Pith. sign in

REVIEW 5 major objections 6 minor 21 references

Evaluating Investment Logic in Large Language Models: A Real-World Benchmark Towards Personalzied Financial Agents

T0 review · 5 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

Pith's one-line read The paper claims that leading LLMs write fluent investment theses but score only about 1.8/5 against expert reasoning traces, and that this event-grounding gap, not missing financial knowledge, explains their live-trading underperformance.

desk verdict A well-framed process-diagnostic benchmark idea undermined by unvalidated ground truth, judge-model entanglement, and internal numeric contradictions. read the letter →

arxiv 2608.06108 v1 pith:LTDX4EYA submitted 2026-08-06 cs.AI

classification cs.AI
keywords investmentlogicLLMevaluationfinancialagentsprocess-nativebenchmarkP-E-R-D-Otraceeventgroundinglivetrading
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper's central claim is that the standard rulers for financial LLMs—static question answering and terminal profit and loss—measure the wrong thing. To show this, the authors built InvestLogicBench2026, a benchmark of 201,247 structured decisions from 151 real-world investors, each recorded as a Profile → Events → Reasoning → Decision → Outcome trace. On the task of generating an investment thesis from market events and an investor profile, four leading LLMs scored near 4/5 for internal logical plausibility but only 0.8–2.8/5 for grounding in the actual events. The paper interprets this as an investment logic gap: models talk coherently without tracking the evidence, and outcome-only evaluation cannot see the difference. A seven-week live trading arena shows the same models underperforming lower-scoring peers and the S&P 500, which the authors attribute to consensus bias and reasoning fragmentation.

What carries the argument

The load-bearing object is the P→E→R→D→O trace—Profile, Events, Reasoning, Decision, Outcome—a structured instantiation of one investment thesis. It is what the benchmark stores for each of its 201,247 decisions and what the evaluation tasks ask models to reproduce or continue. A second piece of machinery is the conditional scoring framework that separates reasoning quality (event coherence plus logical plausibility) from decision outcome (win rate plus total profit), so the two can disagree; this separation is what exposes the polished-but-weakly-grounded pattern. The Stock Investment Logic (SIL) Engine is the device that fills in implicit reasoning from trading history when an investor's explicit thesis is missing.

What would settle it

Re-score the same models on a subset whose ground truth comes only from verbatim, contemporaneous investor notes, with no reconstructed logic; if event-coherence scores rise above 3/5 or cease to predict profitability, the logic-gap conclusion would be refuted.

Watch

Extended reading notes

Core claim

The discovery the paper is trying to establish is that current LLMs exhibit an investment logic gap: their reasoning is fluent and internally coherent but systematically under-anchored to real market events. On InvestLogicBench2026, logical plausibility scores stay near 4/5 across models while event coherence falls as low as 0.8/5, and the reported average reasoning score against expert traces is about 1.8/5. The same models show a capabilities-performance paradox in live trading: higher general-benchmark scores do not translate into returns, and models with lower foundation scores produced the best live results. Two failure mechanisms are identified: consensus bias (reproducing the priced-in mean rather than forming a variant view) and reasoning fragmentation (losing a long-horizon thesis in real-time noise). The intended upshot is that process quality and outcome must be measured together, and that the P→E→R→D→O chain should become the data interface for building and evaluating personalized financial agents.

Load-bearing premise

The benchmark's ground-truth reasoning chains must faithfully represent what real investors actually thought, even though many are reconstructed from YouTube commentary and trading histories by the authors' own inference engine rather than by the investors themselves.

Editorial extensions

If this is right

  • If the logic gap is real, financial benchmarks should stop reporting a single overall score and should report event coherence and logical plausibility separately, because the two can move in opposite directions.
  • Agent builders who want market-aligned behavior need explicit event-noise discrimination and long-horizon consistency checks, rather than relying on knowledge recall or P&L.
  • P&L-only live arenas will keep rewarding lucky or consensus trades; process diagnostics are required to tell a good decision from a good outcome.
  • The P→E→R→D→O schema, as a versioned, temporally provenanced data interface, gives personalization a concrete target: the same market evidence can support different actions for different investor profiles.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The same process-native logic could apply to other consequential agents—medicine, law, engineering—where the risk is that coherent-sounding plans may not be grounded in traceable evidence even when the final outcome looks fine.
  • A testable extension is to inject salient but irrelevant headline events into the input and measure whether event-coherence scores drop; the paper's noise-discrimination claim predicts a clear drop.
  • The benchmark's ground-truth reconstruction could be probed by comparing SIL-inferred reasoning against verbatim contemporaneous notes from the same investor on a held-out subset, if such notes exist.
  • The geographical bias visible in model asset picks suggests event grounding may partly reflect training-data familiarity rather than pure reasoning quality, so cross-locale event sets would sharpen the diagnosis.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

5 major / 6 minor

Summary. The paper introduces InvestLogicBench2026, a benchmark of 201,247 investment decisions from 151 real-world investors, formalized as P→E→R→D→O (Profile, Events, Reasoning, Decision, Outcome) traces. The authors report a seven-week live equity trading study purporting to show a "capabilities-performance paradox," then evaluate four LLMs on a logic-generation task, finding that logical plausibility averages near 4/5 while event coherence is only 0.8–2.8/5. The central claim (Contribution 3) is that LLMs show a significant investment logic gap that exposes weak event grounding. The paper also proposes a data-system interface for personalized financial agents.

Significance. The core idea is timely and valuable: static financial QA benchmarks and terminal profit/loss measures are insufficient for evaluating the reasoning processes of consequential financial agents. The P→E→R→D→O schema and the three-task suite are a sensible step toward process-diagnostic evaluation, and the live trading arena is a useful stress test. If the ground-truth traces were validated, the benchmark could fill a real gap. However, the paper as written does not support its headline quantitative claims: the ground truth is partly reconstructed by a proprietary, unreleased engine; the scoring formula contains a degeneracy; several reported numbers contradict each other; and the human "expert anchors" are ETFs rather than expert decision traces. These are load-bearing issues that affect the validity of every reported score, so the contribution as presented is not yet a reliable benchmark.

major comments (5)
  1. [§3.1, Appendix A.1] The ground-truth traces are not independently validated as genuine expert reasoning. Section 3.1 states that when explicit theses are missing, the proprietary Stock Investment Logic (SIL) Engine "infers implicit logic patterns from an agent's trading history," while Appendix A.1 reveals that the primary source material is YouTube commentary. The paper provides no human annotation, inter-annotator agreement, or release of SIL outputs. Since the same reconstructed traces are then used as the reference for scoring LLM outputs, every reported score measures agreement with the authors' pipeline rather than with documented expert logic. This is a fundamental validity threat to the central claim.
  2. [§4.0.1, Eq. for S_profit] The conditional profit scoring function is degenerate as written. The text defines P_m = max(0, P_h); if P_h is positive, then P_m = P_h, so the interval P_m ≤ P_a < P_h is empty and the second branch has denominator P_h − P_m = 0. This makes the formula ill-defined for the most common case of a positive human benchmark profit. The authors must correct the definition of P_m (e.g., P_m = 0 or P_h/2) and re-derive all outcome scores. As it stands, Table 2's outcome scores cannot be reproduced.
  3. [§2.2 vs. Figure 1 caption] The live trading results are internally inconsistent. Section 2.2 reports DeepSeek-V3's final portfolio value as $11,420.3 (+14.2%), but the Figure 1 caption reports $16,362.1 for the same model; GPT-5 is $9,901.1 in the text but $10,268.8 in the caption; Claude-Sonnet-4.5 is reported as $10,072.8 in the text but $12,582.8 in the caption. Additionally, Section 2.2 defines a 7-week window (Nov 19, 2025–Jan 9, 2026), while the Figure 1 x-axis extends from Nov 19 to May 24, 2026. These contradictions undermine the empirical basis of the "capabilities-performance paradox" that motivates the entire benchmark.
  4. [§1, Contribution 3 vs. Table 2] The headline claim of "only 1.8/5 reasoning score" is not supported by Table 2. Table 2 reports Event Coherence between 0.8 and 2.8 and Logical Plausibility between 3.9 and 4.2; the reasoning-quality average (SEC + S_LP)/2 ranges from 2.5 to 3.45 across the four models. The value 1.8 is GPT-5's event-coherence subscore, not a reasoning score. The abstract's separate characterization (logical plausibility near 4/5 and event grounding 0.8–2.8/5) is accurate, but the contribution statement overstates the gap. The authors should consistently report the composite reasoning score or clearly label the subscore.
  5. [§2.4 and §4.1/Figure 4] The "human expert logic anchors" are not expert decision traces. Section 2.4 emulates Bill Miller and Cathie Wood through their firms' ETFs (Miller Value Partners Appreciation ETF and ARK Innovation ETF), yet the text attributes the ETFs' returns to expert "counter-cyclical aggression" and "variant perception" without evidence that those specific investment decisions generated the returns. Furthermore, the case study in Section 4.1 says the scenario concerns a semiconductor rally, but Figure 4 shows Cathie Wood's decision is to buy GLD/TLT (gold and Treasury ETFs), and the table's outcomes are inconsistent with the surrounding text. These examples do not support the claimed diagnostic insights.
minor comments (6)
  1. [Title] The title contains a typo: "Personalzied" should be "Personalized."
  2. [§3.1 / Appendix A.1] The paper states the dataset contains 201,247 decisions from 151 experts, but Table 3 lists only 43 KOLs and Table 1 reports an average of 1332.8 decisions per KOL. The authors should clarify how the 151 experts were sampled and how the remaining expert types (fund managers, etc.) map to the 201,247 decisions.
  3. [§4.0.1, Table 2] To reproduce the outcome scores, the paper needs to report the human benchmark reference values P_h and W_h for each task; Table 2 only gives the models' outcomes. Without these values, the reported "Overall Score /10" cannot be independently verified.
  4. [§2.2] The "Foundation Aptitude Score" is described as an average of MMLU-Pro, SWE-bench, AIME 2025, and SimpleQA, but no details are given on normalization, task versions, or model-specific subscores. This makes the right-hand panel of Figure 1 difficult to interpret.
  5. [Figure 1] The left panel lacks labeled legend entries for individual curves; only a partial list of models is shown in the text. The x-axis extends well beyond the declared seven-week window, so the figure should either be regenerated with the correct period or the window definition should be revised.
  6. [§3.5] The automated judge is a high-capacity LLM (e.g., GPT-5), which is also one of the evaluated models in Table 2. Even if this is not circular in a strict sense, it introduces a potential conflict; the authors should report the judge prompt, judge-model version, and agreement with human raters on a held-out set.

Circularity Check

2 steps flagged · score 7.0 of 10

The benchmark's 'expert logic' ground truth is partly generated by the authors' proprietary SIL Engine and scored by GPT-5, which is itself an evaluated model, making the central reasoning-gap result self-referential.

  1. self definitional [Section 3.1 (Agent Logic Processing) and Section 3.5 (Fidelity scoring)]
    "using our proprietary Stock Investment Logic (SIL) Engine to infer implicit logic patterns from an agent’s trading history, constructing a holistic behavioral profile. ... Fidelity: Similarity to ground-truth agent logic (Task 1)."

    The benchmark's ground-truth reasoning traces are, per Section 3.1, partially produced by the authors' own SIL Engine from trading histories when explicit theses are absent. Section 3.5 then uses those same traces as the reference for Task 1's Fidelity score. Consequently, the reported 'reasoning score against expert traces' (e.g., the 1.8/5 event-coherence headline in Contribution 3) measures agreement with SIL-inferred logic, not with independently documented investor cognition. The central external quantity—expert investment logic—is defined internally by the authors' pipeline, so the benchmark's process-diagnostic results reduce to a self-comparison without external validation of SIL's reconstructions.

  2. other [Section 3.5 (Automated Evaluation via Judge LLM) and Table 2]
    "A high-capacity LLM (e.g., GPT-5) serves as an automated judge, scoring outputs based on: -Fidelity: Similarity to ground-truth agent logic (Task 1). -Coherence & Plausibility: Logical soundness and market relevance (Tasks 2, 3)."

    GPT-5 is both the automated judge described in Section 3.5 and an evaluated model listed in Table 2. Therefore GPT-5's reported Event Coherence (1.8) and Logical Plausibility (4.1) are self-assessed by the same model family, with no independent judge or human adjudication. This makes the measured reasoning quality of a headline model partially self-defined, conflating the evaluator and the evaluatee.

full rationale

The paper's most distinctive contribution is a process-diagnostic benchmark that quantifies an 'investment logic gap' by comparing LLM reasoning to expert traces. The circularity arises because those expert traces are, in part, constructed by the authors' proprietary SIL Engine from trading histories, and then the same traces are used as the gold standard for Fidelity scoring. Thus the claimed external benchmark (real investor cognition) is partly internal output of the authors' pipeline, and the headline 1.8/5 reasoning score reduces to agreement with SIL-inferred logic. A second, narrower circularity is that GPT-5 serves as the automated judge while also being one of the evaluated models, so its own scores are self-assessed. The live-trading return data and market outcomes are externally grounded, but the paper's central process-level diagnosis—the reasoning gap—is not independently anchored. No human annotation, inter-annotator agreement, or external validation of SIL's reconstructions is provided, so the benchmark's core measurements are self-referential as presented. If the SIL-inferred traces were validated against verbatim investor reasoning or independent annotations, the circularity would be substantially reduced; as reported, the central claim is not fully independent of the authors' own framework.

Assumptions & free parameters 3 free parameters · 4 assumptions · 3 invented entities

The central claim depends on the SIL Engine's inferred reasoning being a faithful stand-in for real investor cognition, on the source videos capturing true reasoning, on the GPT-5 judge matching human judgment, and on look-ahead-free outcomes. None of these is validated. The scoring weights and foundation composite are hand-chosen, and the proprietary SIL Engine blocks external audit.

free parameters (3)
  • Decision outcome weights = 0.7 profit, 0.3 win rate
    Section 4.0.1: Soutcome = 0.7 × Sprofit + 0.3 × Swin. The weights are asserted by hand, not derived or fitted to data.
  • Neutral profit midpoint P_m = max(0, P_h)
    Section 4.0.1 defines P_m = max(0, P_h) as the neutral midpoint. This anchor determines the shape of the profit score and is chosen by hand.
  • Foundation Aptitude Score = unweighted average of MMLU-Pro, SWE-bench, AIME 2025, SimpleQA
    Section 2.2: the composite used to demonstrate the paradox is an average of four selected benchmarks; the selection and equal weighting are author choices that drive the paradox plot.
assumptions (4)
  • domain assumption Articulated YouTube commentary and fund reports faithfully represent the investor's actual reasoning.
    Appendix A.1 and Section 3.1: the benchmark is built from KOL videos and fund reports, assuming these verbalizations capture true decision logic.
  • ad hoc to paper The SIL Engine's inferred implicit logic is a valid reconstruction of investor reasoning.
    Section 3.1: 'infer implicit logic patterns from an agent's trading history.' No validation against the investors themselves or external judges.
  • domain assumption LLM judge scores correspond to human judgment of reasoning quality.
    Section 3.5: GPT-5 serves as automated judge; no human agreement study is reported.
  • domain assumption Market-validated outcomes are correctly point-in-time bound and look-ahead free.
    Section 3.2: outcomes are tied to dates, but the paper does not demonstrate that decision-time information excluded future data.
invented entities (3)
  • P→E→R→D→O investment logic chain
    purpose: Organizing schema for each investment decision in the benchmark
    Section 1: introduced as the organizing construct; closely mirrors the cited Chain-of-Decision framework and is not independently verified as a faithful model of cognition.
  • Stock Investment Logic (SIL) Engine
    purpose: Infers implicit investment logic from trading history to construct ground-truth traces
    Section 3.1: proprietary, not released, and unvalidated; the benchmark's ground truth depends on it.
  • InvestLogicBench2026 dataset independent evidence
    purpose: Benchmark of 201,247 decisions for evaluating LLM investment logic
    Section 3.3: released via anonymous link with market-verified outcomes, so outcome labels are externally checkable in principle, though the reasoning traces are not.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Evaluating Investment Logic in Large Language Models: A Real-World Benchmark Towards Personalzied Financial Agents." pith.science (2026). https://pith.science/paper/LTDX4EYA

@misc{pith2026260806108,
  author       = {Pith},
  title        = {Pith review of: Evaluating Investment Logic in Large Language Models: A Real-World Benchmark Towards Personalzied Financial Agents},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LTDX4EYA}},
  note         = {Machine review of arXiv:2608.06108}
}
abstract

Investment competence is inherently personalized: the same market evidence can justify different actions for investors with different goals, horizons, portfolios, and risk boundaries. Yet financial LLMs are evaluated either by static question answering or by terminal profit and loss. The former omits agency; the latter cannot reveal whether a profitable action was grounded, profile-consistent, or merely lucky. We ask whether the community is using the wrong ruler for consequential agents. We introduce \textsc{InvestLogicBench}, a process-native benchmark containing 201,247 documented decisions from 151 real-world investors. Each episode instantiates a \textbf{P$\rightarrow$E$\rightarrow$R$\rightarrow$D$\rightarrow$O} trace: investor \textit{Profile}, observable market \textit{Events}, investment \textit{Reasoning}, executable \textit{Decision}, and delayed \textit{Outcome}. The release includes profile construction, point-in-time event binding, structured logic, horizons, outcomes, and post-mortems, and supports comprehension, profile-conditioned generation, and end-to-end replay. Across four leading LLMs, logical plausibility remains near 4/5 while event grounding is only 0.8--2.8/5; return and process quality also disagree. These results expose polished but weakly grounded reasoning that outcome-only evaluation hides. We further argue that P$\rightarrow$E$\rightarrow$R$\rightarrow$D$\rightarrow$O should be a data-system interface, requiring versioned profiles, temporal provenance, inspectable retrieval, decision ledgers, and replayable outcomes. Finance is our stress test for a broader class of personalized, consequential agents.

Figures

Figures reproduced from arXiv: 2608.06108 by the authors.

Figure 1
Figure 1. The “Capabilities-Performance Paradox” in financial domain. (Left) Cumulative equity curves between autonomous LLM agents and human expert proxies in a 7-week live U.S. stock trading competition. DeepSeek-V3 achieved the highest cumulative return, significantly outperforming the S&P 500 benchmark, while models like GPT-5 and Claude-Sonnet-4.5 demonstrated lackluster performance in seizing opportunities. Key market i… view at source ↗
Figure 2
Figure 2. The two-stage construction pipeline of INVESTLOGICBENCH2026. Stage 1 creates a static investor profile by aggregating multiple data points from a single agent. Stage 2 performs per-decision annotation, which includes a quality control loop where logic is verified and potentially fed back to refine the profile. • finLogic: The complete E→R→D chain (e.g., ’[Event] → [Thesis] → Action: [Action]’). • investDecision: The… view at source ↗
Figure 3
Figure 3. A working example from the Stock Investment Logic (SIL) Engine. It demonstrates how a specific market event (here, an antitrust probe announcement by China’s SAMR targeting Nvidia on 2025-09-15) is temporally linked to a security (NVDA) and its subsequent price series. The engine structures and stores such Event-Security-Time tuples to form a queryable knowledge base for inferring recurring investment logic patterns… view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: illustrates a simulated event-driven investment sce￾nario where Cathie Wood correctly predicted a rally in semi￾conductor stocks by integrating nuanced observations from multiple data sources. In contrast, LLMs exhibit systematic reasoning failures when provided the sa…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

21 extracted references · 19 canonical work pages

  1. [1]

    arXiv preprint arXiv:2408.06361 , year=

    Large Language Model Agent in Financial Trading: A Survey , author=. arXiv preprint arXiv:2408.06361 , year=

  2. [2]

    Chen, Zhiyu and Chen, Wenhu and Smiley, Charese and Shah, Sameena and Borova, Iana and Langdon, Dylan and Moussa, Reema and Beane, Matt and Huang, Ting-Hao and Routledge, Bryan and Wang, William Yang , journal=. Fin. 2022 , url=

  3. [3]

    2021 , url=

    Zhu, Fengbin and Lei, Wenqiang and Huang, Youcheng and Wang, Chao and Zhang, Shuo and Lv, Jiancheng and Feng, Fuli and Chua, Tat-Seng , journal=. 2021 , url=

  4. [4]

    Islam, Pranab and Kannappan, Anand and Kiela, Douwe and Qian, Rebecca and Scherrer, Nino and Vidgen, Bertie , journal=. Finance. 2023 , url=

  5. [5]

    2024 , url=

    Koncel-Kedziorski, Rik and Krumdick, Michael and Lai, Viet and Reddy, Varshini and Lovering, Charles and Tanner, Chris , journal=. 2024 , url=

  6. [6]

    Advances in Neural Information Processing Systems 38 (NeurIPS 2024) , year=

    FinBen: A Holistic Financial Benchmark for Large Language Models , author=. Advances in Neural Information Processing Systems 38 (NeurIPS 2024) , year=

  7. [7]

    Liu, Shu and Zhao, Shangqing and Jia, Chenghao and Zhuang, Xinlin and Long, Zhaoguang and Zhou, Jie and Zhou, Aimin and Lan, Man and Chong, Yang , booktitle=. Fin. 2024 , url=

  8. [8]

    2025 , url=

    Zhang, Zhihan and Cao, Yixin and Liao, Lizi , journal=. 2025 , url=

Show all 21 references
  1. [9]

    2023 , url=

    Wu, Shijie and Irsoy, Ozan and Lu, Steven and Dabravolski, Vadim and Dredze, Mark and Gehrmann, Sebastian and Kambadur, Prabhanjan and Rosenberg, David and Mann, Gideon , journal=. 2023 , url=

  2. [10]

    Yang, Hongyang and Liu, Xiao-Yang and Wang, Christina Dan , journal=. Fin. 2023 , url=

  3. [11]

    2025 , url=

    Fan, Tianyu and Yang, Yuhao and Jiang, Yangqin and Zhang, Yifei and Chen, Yuxuan and Huang, Chao , journal=. 2025 , url=

  4. [12]

    Li, Changlun and Shi, Yao and Luo, Yuyu and Tang, Nan , journal=. Will. 2025 , url=

  5. [13]

    arXiv preprint arXiv:2410.07225 , year=

    Distilling Analysis from Generative Models for Investment Decisions , author=. arXiv preprint arXiv:2410.07225 , year=

  6. [14]

    Sehgal, Neil K. R. and Guntuku, Sharath Chandra and Ungar, Lyle , journal=. Real-Time Deadlines Reveal Temporal Awareness Failures in. 2026 , url=

  7. [15]

    and Cui, Zhenyu and Liu, Rong and Xu, Zhaozhuo and Subbalakshmi, Koduvayur and Xiong, Guojun and He, Yueru and Huang, Jimin and Li, Dong and Xie, Qianqian , booktitle=

    Yu, Yangyang and Yao, Zhiyuan and Li, Haohang and Deng, Zhiyang and Jiang, Yuechen and Cao, Yupeng and Chen, Zhi and Suchow, Jordan W. and Cui, Zhenyu and Liu, Rong and Xu, Zhaozhuo and Subbalakshmi, Koduvayur and Xiong, Guojun and He, Yueru and Huang, Jimin and Li, Dong and X...

  8. [16]

    Generative

    Alaminos, David and Guill. Generative. International Review of Economics & Finance , year=

  9. [17]

    Klimaszewski, Mateusz and Chen, Pinzhen and Guillou, Liane and Papaioannou, Ioannis and Haddow, Barry and Birch, Alexandra , booktitle=. Aveni. 2025 , url=

  10. [18]

    2025 , url=

    Iacovides, Giorgos and Zhou, Wuyang and Mandic, Danilo , journal=. 2025 , url=

  11. [19]

    Extracting Structured Insights from Financial News: An Augmented

    Dolphin, Rian and Dursun, Joe and Chow, Jonathan and Blankenship, Jarrett and Adams, Katie and Pike, Quinton , journal=. Extracting Structured Insights from Financial News: An Augmented. 2024 , url=

  12. [20]

    2025 , url=

    Fatouros, George and Metaxas, Kostas and Soldatos, John and Karathanassis, Manos , journal=. 2025 , url=

  13. [21]

    arXiv preprint arXiv:2312.11970 , year=

    Large Language Models Empowered Agent-based Modeling and Simulation: A Survey and Perspectives , author=. arXiv preprint arXiv:2312.11970 , year=

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.