Anticipating Social Consequences with Distributional World Models

1 Introduction

• Social world models describe how interactions may change after an agent acts, with other parties’ beliefs and goals helping determine their responses.Outcome measures such as revenue, goal completion, and cost make the consequences of different actions comparable.Source: §1 · p. 1
• Social interactions can yield several plausible futures from the same situation and action; predicting one response can leave alternatives implicit, while repeated simulations add inference-time computation.Prior examples include S3AP Foresee-and-Act, which predicts one next state per step, and GDP-Zero, which uses repeated simulations during search (Zhou et al., 2025).Source: §1 · pp. 1–2
• DSWM predicts response clusters, their probabilities, representative replies, and immediate and final outcome values for a visible interaction and candidate action.The actor uses these alternatives in one prediction per candidate action when choosing an action.Source: §1 · p. 2
• A 4B DSWM is trained on targets from repeated simulations using supervised fine-tuning and GRPO.The training procedure constructs targets from repeated interactions.Source: §1 · p. 2
• Across three seen and two unseen environments with unseen agent backbones, DSWM achieves the best mean on 13 of 15 metrics.Ablations examine probability assignments, response information, and diverse outcome supervision.Source: §1 · p. 2

2.1 Social World Models

• For each visible context, candidate action, and query, DSWM predicts response clusters with probabilities and conditional mean outcomes at immediate and final horizons.Final outcomes assume ordinary continuation after the response segment.Source: §2.1 · p. 2

2.3 Applying DSWM at Inference Time

• At each decision point, one prediction per candidate yields expected outcomes and response guidance; constraints and dominance filter candidates before the actor selects an eligible action.If all predictions are invalid, the actor selects without model guidance.Source: Figure 2 · p. 3, §2.3 · p. 4

2.1 Social World Models

• Probability-weighted cluster means estimate expected outcomes, but nonlinear objectives are approximate because within-cluster variation and covariance are omitted.The expected value averages conditional cluster values, including the residual bucket when present.Source: §2.1 · p. 3

2.2 Training DSWM

• From 50 same-action simulations, action-partitioned semantic clusters supply empirical probabilities and conditional means; SFT serializes targets, then GRPO evaluates generated distributions against raw continuations.SFT inputs use visible context, candidate action, and query, excluding reference continuations and observed labels.Source: §2.2 · pp. 3–4

3.1 Benchmarks

• Five benchmark adaptations cover negotiation, open-ended social interaction, information reasoning, mediation, and multi-agent coordination and privacy-aware sharing.ProMediate and SOTOPIA-ToM are unseen to the trained world model.Source: §3.1 · p. 5

3.2 Training Data; 3.3 Experimental Setup

• The training set contains 12,247 query rows from Magentic-NL, SOTOPIA, and AgentSense, while evaluation uses backbones disjoint from rollout collection.The world model is Qwen3-4B-Instruct-2507; evaluation pools Ministral 3 14B, MiniMax H3, and Claude Sonnet 5.Source: §3.2 · p. 5, §3.3 · p. 5

3.4 Main Results

• DSWM (SFT + GRPO) achieves the best average performance on six of seven held-out-environment metrics and seven of eight joint-transfer metrics, leading on 13 of 15 overall.The held-out evaluation uses backbones absent from training rollouts; joint transfer holds out both environments and backbones.Source: §3.4 · p. 6, Table 1 · p. 6, Table 2 · p. 7

3.5 Discussion

• Retaining only the most likely response lowers every primary metric across five environments, including Magentic-NL customer utility from 272.82 to 243.85 and SOTOPIA goal achievement from 8.21 to 7.87.The intervention changes numerical guidance and eligibility as well as removing alternative futures.Source: §3.5 · p. 7, Table 3 · p. 7
• Uniform probabilities lower all five primary metrics, and probability perturbations at pooled average JS divergence 0.28 reduce performance to 77–90% of original values.Reweighting changes how consequences contribute to expected utility and eligibility; calibration against observed frequencies remains a separate question.Source: Figure 4 · p. 7, §3.5 · p. 8
• Response text benefits all four environments tested with numerical guidance and filters fixed, while cluster-wise utility aggregation raises InfoMgmt from 0.2371 to 0.2462 as an eligible-set approximation.Magentic-NL remains unchanged under payoff-only filtering.Source: Figure 3 · p. 6, §3.5 · p. 8
• Training on 25 rather than 5 outcome measures raises consensus change from 0.0506 to 0.0628 and InfoMgmt from 0.1851 to 0.2508, with diminishing gains across intermediate subsets.Shared rollouts and matched supervision volume support attributing the gains to broader outcome supervision.Source: Figure 4 · p. 7, §3.5 · p. 8
• On SOTOPIA, DSWM prediction uses 16.2× fewer tokens and 4.8× lower latency than 30 sampled futures, including retries; direct expectation remains cheapest.With matched SFT budgets, SFT-only DSWM exceeds sampled futures on all 15 metrics; GRPO improves 13 metrics and reduces prediction error, but worsens CPV and leaves information accuracy unchanged.Source: Table 4 · p. 7, §3.5 · p. 8

Social world models and dialogue planning

• Prior social world models include single-next-state prediction, factorized physical and mental updates, and learned collective-belief transitions, whereas DSWM learns response probabilities and conditional outcomes from rollouts for a separate actor.Foresee-and-Act uses one next-state prediction; Fei and Zhao factor updates without training; Yu et al. learn collective-belief transitions.Source: §4 · p. 8
• Dialogue planners either simulate negotiations or search with prompted user simulators, while DSWM learns its response distribution from rollouts.The cited approaches include negotiation simulation and simulator-based search.Source: §4 · p. 8

Decision making with predicted outcomes

• Unlike probabilistic ensembles and distributional reinforcement learning, DSWM estimates conditional mean outcomes within action-conditioned response clusters expressed as generated text.The related-work comparison also distinguishes fixed-option population-answer prediction from DSWM's generated, action-conditioned alternatives.Source: §4 · p. 8

5 Conclusions

• DSWM achieved the best performance on 13 of 15 metrics across five environments and three unseen backbones, outperforming other world-modeling baselines.Ablations support retaining multiple futures and their probability weights; richer actor-facing information improves outcomes in four environments.Source: §5 · p. 8
• DSWM predicts response alternatives, probabilities, and outcome values to guide a separate actor.Source: §5 · p. 8

Limitations

• Transfer to people remains untested, and per-backbone analysis is needed to establish which models benefit.Training and evaluation use language-model interaction partners.Source: Limitations · p. 9
• Fifty continuations per action yield probabilities in 0.02 increments, can miss rare responses, and leave nonlinear utilities approximate because cluster means omit within-cluster variance.Labels also include judge noise, and final-horizon targets come from the ordinary actor policy rather than repeated model-guided selection.Source: Limitations · p. 9

Ethical Statements

• Predicted responses could support pressure or manipulation, so use with people should involve AI disclosure, consent, and objectives accounting for counterpart interests.Eligibility filters based on predicted expected values are not safety mechanisms; partners are simulated and human involvement was limited to one author annotating 400 transcripts.Source: Ethical Statements · p. 9

A Social World Model Interface

• The interface supplies setting, focal-agent profile, visible history, candidate action, outcome definitions, horizons, continuation policy, and output schema.Outputs encode response clusters with representative text, probabilities, and conditional mean outcomes; an optional residual bucket has no response text.Source: A Social World Model Interface · p. 11

B.1 Collecting Future States

• Collection restores the same complete state and samples 50 response segments per candidate action; only selected batches continue every branch to termination for final labels.Outcome values use deterministic calculations where available and otherwise LLM judges; legacy judge responses cannot be fully semantically re-checked.Source: §B.1 · p. 11

B.2 Constructing Distributional Targets

• Targets separate incompatible representative actions before clustering embeddings, retain the largest K clusters, and place the rest in a residual bucket; cluster summaries omit within-cluster variance and covariance.The prompt template lists requested horizons and outcome measures, but its rendered JSON line is abbreviated in the figure caption.Source: §B.2 · pp. 11–12, Figure 5 · p. 12

B.3 Query Views

• Query views vary requested measures, paraphrases, horizons, and cluster counts; complete-label checks require finite labels across every reference sample, with missing labels never imputed.Stored training rows are query rows, and different measure requests can reuse the same context, action, and samples.Source: §B.3 · p. 12, Table 5 · p. 13

B.4 Datasets

• Supplementary collection retained 179 candidate batches with 50 continuations each, totaling 8,950 continuation samples; recovered final-outcome views required no new rollout or judge call.The recovered views include 522 customer-utility and 501 service-revenue views in Magentic-NL, plus 186 AgentSense goal-completion views.Source: §B.4 · p. 13
• The 12,247-row training pool contains query rows that may reuse contexts, actions, and samples, while ProMediate and SOTOPIA-ToM contribute no training rows.The prediction-input inventory counts test queries and shared contexts separately; diagnostic comparisons retain 94% of cases on the same retained set for all methods.Source: C Training · p. 13, §B.4 · p. 13, Table 6 · p. 14

C Training

• SFT learns serialized response-cluster targets before GRPO optimizes prediction-quality reward, with both stages drawing on the training pool.SFT masks prompt and padding tokens, excludes rows longer than 8,192 tokens, and balances eligible rows across environments; GRPO makes one shuffled pass over its eligible pool.Source: C Training · pp. 13–14
• GRPO samples eight completions for each of four prompts per step and uses a clipped surrogate with DAPO token normalization; its loss masks padding and truncated completions.The ratio compares current and sampling-policy token probabilities, with epsilon set to 0.2; beta is zero, so there is no KL penalty or frozen reference model.Source: C Training · p. 14
• The reward combines distribution matching, response coverage and precision, outcome-value accuracy, and response diversity, with a parser gate for rejected predictions.Outcome error is range-normalized and averaged over requested measures and horizons; an immediate-only query has no final-horizon error term.Source: C Training · p. 14

D Benchmark Implementations

• The five benchmarks span negotiation, social coordination, family and workplace interactions, mediation, and privacy-sensitive information sharing.Magentic-NL covers negotiation; SOTOPIA and AgentSense cover social interactions; ProMediate tests mediation, and SOTOPIA-ToM tests information sharing and disclosure restrictions.Source: D Benchmark Implementations · p. 15
• The shared framework saves and restores complete environment states, but its implementation deviations make outcome values incomparable with official leaderboard numbers.Using SOTOPIA-ToM does not mean DSWM contains a theory-of-mind or belief model.Source: D Benchmark Implementations · p. 15
• Evaluation uses interaction backbones disjoint from training collection and evaluates every eligible focal role, except ProMediate, which evaluates only the mediator.The world-model backbone remains Qwen3-4B-Instruct-2507; five runs vary training and evaluation seeds.Source: D Benchmark Implementations · p. 15
• The online evaluation uses every test-split scenario and a shared grid across applicable methods, with three backbones and five seeds.GDP-Zero is evaluated only in Magentic-NL and SOTOPIA.Source: D Benchmark Implementations · p. 15, Table 9 · p. 17

E Applying DSWM at Inference Time

• At each focal decision, the actor samples up to four candidate actions, and DSWM makes a separate prediction request for each candidate.Requests predict up to four response clusters, or one for an action that ends the interaction.Source: E Applying DSWM at Inference Time · p. 16
• The policy computes probability-weighted expected outcomes, then applies constraint and dominance filters to determine candidate eligibility.A constrained candidate remains eligible only when its expected constraint cost is within 0.02 of the lowest cost in the current eligible set.Source: E Applying DSWM at Inference Time · p. 17
• The actor sees candidate actions, outcome values, the two most probable response descriptions and probabilities, and combined utilities before selecting an allowed action ID.The environment executes the selected candidate exactly; the actor does not rewrite it.Source: E Applying DSWM at Inference Time · p. 18
• Online selection is repeated at each later focal turn, although the final predictions model ordinary continuation rather than this repeated policy use.The supplementary collection does not remove this continuation mismatch.Source: E Applying DSWM at Inference Time · p. 18

• Methods within each evaluation panel share the focal actor, partner policies, scenarios, judges, and episode seeds.Trained prediction baselines also share candidate generation and selection procedures.Source: F Baselines and Variants · p. 18
• S3AP and GDP-Zero use their released implementations and are compared as complete systems under shared benchmark interfaces and evaluation backbones.GDP-Zero is used only in two-party settings; both external methods use the evaluation backbones rather than older models in released example commands.Source: F Baselines and Variants · p. 18
• S3AP predicts one next social state and refines a provisional action once, while GDP-Zero uses 20 MCTS simulations with three realizations per open-loop state.These are inference-time baselines; their models are not fine-tuned.Source: F Baselines and Variants · p. 18
• Trained prediction baselines match DSWM's base model, LoRA settings, rollout pool, SFT presentations, and optimizer steps, but omit GRPO.Comparison with DSWM after GRPO also reflects post-training; each trained baseline is retrained for each of five main-experiment seeds.Source: F Baselines and Variants · p. 18
• Prediction variants change one fixed final checkpoint, while the candidates-only baseline presents actions without prediction lines or a utility line.In the all-invalid fallback, the utility line is also omitted.Source: F Baselines and Variants · p. 19, Figure 6 · p. 19
• The prediction formats differ in what they learn: sampled futures generate 30 segments, single-future predicts one, and direct expectation predicts means without response text.DSWM averages utilities computed from cluster means, unlike the specified direct-expectation and sampled-futures aggregation in SOTOPIA-ToM.Source: F Baselines and Variants · p. 19
• Prediction-quality comparisons use identical saved states and candidate sets, but later online histories may diverge after different choices.Table 14 reports saved-state SOTOPIA prediction accuracy and inference cost, not total training cost or end-to-end episode latency.Source: F Baselines and Variants · p. 19

• Tables 1 and 2 report benchmark-native outcome values, and incompatible measures are not averaged across benchmarks.Tables 1–4 report means and standard deviations across five seeds after averaging three evaluation backbones within each seed.Source: G Metrics and Analysis Protocol · p. 19
• Every scheduled episode completed without failures or failure-based exclusions, and overlapping metrics mean counts of best means are not independent statistical tests.SOTOPIA overall includes goal, and InfoMgmt combines DA, IA, CPV, and EFF.Source: G Metrics and Analysis Protocol · p. 20
• Main evaluations report higher-is-better outcomes except CPV, with benchmark-specific scales and role-specific payoff subsets.For example, SOTOPIA goal ranges from 0 to 10, while AgentSense goal completion and actual information-question accuracy lie in [0, 1].Source: G Metrics and Analysis Protocol · p. 20
• ProMediate consensus change and topic efficiency range from −1 to 1, while SOTOPIA-ToM reports DA, IA, CPV, and EFF in [0, 1].InfoMgmt is computed per episode as (DA IA (1 −CPV) EFF)1/4 before averaging, so it cannot be recovered from the other columns' means.Source: G Metrics and Analysis Protocol · p. 20
• Prediction diagnostics cover normalized expected-outcome error, candidate-ranking agreement, and reference-estimated regret in three environments.On saved SOTOPIA states, validity is assessed before parser normalization or clipping, while accuracy includes repaired predictions on the common retained 94% of cases.Source: G Metrics and Analysis Protocol · p. 20
• Probability calibration is not evaluated; the authors say it would require independently defined observable events and fresh reference continuations.Agreement with the training matcher, low outcome-value error, and judge–human correlation do not establish calibrated response probabilities.Source: G Metrics and Analysis Protocol · p. 20

H Reproducibility Details

• The authors state that training data and implementation will be publicly released upon acceptance.The release is to preserve applicable licence notices, attribution, and descriptions of changes.Source: H Reproducibility Details · pp. 20–21
• The world model is Qwen3-4B-Instruct-2507, and clustering uses Qwen3-Embedding-0.6B; the exact TRL version is unrecorded.The workflow used eight A100 GPUs for approximately 80 hours, or 640 allocated GPU-hours, excluding server-side external API compute.Source: H Reproducibility Details · pp. 20–21
• The benchmark resources have differing licences, and repository licences do not independently establish rights to underlying film or television scripts.The planned release will identify such source material separately rather than apply a blanket licence to all benchmark-derived content.Source: H Reproducibility Details · p. 21
• Data-query construction uses seed 0; the main tables average five seeds, while the separate reliability audit uses three training seeds.Seed identifiers and their exact derivation are not tabulated, though manifests record data and model hashes, software, sampled rows, and validation queries.Source: H Reproducibility Details · p. 21

I Implementation Settings

• The implementation appendix points to the interface, data, and inference-time action-selection protocols for details beyond its settings record.ProMediate requires precomputed initial judgments.Source: I Implementation Settings · p. 21, H Reproducibility Details · p. 21

I.1 Collection Settings

• Collection samples 50 continuations per candidate using three interaction backbones, with final labels following an ordinary continuation policy.The supplementary collection continues every branch to termination; the original collection could instead supply immediate labels only for a batch.Source: §I.1 · p. 21

I.2 Clustering Settings

• Response texts are embedded and clustered by average-linkage agglomerative clustering on cosine similarity with a 0.75 threshold.The K largest clusters are retained explicitly; remaining clusters form a residual bucket.Source: §I.2 · p. 21

I.3 Model and Training Settings

• The setup adapts Qwen3-4B-Instruct-2507 (Yang et al., 2025) with rank-64 LoRA (α=128) on attention and MLP projections.Training uses two epochs and keeps the final checkpoint; SFT masks prompt and padding tokens.Source: §I.3 · p. 21
• Rows longer than 8,192 tokens are excluded, and smaller-environment rows are repeated and sampled to match Magentic-NL.Tables 1–4 average over five training/evaluation seeds.Source: §I.3 · p. 21
• GRPO samples eight prediction completions per prompt, matches predictions at cosine similarity ≥0.75 with representative-action agreement, and standardizes rewards within each group.It uses neither a KL penalty nor a reference model; rows with reference entropy above one nat are listed twice.Source: §I.3 · p. 21
• The actor uses predictions to filter and select actions, then executes the chosen action unchanged.Candidate IDs in Figure 7 are illustrative.Source: Figure 7 · p. 22

J Additional Analyses and Evaluation Protocols

• Combining first raises sampled-futures InfoMgmt 0.2114→0.2238 and CPV 0.3097→0.3121; DSWM InfoMgmt rises 0.2371→0.2462 while CPV falls 0.2674→0.2586.The table compares utility-aggregation orders in SOTOPIA-ToM; InfoMgmt is maximized and CPV minimized.Source: J Additional Analyses and Evaluation… · p. 22, Table 12 · p. 23
• Prediction-quality diagnostics compare methods on a shared retained subset of 94% of diagnostic cases, after repair, retries and exclusions.For SOTOPIA, each candidate action has 50 fresh reference continuations; reported latency includes prediction-request retries, not the full episode.Source: J Additional Analyses and Evaluation… · pp. 22–23
• Intervals exclude zero for Magentic-NL, SOTOPIA, AgentSense and SOTOPIA-ToM, but include zero for ProMediate in a separate 300-pair audit.All pairs completed, and the audit uses three training seeds per environment.Source: J Additional Analyses and Evaluation… · p. 23, Table 13 · p. 25

J.1 Mechanism Analyses

• Full DSWM improves over sampled futures on all five primary measures, with intervals above zero, including ProMediate (0.0059, [0.0014, 0.0132]).Against direct expectation, full DSWM exceeds it in four environments; the AgentSense difference of 0.020 remains inconclusive.Source: §J.1 · p. 24, Table 15 · p. 26
• Full guidance exceeds expected-values-only in all four environments outside Magentic-NL; its component benefits vary by environment and comparison.Retaining response probabilities and residual-mass annotations without text improves AgentSense and ProMediate, while the corresponding SOTOPIA and SOTOPIA-ToM intervals include zero.Source: §J.1 · p. 24
• Among SFT-only predictors, DSWM has higher rank agreement, lower regret and lower mean RMSE than direct expectation in all three environments.The SOTOPIA RMSE difference is small even though rank agreement differs (0.69 vs. 0.64).Source: §J.1 · p. 25, Table 16 · p. 27
• Dominance filtering by expected outcome values is a heuristic that can remove an action with higher cluster-averaged nonlinear utility; the conflict frequency and utility cost are unquantified.Auditing those conflicts would require candidate predictions before and after filtering, with constraint eligibility held fixed.Source: §J.1 · p. 26

K Probability Perturbation and Measure Diversity

• Probability perturbation tests sensitivity to assignments while holding response text and conditional means fixed; measure-diversity training varies supervision breadth at fixed volume.The training inventory contains 25 measures, with conditions retaining 5, 10, 15, 20 or 25.Source: K Probability Perturbation and Measure… · p. 28