex cognitive capabilities, such as multi-step reasoning, planning, and tool use, that increasingly position these agents as collaborative partners in human workflows. Effective collaboration, however, requires collaborators to continuously maintain and align mental models of their own reasoning, partners' intentions, and shared goals during the collaborative process. Today's agents rarely develop such capabilities since they are primarily optimized for task completion rather than for maintaining the underlying mental models required for collaborative processes. Moreover, the community lacks authentic human collaboration data with action-level mental model annotations that could guide agents toward collaborative competence. To bridge this gap, we present ALMANAC, a dataset of Action-Level Mental model ANnotations for Agent Collaboration built from the Map Task, a classic dyadic routing task from social science. ALMANAC contains 2,987 collaboration actions, each paired with theory-informed mental model annotations that record the participants' self-reasoning, perceived partner intent, and perceived team goal. We benchmark six LLMs on predicting humans' next-turn behavior and mental models. Our results demonstrate ALMANAC's utility in evaluating models' ability to simulate human collaborative behaviors and infer their underlying mental models. The project page is available at https://human-agent-collab.hailab.io/almanac.
Jiaju Chen, Yuxuan Lu, Jiayi Su +10
Northeastern University · University of Notre Dame · Adobe +3
Supporting a person over repeated interactions requires interpreting their current state in context and anticipating its subsequent evolution. We formulate the Human-Oriented World Model (HOWM) as a conceptual framework centered on task-relevant human states and their evolution. HOWM connects current observations and individual history with environmental and social context to organize continuing state estimation, personalized dynamics prediction, and action support. Dynamic expectations and new observations jointly inform the current estimate, which in turn anchors personalized forecasts based on shared knowledge and individual experience. An external policy relates forecasts under alternative conditions to the person's goals, preferences, and constraints when considering support. Actual responses and outcomes provide evidence for revising the state estimate and personal dynamic patterns, thereby informing subsequent estimation, prediction, and support. We illustrate the framework through hypothetical training and learning scenarios. Our research agenda asks whether accumulated interaction experience improves subsequent estimation and prediction, and whether using the model helps the person pursue their goals over time.