arXiv subjects

Tengfei Shao

Publications and source records attributed to Tengfei Shao.

3 recordsLinked to original sources

arXiv2026

Testing, not presuming, adequacy: calibrating generative social simulators against emergent network structure

Validation of generative social simulators often stops at face validity: emergent network structure is compared descriptively, without quantified parameter uncertainty or an adequacy check. We present an adequacy-aware calibration protocol that couples amortized posterior estimation with a synthetic identifiability assessment, a matched-sample-size adequacy check (prior-predictive reachability plus per-statistic posterior-predictive localization), a diagnosis-guided repair, and a statistic-held-out audit. We demonstrate it on a real second-hand luxury resale market with four channel-by-residency cells, each a bipartite buyer-brand network, using a forward model built from persona profiles elicited once, offline, by a language model. The behavioural parameters are recoverable in all four cells, though calibration is approximate and overconfident for one parameter. The observed summary falls outside the simulator's reachability reference in every cell, with the mean purchased tier as the pervasive discrepancy. The repair meets the value-block criterion in two of four cells but does not restore adequacy, and the held-out audit surfaces a buyer-breadth-dispersion miss no earlier diagnostic detected. A profile-source ablation finds the language-model profiles beat a flat rule baseline in all four cells, yet within-category brand relabelling causes no consistent degradation, so the profiles are a partially validated input whose value rests on structure, not brand identity. Making no causal claim, we conclude that an independent-aggregation account, without agent interaction or a buyer-breadth mechanism, cannot jointly reproduce the market's purchased-tier level, head-brand concentration, community structure and buyer-breadth heterogeneity.

cs.AI↗

arXiv2026

Auditing bipartite motif interpretations: a worked example with conservation checks and open-path decomposition

Motif profiles of bipartite agent-object networks, such as tourist-site visits and customer-item transactions, are read as evidence about structural roles and about differences between networks, often without asking what the two degree sequences already fix. In a simple bipartite graph the induced k-fan count on one node type is a sum of degree combinations, so it has zero variance under a null that preserves both degree sequences. We apply this known result to a reconstructed tourism rating network of 17 tourists, 80 sites and 637 edges, the sole inferential worked example, and, as a provenance-limited illustration, to published motif-instance aggregates over 36 monthly luxury customer-item networks. The four fan classes are exact functions of the degree sequences: in the tourism network the raw fan counts and the size-3 two-fan ratio (84.7% fan-out) restate those sequences. The published luxury counts require at least 89,502 customer-item edges against 26,451 reported transactions, so their 99.8% fan-in is reported as a descriptive value only. Against a hard bipartite configuration null, the four-cycle count is degree-consistent (z about +1.0) and the open path is deficient (z about -5.8) by 2,243 instances, 2.4% of the null mean; the deficit survives every leave-one-tourist-out re-run (z -4.5 to -8.1). An exact identity splits it at the point estimate into 64.4% mixing and 35.6% four-cycle, but that split is not an attribution: the observed mixing term lies below all 500 null samples, the two components are almost collinear under the null (r = 0.968), and the mixing share ranges from 36.6% to 116.5% under leave-one-tourist-out deletion. The deficit is extreme relative to the sampled null, while its class-level interpretation is undetermined and unstable. We give a four-step pre-interpretation check and a reference implementation.

cs.SI↗

arXiv2026

Language-model groups overstate consensus when replaying human deliberation on a reasoning task

Full-consensus rates are often treated as indicators of collective cognition, yet depend on how participation and final states are operationalized. We replayed 100 held-out human Wason groups with matched large language model (LLM) agent groups, seeding one belief-anchored agent per participant's pre-discussion answer and scoring agents and people with the same code. Across human scoring definitions, estimates ranged from 24.0% to 57.0%; about one fifth of participants never posted, whereas agents almost always did. Agent groups remained more consensual in two post-unblinding sensitivity analyses: the submit-based comparison (n = 98) yielded gaps of 34.0 and 43.9 percentage points for chat and reasoning modes, and the participation-matched comparison (n = 45) yielded gaps of 34.1 and 44.4 points. These complementary routes reduced different measurement asymmetries yet converged within 0.5 percentage points. The gap persisted without early stopping and under a reparameterization removing the memorizable answer; reasoning-mode groups then agreed nearly unanimously, mostly on incorrect answers. Simulated consensus did not track collective accuracy, and belief-anchored agent groups were biased estimators of the human group-outcome distribution in this setting. These analyses provide a scoring-explicit basis for assessing simulated-group estimates of human deliberative outcomes.

cs.AI↗