Anticipating Social Consequences with Distributional World Models

Viet-Thanh Pham; Thuy-Trang Vu; Zhuang Li; Lizhen Qu; Gholamreza Haffari

cs.AIOct 22, 2026

Anticipating Social Consequences with Distributional World Models

Authors: Viet-Thanh Pham, Thuy-Trang Vu, Zhuang Li, Lizhen Qu, Gholamreza Haffari

Organizations: Department of Data Science & AI, Monash University, Australia · Royal Melbourne Institute of Technology, Australia

Abstract

Social world models predict how interactions develop after an action, helping agents negotiate, cooperate and mediate. In these situations, one action can lead to several plausible futures, hence predicting only one future can overlook consequential alternatives, while exploring them through repeated simulations increases inference cost. We propose DSWM, a Distributional Social World Model that predicts alternative responses, their probabilities and immediate and final outcome values in one output per candidate action. We train a 4B model on repeated simulations using supervised fine-tuning and GRPO. At inference time, the actor uses these possible consequences to choose an action. Across three seen and two unseen environments, with interaction backbones absent from training collection, DSWM achieves the best average performance on 13 of 15 metrics. Ablations show that retaining multiple futures and their predicted probabilities improves outcomes in all five environments.

Figures & tables

Figure 1: How distributional future states guide decisions.

Figure 2: Training and inference with DSWM: repeated simulations yield response clusters, probabilities and outcome values for training; at inference, predicted future states guide filtering and actor selection. Some of the 50 trials are drawn. Dotted branches indicate ordinary continuation for final labels; $\mu_{i}$ and $\mu_{f}$ denote immediate and final means. Values and action IDs are illustrative.

	Magentic-NL			SOTOPIA		AgentSense
Method	Cust. util.	Serv. rev.	Needs met	Goal	Overall	Goal	Info. acc.
Actor alone	179.98 $\,\pm\,$ 8.30	116.84 $\,\pm\,$ 7.34	0.35 $\,\pm\,$ 0.021	6.47 $\,\pm\,$ 0.24	2.45 $\,\pm\,$ 0.15	0.66 $\,\pm\,$ 0.027	0.86 $\,\pm\,$ 0.025
Candidates only	195.46 $\,\pm\,$ 11.44	125.20 $\,\pm\,$ 8.51	0.37 $\,\pm\,$ 0.023	6.77 $\,\pm\,$ 0.19	2.48 $\,\pm\,$ 0.11	0.66 $\,\pm\,$ 0.024	0.88 $\,\pm\,$ 0.027
Single future	206.05 $\,\pm\,$ 8.20	125.15 $\,\pm\,$ 8.29	0.37 $\,\pm\,$ 0.028	6.49 $\,\pm\,$ 0.23	2.63 $\,\pm\,$ 0.13	0.67 $\,\pm\,$ 0.019	0.91 $\,\pm\,$ 0.026
Direct expectation	208.68 $\,\pm\,$ 12.20	129.33 $\,\pm\,$ 6.05	0.36 $\,\pm\,$ 0.030	7.59 $\,\pm\,$ 0.23	2.49 $\,\pm\,$ 0.14	0.70 $\,\pm\,$ 0.022	0.88 $\,\pm\,$ 0.029
Sampled futures	255.56 $\,\pm\,$ 9.65	150.01 $\,\pm\,$ 6.65	0.38 $\,\pm\,$ 0.024	7.10 $\,\pm\,$ 0.23	2.92 $\,\pm\,$ 0.15	0.59 $\,\pm\,$ 0.021	0.80 $\,\pm\,$ 0.024

Table 1: Held-out scenario families of the training environments, evaluated with backbones absent from training rollouts. Mean

\pm

SD over five seeds, each averaging three backbones; higher is better on every metric. Customer utility and service revenue are reported for the corresponding focal role. Green shading and bold mark the best mean; underlining marks the second-best distinct mean. Ties share formatting. GDP-Zero applies only to two-party settings.

Figure 3: SOTOPIA choices and recorded reasoning with expected values and eligibility fixed. C achieves the best observed overall outcome. Two clusters (main evaluation: four); text abbreviated. Details: Appendix J.1 .

	ProMediate			SOTOPIA-ToM
Method	Cons. chg.	Topic eff.	Med. intel.	InfoMgmt	DA	IA	CPV $\downarrow$	EFF
Actor alone	0.0576 $\pm$ 0.0058	0.0030 $\pm$ 0.0004	3.2179 $\pm$ 0.200	0.1769 $\pm$ 0.0174	0.3538 $\pm$ 0.0205	0.1125 $\pm$ 0.0138	0.3487 $\pm$ 0.0218	0.6961 $\pm$ 0.0294
Candidates only	0.0616 $\pm$ 0.0053	0.0031 $\pm$ 0.0004	3.3219 $\pm$ 0.153	0.1789 $\pm$ 0.0126	0.3764 $\pm$ 0.0227	0.1151 $\pm$ 0.0153	0.3689 $\pm$ 0.0179	0.7525 $\pm$ 0.0328
Single future	0.0452 $\pm$ 0.0039	0.0028 $\pm$ 0.0003	3.2051 $\pm$ 0.137	0.1795 $\pm$ 0.0115	0.3692 $\pm$ 0.0244	0.1278 $\pm$ 0.0127	0.3475 $\pm$ 0.0282	0.8154 $\pm$ 0.0313
Direct expectation	0.0440 $\pm$ 0.0050	0.0034 $\pm$ 0.0003	3.2919 $\pm$ 0.136	0.1840 $\pm$ 0.0133	0.4736 $\pm$ 0.0274	0.1352 $\pm$ 0.0147	0.3376 $\pm$ 0.0177	0.8813 $\pm$ 0.0391
Sampled futures	0.0569 $\pm$ 0.0054	0.0027 $\pm$ 0.0003	3.2914 $\pm$ 0.189	0.2160 $\pm$ 0.0146	0.4650 $\pm$ 0.0192	0.1249 $\pm$ 0.0145	0.3071 $\pm$ 0.0209	0.8509 $\pm$ 0.0382

Table 2: Transfer to unseen environments and backbones; aggregation follows Table 1 . Mean

\pm

s.d. over seeds. Higher is better except CPV; highlighting follows Table 1 . GDP-Zero is inapplicable to these multi-party settings.

	Magentic-NL	SOTOPIA	AgentSense	ProMediate	SOTOPIA-ToM
Selection variant	Cust. util. $\uparrow$	Goal $\uparrow$	Goal $\uparrow$	Cons. chg. $\uparrow$	InfoMgmt $\uparrow$
Full DSWM	272.82 $\,\pm\,$ 7.91	8.21 $\,\pm\,$ 0.20	0.72 $\,\pm\,$ 0.023	0.0628 $\,\pm\,$ 0.0047	0.2508 $\,\pm\,$ 0.0119
Without response text	272.82 $\,\pm\,$ 7.91	8.02 $\,\pm\,$ 0.25	0.63 $\,\pm\,$ 0.025	0.0574 $\,\pm\,$ 0.0044	0.2348 $\,\pm\,$ 0.0112
Expected values only	272.82 $\,\pm\,$ 7.91	7.98 $\,\pm\,$ 0.19	0.60 $\,\pm\,$ 0.025	0.0511 $\,\pm\,$ 0.0038	0.2306 $\,\pm\,$ 0.0140
Uniform response probabilities	254.60 $\,\pm\,$ 8.35	8.16 $\,\pm\,$ 0.20	0.68 $\,\pm\,$ 0.028	0.0552 $\,\pm\,$ 0.0058	0.2295 $\,\pm\,$ 0.0148
Most likely response only	243.85 $\,\pm\,$ 10.00	7.87 $\,\pm\,$ 0.17	0.67 $\,\pm\,$ 0.025	0.0526 $\,\pm\,$ 0.0044	0.2187 $\,\pm\,$ 0.0112

Table 3: Component ablations; aggregation follows Table 1 . The future representations preserve numerical guidance and filters, while probability interventions change them. Higher is better. Highlighting follows Table 1 .

Predictor	Magentic	SOTOPIA	AgentSense
Single future	0.201 $\,\pm\,$ 0.010	0.178 $\,\pm\,$ 0.008	0.164 $\,\pm\,$ 0.013
Direct expectation	0.171 $\,\pm\,$ 0.009	0.134 $\,\pm\,$ 0.010	0.158 $\,\pm\,$ 0.014
Sampled futures	0.145 $\,\pm\,$ 0.009	0.151 $\,\pm\,$ 0.008	0.169 $\,\pm\,$ 0.011
DSWM (SFT)	0.139 $\,\pm\,$ 0.010	0.133 $\,\pm\,$ 0.008	0.142 $\,\pm\,$ 0.012
Full DSWM	0.127 $\,\pm\,$ 0.007	0.116 $\,\pm\,$ 0.011	0.135 $\,\pm\,$ 0.011

Table 4: Normalized prediction RMSE (

\downarrow

) on shared retained cases (94%); aggregation follows Table 1 . Tables 14 and 16 add cost, validity, rankings and regret. Highlighting follows Table 1 .

Panel of Figure 4 — Figure 4: Sensitivity results (% of full-condition mean), with SD bands. Protocols and raw values: Appendix K .

Appendix figures & tables16 assets

Supplementary material from the paper’s appendix.

Appendix

Data	Mag.-NL	SOTOPIA	AgentS.	Total
Original	4,975	4,766	1,118	10,859
Recovered	1,023	0	186	1,209
New	43	44	92	179
Full pool	6,041	4,810	1,396	12,247

Appendix

Table 5: Composition of the full training pool. Both training stages use all three sources.

Environment	Test queries	Final queries	Visible contexts	Context–action pairs	Held-out-spec. queries
Magentic-NL	1,316	485	206	936	936
SOTOPIA	1,036	335	133	511	518
AgentSense	379	106	95	379	379
Total	2,731	926	434	1,826	1,833

Appendix

Table 6: Held-out prediction-input inventory. Final queries are test rows explicitly requesting the final horizon. Query files share the contexts and candidate actions shown here. Latest diagnostic references use 50 continuations per action. These source counts are distinct from the online episode counts in Table 9 and from post-exclusion diagnostic denominators.

Setting	Value
LoRA	rank 64, $\alpha$ 128, dropout 0.05; attention and MLP projections
SFT schedule	2 balanced epochs; final checkpoint
SFT learning rate	$10^{-4}$ ; cosine, 3% warmup
SFT batch / length	1 $\times$ accumulation 16; 8,192 tokens
GRPO schedule	1 pass over the eligible weighted pool
GRPO learning rate	$2\times 10^{-6}$ ; constant after 10 warmup steps

Appendix

Table 7: Settings for each SFT and SFT + GRPO training run. Main Tables 1 – 4 average five training/evaluation seeds; the separate reliability audit uses three.

Benchmark	Use	Runner and deviations	Main metrics (native scale)	Test
Magentic-NL	train	New two-party customer/service negotiation built from Magentic Marketplace data ( Microsoft, 2025 ) (restaurant and home-contractor scenario families). It is not the official marketplace benchmark. Each turn is free text plus an optional act (offer, counter, accept, decline); at most 12 messages; scenario-defined price cap; deterministic negotiation style per scenario.	customer utility, service revenue, needs met (deterministic)	90
SOTOPIA	train	450 benchmark combinations and prompts ( Zhou et al., 2024 ) ; at most 20 turns or until an agent leaves. The local judge uses the SOTOPIA-EVAL rubrics in an object schema.	seven SOTOPIA-EVAL dimensions, including goal (0–10)	70
AgentSense	train	Official template ( Mou et al., 2025 ) ; random next speaker with no repeats; 15 messages; first paragraph only, 128 tokens. Official goal questions, but a single judge instead of the official three-model ensemble.	goal completion (0–1); information questions where available	270
ProMediate	unseen	Official cases, personas, conflict-mode, mediator and judge prompts and metric formulas ( Liu et al., 2026 ) . Compact motivation-ranked turn-taking replaces the released engine; at most 40 utterances (official: $4\times$ parties $\times$ topics); known released-code bugs avoided; initial judgments precomputed.	consensus change; topic-level efficiency; mediator intelligence	18
SOTOPIA-ToM	unseen	We reconstruct the runner and judge from the released gold set and the paper ( YS et al., 2026 ) : round-robin turns, public and direct-message channels, at most 60 turns, early stop. The DA/IA/CPV/EFF judge prompts are our reimplementation; the formulas follow the paper; metrics are computed for the focal agent.	InfoMgmt; DA, IA, CPV, EFF	160

Appendix

Table 8: Benchmarks as implemented in our shared framework. “Test” counts held-out scenario configurations, before crossing focal roles, five seeds and three backbones (Table 9 ).

Environment	Total	Train	Val.	Test	Test role configs.	Episodes / method
Magentic-NL	439	306	43	90	180	2,700
SOTOPIA	450	325	55	70	140	2,100
AgentSense	1,225	865	90	270	650	9,750
ProMediate	18	0	0	18	18	270
SOTOPIA-ToM	160	0	0	160	536	8,040

Appendix

Table 9: Scenario inventory and complete online evaluation grid. Total, Train, Val. and Test count scenario configurations; Train describes the eligible scenario split, not the number of prediction-query rows. Test role configurations count each eligible focal role separately, per backbone and seed. Episodes multiplies this count by three backbones and five seeds. Counts apply to Tables 1 – 3 for each applicable method, not to the separate reliability audit or saved-state prediction diagnostics.

Method	Information at decision time	WM training
Actor alone	Ordinary action prompt, without candidate selection	none
Candidates only	Up to four actor proposals; selection without predictions	none
Single future	One predicted response segment and its outcome values	matched SFT
Direct expectation	Expected value of each requested outcome measure and horizon; no response text	matched SFT
Sampled futures	30 independent response segments with outcome values per candidate	matched SFT
S3AP F&A	A prompted next-state prediction and the Foresee-and-Act procedure ( Zhou et al., 2025 )	none

Appendix

Table 10: Evaluated decision-time methods. Matched SFT uses the same base model, LoRA settings, source rollout pool and SFT budget. It does not include the additional GRPO stage.

Method	Component	Setting
S3AP F&A	Planning	One provisional action, one next-state prediction, one refinement
	Prediction / refinement	Temperature 0.7 / 0.7; structured output
	Models	Internal world model matches the evaluation backbone of the focal actor
GDP-Zero	Search	20 simulations; 3 realizations per state; $c_{\mathrm{puct}}=1$ ; $Q_{0}=0$
	Simulated utterances	Actor temperature 0.7; partner temperature 1.1; 128-token limit; 5-turn history window
	Action prior	15 samples at temperature 1.0; additive smoothing 1.0

Appendix

Table 11: External-baseline configurations. GDP-Zero uses the 20-simulation, zero-initial-value setting from the experimental runner. Other algorithmic settings follow the released implementations; benchmark interfaces, evaluation backbones and judge are shared as described in Appendix D .

Figure 7: Model-guided action selection. Predictions guide filtering and inform action selection by the same actor; the chosen action is executed unchanged. Candidate IDs are illustrative.

Prediction	Order	InfoMgmt $\uparrow$	CPV $\downarrow$
Direct expectation	Average first	0.1846	0.3382
Sampled futures (30)	Average first	0.2114	0.3097
Sampled futures (30)	Combine first	0.2238	0.3121
DSWM	Average first	0.2371	0.2674
DSWM	Combine first	0.2462	0.2586

Appendix

Table 12: SOTOPIA-ToM utility-aggregation comparison, separate from the five-seed main tables; DSWM uses SFT + GRPO. Average first computes utility from expected outcome values; combine first averages utilities of individual samples or cluster means. Only the utility line changes within each prediction format. InfoMgmt is maximized; CPV is minimized. Highlighting follows Table 1 .

Environment	Seeds	Pairs	$\Delta\uparrow$ [95% CI]	Failure (%) $\downarrow$		Judge $\rho\uparrow$	Human $n$
				DSWM	Samples
Magentic-NL	3	300	$14.8$ $[4.2,\,25.6]$	0.0	0.0	n/a	n/a
SOTOPIA	3	300	$1.11$ $[0.90,\,1.32]$	0.0	0.0	0.76	100
AgentSense	3	300	$0.08$ $[0.02,\,0.14]$	0.0	0.0	0.71	100
ProMediate	3	300	$0.0043$ $[-0.0011,\,0.0097]$	0.0	0.0	0.68	100
SOTOPIA-ToM	3	300	$0.026$ $[0.008,\,0.044]$	0.0	0.0	0.73	100

Appendix

Table 13: Reliability audit for DSWM versus 30 sampled futures on a separate evaluation schedule. The audit uses 300 scheduled pairs across three training seeds per environment. All 300 pairs completed in each environment; both methods have zero episode failures. Differences and intervals describe all 300 completed pairs. Human

n

counts interactions, all annotated by one author. Judge agreement is inapplicable to deterministic Magentic-NL payoffs. Primary measures, sampling and failure rules are specified in Appendix J .

Predictor	RMSE $\downarrow$	Rank $\uparrow$	Raw valid (%) $\uparrow$	Tokens / cand. $\downarrow$	Latency (ms) $\downarrow$
Single future	0.178 $\,\pm\,$ 0.008	0.49 $\,\pm\,$ 0.024	98.8 $\,\pm\,$ 0.6	438 $\,\pm\,$ 12	610 $\,\pm\,$ 46
Direct expectation	0.134 $\,\pm\,$ 0.010	0.64 $\,\pm\,$ 0.029	99.6 $\,\pm\,$ 0.6	112 $\,\pm\,$ 3	185 $\,\pm\,$ 13
Sampled futures (30)	0.151 $\,\pm\,$ 0.008	0.69 $\,\pm\,$ 0.026	92.1 $\,\pm\,$ 0.6	13140 $\,\pm\,$ 328	6200 $\,\pm\,$ 474
DSWM (SFT)	0.133 $\,\pm\,$ 0.008	0.69 $\,\pm\,$ 0.027	96.8 $\,\pm\,$ 0.5	760 $\,\pm\,$ 17	1160 $\,\pm\,$ 96
DSWM (SFT + GRPO)	0.116 $\,\pm\,$ 0.011	0.71 $\,\pm\,$ 0.030	98.1 $\,\pm\,$ 0.5	812 $\,\pm\,$ 25	1280 $\,\pm\,$ 91

Appendix

Table 14: Final-horizon SOTOPIA diagnostics on shared saved states, reported as mean

\pm

SD across five seeds after averaging three backbones: RMSE of normalized expected outcome values (

\downarrow

), candidate-ranking agreement (Kendall’s

\tau_{b}

\uparrow

), valid raw requests (

\uparrow

), generated tokens and latency per candidate. Accuracy uses the shared retained cases; costs include retries. Highlighting follows Table 1 .

	Magentic-NL	SOTOPIA	AgentSense	ProMediate	SOTOPIA-ToM
Comparison	Cust. util. $\uparrow$	Goal $\uparrow$	Goal $\uparrow$	Cons. chg. $\uparrow$	InfoMgmt $\uparrow$
SFT $-$ direct expectation	$52.60$ $[43.10,61.90]$	$0.240$ $[0.050,0.430]$	$-0.030$ $[-0.060,0.005]$	$0.0173$ $[0.0101,0.0245]$	$0.0333$ $[0.0220,0.0446]$
Full $-$ direct expectation	$64.14$ $[55.20,73.40]$	$0.620$ $[0.430,0.820]$	$0.020$ $[-0.006,0.046]$	$0.0188$ $[0.0114,0.0262]$	$0.0668$ $[0.0554,0.0781]$
Full $-$ sampled futures	$17.26$ $[8.10,26.40]$	$1.110$ $[0.900,1.320]$	$0.130$ $[0.099,0.161]$	$0.0059$ $[0.0014,0.0132]$	$0.0348$ $[0.0239,0.0458]$
Full $-$ SFT	$11.54$ $[3.80,19.20]$	$0.380$ $[0.210,0.560]$	$0.050$ $[0.019,0.081]$	$0.0015$ $[-0.0027,0.0058]$	$0.0335$ $[0.0231,0.0439]$
Full $-$ without response text	$0.00$ $[-1.20,1.10]$	$0.190$ $[0.006,0.374]$	$0.090$ $[0.055,0.125]$	$0.0054$ $[0.0009,0.0115]$	$0.0160$ $[0.0052,0.0269]$

Appendix

Table 15: Paired differences on the main evaluation grid. Each cell shows the difference between reported means (first method minus second) above its 95% confidence interval from 10,000 bootstrap resamples over scenarios and five training/evaluation seeds. Full denotes DSWM (SFT + GRPO); SFT denotes DSWM (SFT only). Positive differences favor the first method. Intervals account for scenario and seed variation, conditional on the three evaluation backbones.

Environment	Predictor	RMSE $\downarrow$	$\tau_{b}\uparrow$	Regret $\downarrow$
Magentic-NL	Single future	0.201 $\,\pm\,$ 0.010	0.43 $\,\pm\,$ 0.038	0.118 $\,\pm\,$ 0.008
	Direct expectation	0.171 $\,\pm\,$ 0.009	0.58 $\,\pm\,$ 0.027	0.086 $\,\pm\,$ 0.006
	Sampled futures (30)	0.145 $\,\pm\,$ 0.009	0.66 $\,\pm\,$ 0.029	0.071 $\,\pm\,$ 0.006
	DSWM (SFT)	0.139 $\,\pm\,$ 0.010	0.68 $\,\pm\,$ 0.032	0.065 $\,\pm\,$ 0.007
	DSWM (SFT + GRPO)	0.127 $\,\pm\,$ 0.007	0.73 $\,\pm\,$ 0.030	0.054 $\,\pm\,$ 0.009
SOTOPIA	Single future	0.178 $\,\pm\,$ 0.008	0.49 $\,\pm\,$ 0.024	0.094 $\,\pm\,$ 0.005

Appendix

Table 16: Final-horizon prediction diagnostics, mean

\pm

SD across five seeds after averaging three evaluation backbones. RMSE and regret use normalized scales; regret concerns the highest predicted-utility candidate before filtering. All methods use the same retained cases (94% of the diagnostic pool). Higher

\tau_{b}

and lower RMSE and regret are better. Highlighting follows Table 1 .

Figure 8: Hypothetical utility example. Equal expected outcome values need not imply equal cluster-averaged nonlinear utility. The calculation illustrates aggregation, not an observed decision.

		Magentic-NL	SOTOPIA	AgentSense	ProMediate	SOTOPIA-ToM
$\alpha$	JS	Cust. util. $\uparrow$	Goal $\uparrow$	Goal $\uparrow$	Cons. chg. $\uparrow$	InfoMgmt $\uparrow$
0.00	0.00	272.82 $\,\pm\,$ 7.91	8.21 $\,\pm\,$ 0.20	0.720 $\,\pm\,$ 0.023	0.0628 $\,\pm\,$ 0.0047	0.2508 $\,\pm\,$ 0.0119
0.25	0.04	265.73 $\,\pm\,$ 8.24	8.23 $\,\pm\,$ 0.22	0.700 $\,\pm\,$ 0.025	0.0622 $\,\pm\,$ 0.0049	0.2460 $\,\pm\,$ 0.0124
0.50	0.10	251.27 $\,\pm\,$ 8.86	8.02 $\,\pm\,$ 0.24	0.680 $\,\pm\,$ 0.028	0.0602 $\,\pm\,$ 0.0052	0.2347 $\,\pm\,$ 0.0131
0.75	0.18	230.81 $\,\pm\,$ 9.42	7.73 $\,\pm\,$ 0.27	0.650 $\,\pm\,$ 0.030	0.0581 $\,\pm\,$ 0.0056	0.2202 $\,\pm\,$ 0.0143
1.00	0.28	209.53 $\,\pm\,$ 10.17	7.36 $\,\pm\,$ 0.30	0.610 $\,\pm\,$ 0.033	0.0536 $\,\pm\,$ 0.0060	0.2021 $\,\pm\,$ 0.0155

Appendix

Table 17: Probability perturbation: mean

\pm

SD across five seeds after averaging three backbones. JS is achieved divergence from the original prediction, pooled across environments. The five primary outcome measures match Table 3 ; higher is better. Highlighting follows Table 1 .

	ProMediate			SOTOPIA-ToM
Measures	Cons. chg. $\uparrow$	Topic eff. $\uparrow$	Med. intel. $\uparrow$	InfoMgmt $\uparrow$	DA $\uparrow$	IA $\uparrow$	CPV $\downarrow$	EFF $\uparrow$
5 (20%)	0.0506 $\pm$ 0.0062	0.0030 $\pm$ 0.0006	3.5742 $\pm$ 0.244	0.1851 $\pm$ 0.0176	0.4512 $\pm$ 0.0311	0.1286 $\pm$ 0.0187	0.3041 $\pm$ 0.0356	0.8214 $\pm$ 0.0412
10 (40%)	0.0546 $\pm$ 0.0058	0.0032 $\pm$ 0.0005	3.8147 $\pm$ 0.223	0.2079 $\pm$ 0.0158	0.4807 $\pm$ 0.0297	0.1423 $\pm$ 0.0172	0.2870 $\pm$ 0.0311	0.8538 $\pm$ 0.0376
15 (60%)	0.0582 $\pm$ 0.0054	0.0035 $\pm$ 0.0005	4.0613 $\pm$ 0.205	0.2280 $\pm$ 0.0144	0.5085 $\pm$ 0.0279	0.1551 $\pm$ 0.0159	0.2674 $\pm$ 0.0254	0.8846 $\pm$ 0.0340
20 (80%)	0.0610 $\pm$ 0.0049	0.0037 $\pm$ 0.0004	4.2418 $\pm$ 0.193	0.2423 $\pm$ 0.0128	0.5284 $\pm$ 0.0256	0.1652 $\pm$ 0.0139	0.2519 $\pm$ 0.0210	0.9041 $\pm$ 0.0319
25 (100%)	0.0628 $\pm$ 0.0047	0.0038 $\pm$ 0.0004	4.3321 $\pm$ 0.185	0.2508 $\pm$ 0.0119	0.5398 $\pm$ 0.0243	0.1695 $\pm$ 0.0128	0.2450 $\pm$ 0.0173	0.9138 $\pm$ 0.0300

Appendix

Table 18: Training-measure diversity: mean

\pm

SD across five seeds after averaging three backbones. The inventory contains 25 outcome measures; training supervision is matched across conditions. Higher is better except CPV (

\downarrow

). Highlighting follows Table 1 .

World Model-Based Planning

World model-based planning uses models of how environments change to compare imagined actions and choose goal-directed behavior.

See all 261 papers

Jun 9, 2026

Building Social World Models with Large Language Models

Understanding and predicting how social beliefs evolve in response to events -- from policy changes to scientific breakthroughs -- remains a fundamental challenge in social science. Given LLMs' commonsense knowledge and social intelligence, we ask: Can LLMs model the dynamics of social beliefs following social events? In this work, we introduce the concept of the Social World Model (SWM), a general framework designed to capture how social beliefs evolve in response to major events. SWM learns state-transition functions for social beliefs by mining temporal patterns in social data and optimizing the evidence lower bound, without the need for explicit human annotations linking events to belief shifts, or for expensive census data. To evaluate SWM, we introduce a benchmark, SWM-bench, derived from real-world prediction markets, specifically Kalshi and Polymarket. SWM-bench includes over 12k data points for social belief prediction tasks spanning diverse domains such as politics, finance, and cryptocurrency. Our experimental results show that SWM significantly outperforms time-series foundation models, achieving state-of-the-art results on Kalshi data and demonstrating competitive performance on Polymarket data, while offering interpretable insights into the underlying mechanisms of social belief dynamics.

Haofei Yu, Yining Zhao, Guanyu Lin +1

University of Illinois Urbana-Champaign · Carnegie Mellon University

World Model Learning World Models Computational Social Science Social Reasoning

Ask PDF

Oct 21, 2026

From Single-Actor Environments to Multi-Actor Event Worlds: Can LLMs Be Event World Models?

World models help agents assess the consequences of their actions before acting. Many real systems, however, involve multiple actors who make their own decisions, can evolve even when the agent takes no action, and are only partially observed through event logs. We formalize these systems as event worlds and ask whether large language models can predict future observations from the log and the agent's proposed actions. We build complementary benchmarks from human Diplomacy games and real Polymarket trades. Our hierarchical framework compares prediction with and without other actors' actions, then tests multi-step rollouts, predictions for actions never taken, and usefulness to agents. Fine-tuning and more data improve prediction when all actors' recorded actions are given. With only the agent's action, models struggle when other actors strongly affect the outcome, and larger models and longer histories help little. This gap also limits multi-step prediction and agent decisions, with models overestimating plan success. Yet imagined experience and initialization from world models can still improve policy learning. Anticipating other actors' actions remains a central challenge for event world modeling.

Juntong Ni, Zewen Liu, Bohan Wang +2

Emory University

LLM World Models World Models

Ask PDF

Aug 4, 2026

SocietyBench: Forecasting Counterfactual Social-World Evolution

Large language models (LLMs), and the agents built on top of them, are now benchmarked heavily on whether they can finish a task -- fix a bug, drive a browser, operate a GUI. A complementary social ability, namely how well a model understands and forecasts the way real social events unfold, has barely been measured. We introduce SocietyBench, an end-to-end benchmark that takes a one-line event topic, collects Web news and social-media posts across five platforms, distills them into a date-indexed timeline that keeps factual events and a public-opinion layer separate, and then turns every cutoff date on that timeline into an audited bank of forecasting questions. Questions are scored on two orthogonal 100-point axes: probability calibration and temporal accuracy. Before any model sees a timeline, a three-phase procedure replaces every named entity and shifts every date by a per-event constant, turning a real arc into a counterfactual social world -- structurally identical to what happened, but stripped of the surface labels a model could match against pre-training memory. On five heterogeneous events and 125 prediction points in Chinese and English editions, the strongest of six frontier LLMs reaches only 75.0 out of 100, against a trivial anchor of 50. The two axes come apart: a model can be calibration-strong but time-weak, or the reverse. Three agent frameworks built on a shared base model fail to improve on that base, and two model-free heuristics trail every LLM. Per-event gaps reach 21.4 points on a single axis, which is our main argument for evaluating on several events rather than one. All anonymized timelines, question banks, ground truth, and scoring code are released.

Zhenran Wang, Zhonghan Bian, Jinsong Li +1

Counterfactual Evaluation Forecasting Benchmarks Probabilistic Forecasting Language Model Calibration

Ask PDF

Abstract

World Model-Based Planning

Explore similar work

Building Social World Models with Large Language Models

From Single-Actor Environments to Multi-Actor Event Worlds: Can LLMs Be Event World Models?

SocietyBench: Forecasting Counterfactual Social-World Evolution