At a Glance
Learning compact summaries and generating synthetic summary–parameter pairs can replace many direct simulator runs: one tested setup recovered full posterior structure with 10x fewer direct simulations and lower total runtime.
ON THIS PAGE
Key Findings
A two-stage approach first learns a low-dimensional summary of time-series outputs, then trains a conditional generator to produce synthetic summaries from parameters and a second model to recover parameters from those summaries. Using synthetic summaries to augment a small pool of real simulations can give accurate posterior estimates while cutting simulator calls and runtime. The best allocation and training order depend on problem noise and model structure: a mixed strategy (1,000 real sims + synthetic to reach 10,000 total) worked well overall, while an ordered synthetic-then-real strategy beat others in a noisier setting. Event-Driven Agent Pattern.
Key Data
110x fewer direct simulations: the Mixed 1:9 setup used 1,000 direct simulator runs versus 10,000 for the standard sequential baseline.
2Posterior comparisons used 1,000 reference and 1,000 estimated draws (n = m = 1,000) to compute Wasserstein and MMD2 distances.
3Synthetic-to-Simulation with a 3:7 allocation achieved the lowest median posterior discrepancy in the noisy Brock–Hommes setting.
Why It Matters
Engineers and teams that calibrate agent-based simulations or other costly simulators and want to cut simulation cost without losing posterior quality. Technical leaders choosing inference workflows who need trade-offs between simulator budget and wall-clock time. Researchers in simulation-based inference exploring surrogate or synthetic-data augmentation strategies. Evaluation-Driven Development (EDDOps)
Not sure where to start?Get personalized recommendations
Key Figures

Fig 2: (a) Approximate ground truth

Fig 4: (a) Approximate ground truth

Fig 9: (a) Smooth BH: Syn. → \to Sim.
Ready to evaluate your AI agents?
Learn how ReputAgent helps teams build trustworthy AI through systematic evaluation.
Learn MoreLimitations
Results come from two settings of a single benchmark model with one observed trajectory per setting, so transfer to other agent-based models is not guaranteed. Gains depend on the encoder retaining parameter information; a poor summary will harm posterior recovery even with many synthetic samples. Synthetic augmentation can reduce or increase error depending on allocation and training order—validate with posterior-distance checks and predictive checks for your model before deploying. Orchestration
The Details
The method learns an 8-dimensional learned summary of time-series simulator outputs using a recurrent encoder. A conditional generator (the latent surrogate) learns to produce those summaries from parameter draws; a second model learns the reverse mapping from summaries to parameters (the posterior). During sequential rounds the encoder and surrogate are updated from a limited set of real simulator runs; stored trajectories are re-encoded and synthetic summaries are regenerated so all training pairs share the current representation. Experiments on two settings of the Brock–Hommes model (which has a tractable reference posterior) compare three posterior-training strategies and several allocations between directly-simulated and synthetic pairs. A mixed training scheme that combined 1,000 direct simulations with 9,000 synthetic summaries (Mixed 1:9) matched or improved mean posterior distances while reducing runtime compared with a 10,000-direct-simulation baseline. However, the optimal strategy depends on noise and model specifics: an ordered synthetic-then-simulation approach yielded the best median discrepancy for the noisy setting. Practical takeaway: synthetic summaries can substantially reduce simulator demand, but you must monitor posterior recovery metrics and test allocations/ordering for your model because encoder quality and feedback across sequential rounds shape outcomes. Planning Pattern Agent Registry Pattern
Avoid common pitfallsLearn what failures to watch for
Credibility Assessment:
ArXiv preprint with no visible affiliations, citations, or clearly known author reputations — limited information.