Explore the evidence
All 599 papers in the review’s corpus, with the taxonomy, reading depth and coding from Online Resource 1. Select a bar or a cell to list the papers behind it.
Corpus composition and Figure 3
Papers by primary category and year
Counts from Table 1 and Online Resource 1, with one panel per category on a shared scale. Select a bar to list its papers.
- 2023
- 2024
- 2025
- 2026 (to 27 September)
Table 1 and counting rules
Each paper is counted once, by primary category and the year of its recorded version.
Table 1 Composition of the corpus by category, year, and publication status
| Category | Nodes (papers with that primary node) | Total | 2023 | 2024 | 2025 | 2026 | WP | CP | JA |
|---|---|---|---|---|---|---|---|---|---|
| a. Synthetic economic subjects | Anchoring (5); loss aversion, prospect theory, and reference dependence (9); disposition effect and attention-driven trading (3); herding, social influence, and extrapolation (2); risk and time preferences, beliefs, and rationality tests (30); framing, heuristics, and multi-bias batteries (12); synthetic respondents and expectation formation (8) | 69 | 5 | 14 | 28 | 22 | 39 | 12 | 18 |
| b. Market simulations and agent-based models | Experimental asset-market replications (7); market simulators and platforms (20); microstructure, aggregation, and mechanisms (12); bubbles, runs, and systemic dynamics (8) | 47 | 1 | 5 | 19 | 22 | 31 | 9 | 7 |
| c. Agents performing financial tasks | Trading agents (73); portfolio and asset allocation (48); analysis and forecasting (61); valuation, credit, underwriting, and treasury (27); advice (23); quantitative-research automation (21); benchmarks and arenas (44); reviews and surveys (29) | 326 | 14 | 44 | 127 | 141 | 172 | 87 | 67 |
| d. Audited financial judgments | Demographic, name, and identity cues (13); prompt-embedded reference points and context effects (11); intrinsic priors and preference biases (11); advice quality, suitability, and warnings (14); misconduct, alignment, and compliance (17); calibration, rationality, and coherence (12) | 78 | 3 | 9 | 23 | 43 | 48 | 16 | 14 |
| e. Validity and methodology | Look-ahead, memorization, and contamination (25); reproducibility, output variability, and drift (10); persona and identity prompting (3); calibration to human benchmarks (10); evaluation critiques, protocols, and robustness (24); theory and infrastructure for LLM experiments (7) | 79 | 2 | 3 | 24 | 50 | 60 | 9 | 10 |
| Total | 31 nodes | 599 | 25 | 75 | 221 | 278 | 350 | 133 | 116 |
Note: Each paper is counted once under its primary category and node; 260 also have secondary nodes. The years identify the version recorded, which differs from the arXiv identifier’s year for 21 of the 392 records. The 2026 column ends on September 27. WP = working papers (including arXiv, SSRN, and institutional series); CP = conference proceedings; and JA = journal articles (versions of record).
Evidence map of the 271 primary studies
Fig. 3 of the review. Each cell counts studies with a feature over those in which it could be assessed. Select a cell to list them.
| Interaction structure | Any human comparator | Concurrent same-task humans | Heterogeneity vs humans | Dated model snapshot | Out-of-time validation | Endogenous prices |
|---|---|---|---|---|---|---|
| Independent elicitationn = 215 | ||||||
| Coordination within a systemn = 20 | ||||||
| Strategic interactionn = 40 | ||||||
| Network transmissionn = 6 | ||||||
| Human-AI co-participationn = 4 |
Figure caption
Fig. 3 Evidence map of primary studies in the synthesis set by interaction structure. Each cell reports the number of studies that have the feature, along with the total number of studies in the corresponding row that could be assessed; the share is indicated in parentheses and through the shading. A positive code can come from any reading depth, but an absence counts only for studies read in full, and dated snapshots are assessed only in the full texts. Of the 271 primary studies, 263 had an identified interaction structure, and 21 appeared in more than one row. Heterogeneity was considered when it was compared with human data, regardless of the outcome, among studies whose source of heterogeneity was known. The concurrent same-task column counts humans observed on the same task in the same study. Both comparator columns count against studies whose comparator type is coded or whose full text does not report any. Endogenous prices are included in studies whose price formation can be coded. The out-of-time column keeps designs without historical data in its denominators (Section 3.2), so light cells partly reflect designs to which the check does not apply. Among designs that use historical data, the counts are 35 of 69, 6 of 11, 1 of 10, 0 of 4, and 0 of 1 in row order. Shares in rows with few assessable studies are imprecise
The corpus
Search titles, authors, models and tasks, or narrow the list with the filters. Open a paper to see its identifiers, how it was verified and read, and, for the synthesis and system-audit sets, the coding recorded in Online Resource 1.
Loading the corpus…
The coding was carried out by LLM agents under written protocols, as described in Section 2.3; “not extracted” marks an abstract-only reading and “not reported” means the full text was searched. Online Resource 1 records the evidence locators. The synthesis-set filters cover all 276 synthesis entries, including five secondary syntheses, so their counts can differ slightly from the figures for the 271 primary studies.