Paper: arXiv 2512.01354

Authors: Zhongjie Jiang

Abstract

Although synthetic data is widely promoted as a remedy, its prevailing production paradigm – one optimizing for statistical smoothness – systematically removes the long-tail, cognitively grounded irregularities that characterize human text. Prolonged training on such statistically optimal but cognitively impoverished data accelerates model collapse. This paper proposes a paradigm shift: instead of imitating the surface properties of data, we simulate the cognitive processes that generate human text. We introduce the Prompt-driven Cognitive Computing Framework (PMCSF), whose core consists of a Cognitive State Decoder (CSD) that reverse-engineers unstructured text into structured cognitive vectors, and a Cognitive Text Encoder (CTE) that re-materializes these states into text enriched with human-typical imperfections via mathematically defined Cognitive Perturbation Operators. The framework is validated through a two-stage objective evaluation pipeline. First, in cognitive codec verification, CTE text yields a Jensen-Shannon divergence of 0.0614 from human text (vs. 0.4431 for standard LLM output), passes double-blind professional media review, and achieves an intraclass correlation coefficient ICC > 0.9 for cognitive profile alignment across heterogeneous models. Second, in functional gain evaluation, isomorphic stress tests in the A-share market show that strategies incorporating CTE-generated data reduce maximum drawdown by 47.4% during the 2015 crash and deliver 8.6% Defensive Alpha, exceeding transaction costs by a factor of 33. Our findings demonstrate that modelling human cognitive limitations – not copying surface data – enables synthetic data with genuine functional gain, offering a viable technical pathway toward resolving the AI data-collapse crisis.

Complexity vs Empirical Score

  • Math Complexity: 7.0/10
  • Empirical Rigor: 8.5/10
  • Quadrant: Holy Grail — high math complexity, high empirical rigor

Why this score: The paper employs advanced mathematical concepts including a 17-dimensional vector space, Jensen–Shannon divergence, intraclass correlation coefficients, and a ‘Cognitive Perturbation Operators’ framework. It demonstrates high empirical rigor through specific quantitative results (e.g., 47.4% drawdown reduction, 8.6% Defensive Alpha) in a real-world market stress test (A-share market 2015 crash) with validated data pipelines.

Research Flowchart

  flowchart TD
  A["Research Goal:<br>Reverse Model Collapse via<br>Simulating Cognitive Boundedness"] --> B["Methodology:<br>Prompt-driven Cognitive Computing Framework PMCSF"]

  subgraph B_Method ["PMCSF Core Mechanisms"]
      B1["Cognitive State Decoder CSD<br>Reverse-engineers text → cognitive vectors"] --> B2["Cognitive Perturbation Operators<br>Mathematically defines human imperfections"]
      B2 --> B3["Cognitive Text Encoder CTE<br>Re-materializes states → enriched text"]
  end

  B --> C["Input Data:<br>Unstructured Human Text"]

  subgraph D_Process ["Two-Stage Evaluation Pipeline"]
      direction LR
      D1["Stage 1: Cognitive Codec Verification<br>Jensen-Shannon Divergence: 0.0614<br>ICC > 0.9"] --> D2["Stage 2: Functional Gain Evaluation<br>A-share Market Stress Test"]
  end

  B3 --> D_Process

  E["Key Findings & Outcomes"] --> F["Technical Pathway:<br>Modeling limitations enables<br>functional synthetic data"]
  E --> G["Performance Gains:<br>- 47.4% Max Drawdown Reduction<br>- 8.6% Defensive Alpha<br>- Exceeds transaction costs 33x"]

  D_Process --> E

Evidence Card

What the authors claim, on what data, and who has checked it. Read from the paper's own text by a language model: every quote is verbatim, nothing was re-run. How to read this card

Proposes

The paper proposes the Prompt-driven Cognitive Computing Framework (PMCSF) with a Cognitive State Decoder and a Cognitive Text Encoder to simulate human cognitive imperfections for synthetic data generation.

Source sentence
We introduce the Prompt-driven Cognitive Computing Framework (PMCSF), whose core consists of a Cognitive State Decoder (CSD) that reverse-engineers unstructured text into structured cognitive vectors, and a Cognitive Text Encoder (CTE) that re-materializes these states into text enriched with human-typical imperfection

Assumes 6

  • A flat transaction friction of 0.26% per trade (including commissions and slippage) is applied to mimic real‑world execution costs. execution
    Source sentence
    Transaction Friction: A total cost of 0.26% per trade—encompassing bilateral commissions and impact slippage—was established to replicate real-world execution costs.
  • An annualized risk‑free rate of 2% (daily 0.008%) is used for Sharpe ratio calculations. market
    Source sentence
    Risk Benchmark: An annualized risk-free rate of 2% (daily 0.008%) was designated for Sharpe Ratio calculation, aligning with industry-standard risk metrics.
  • The experimental design assumes non‑stationarity of financial markets and therefore abandons large‑sample statistical fitting in favor of prototype verification. statistical
    Source sentence
    the experimental design abandons the traditional quantitative research approach of “large-sample statistical fitting” in favor of a prototype verification strategy prioritizing “causality over correlation.”
  • Anti‑Data Leakage principle: all specific historical facts and objective market summaries are stripped from the training data. data
    Source sentence
    Rigorous adherence to the Anti-Data Leakage Principle governs the design of the PMCSF. During construction, the system strips away all specific historical facts and objective market summaries that could induce bias, retaining only universal commentary information.
Show 2 more
  • Scenario Anchors (2015 crash and 2024 bull market) are assumed to have the highest signal‑to‑noise ratios and most pronounced irrational characteristics for testing. market
    Source sentence
    selecting “Scenario Anchors” from A-share market history—specifically the 2015 stock market crash and the 2024 bull market—which exhibit the highest signal-to-noise ratios and most pronounced irrational characteristics as experimental samples.
  • The CTE‑Enhanced group’s performance differences are attributed solely to the data, holding strategy logic constant. other
    Source sentence
    holding the M-Dynamic (dynamic adaptive) strategy logic constant, simulated trading performance was compared between the CTE-Enhanced (cognitive enhancement group) and Human-Baseline (human baseline group) using distinct input data streams.

Data

Backtest sample: A-share market · equities

  • CTE‑A (Emotional Catharsis Mode) synthetic unknown
    2015 crash and 2024 bull market · A‑share market scenario anchors
    Source sentence
    were designated as CTE-A (Emotional Catharsis Mode), CTE-B (In-depth Analysis Mode), and CTE-C (Macro-Chaotic Mode).
  • CTE‑B (In-depth Analysis Mode) synthetic unknown
    2015 crash and 2024 bull market · A‑share market scenario anchors
    Source sentence
    were designated as CTE-A (Emotional Catharsis Mode), CTE-B (In-depth Analysis Mode), and CTE-C (Macro-Chaotic Mode).
  • CTE‑C (Macro‑Chaotic Mode) synthetic unknown
    2015 crash and 2024 bull market · A‑share market scenario anchors
    Source sentence
    were designated as CTE-A (Emotional Catharsis Mode), CTE-B (In-depth Analysis Mode), and CTE-C (Macro-Chaotic Mode).
  • Standard‑AI Group text-news unknown
    2015 crash and 2024 bull market · A‑share market scenario anchors
    Source sentence
    The total sample comprises 461 sentences (Standard Group: 155; CTE Groups: 113/90/103 respectively)

Costs

Modeled Costs are part of the reported headline results.

0.26% per trade—including bilateral commissions and impact slippage.

Source sentence
Transaction Friction: A total cost of 0.26% per trade—encompassing bilateral commissions and impact slippage—was established to replicate real-world execution costs.

Finds 5

Reported performance (CTE-Enhanced strategy): Sharpe 0.25 · max drawdown 12.2%

  • CTE‑generated text achieves a Jensen‑Shannon divergence of 0.0614 from human text, far lower than 0.4431 for standard LLM output. statistical
    Source sentence
    CTE text yields a Jensen‑Shannon divergence of 0.0614 from human text (vs. 0.4431 for standard LLM output)
  • In the 2015 bear‑market test, the CTE‑Enhanced group’s maximum drawdown is 12.2%, a 47.4% reduction versus the Human‑Baseline’s 23.2%. performance
    Source sentence
    ... CTE‑Enhanced group’s maximum drawdown stood at 12.2%, a figure 47.4% lower than the Human‑Baseline group’s 23.2%.
  • The CTE‑Enhanced group generates 8.6% Defensive Alpha, equivalent to a safety buffer 33 times the 0.26% transaction cost per trade. performance
    Source sentence
    Backtest data delineates that the CTE‑Enhanced group generated 8.6% Defensive Alpha
  • During the 2024 bull‑market test, the CTE‑Enhanced group achieves a net return of +17.92%, more than double the Human‑Baseline’s +8.04%. performance
    Source sentence
    ... achieved a net return of +17.92%, 2.2 times that of the Human‑Baseline group’s +8.04%.
  • Expert review pass rate for CTE‑generated submissions is 72.7%, significantly higher than the 13.1% rate of the human control group (p<0.01). performance
    Source sentence
    CTE-generated submissions achieved a 72.7% expert review pass rate—significantly outperforming the 13.1% rate of the human control group (𝑝< 0.01).

Checked by others

No other paper in the archive has been found to support or contradict these findings (compared with 26 related papers). Coverage is partial, so this is neither replication nor refutation.

Code

No repository by the authors found in the paper.

Extracted 2026-11-14 · same card as JSON · A language model chose these items: check the source sentence before relying on one.

Cited References (30)

  1. Distinguishing AI-Generated and Human-Written Text Through Psycholinguistic Analysis (2025) — link
  2. Self-Consuming Generative Models Go MAD (2023) — link
  3. The Curse of Recursion: Training on Generated Data Makes Models Forget (2023) — link
  4. Generative Agents: Interactive Simulacra of Human Behavior (2023) — link
  5. Out of One, Many: Using Language Models to Simulate Human Samples (2022) — link
  6. Human heuristics for AI-generated language are flawed (2022) — link
  7. Advances in Financial Machine Learning: Lecture 10/10 (Presentation Slides) (2018) — link
  8. Heuristic decision making. (2011) — link
and 22 more
  1. The Way We Think: Conceptual Blending and the Mind’’s Hidden Complexities. Basic Books (2002) — link
  2. Advances in prospect theory: Cumulative representation of uncertainty (1992) — link
  3. An argument for basic emotions (1992) — link
  4. Cognitive Load During Problem Solving: Effects on Learning (1988) — link
  5. Working memory (1984) — link
  6. A circumplex model of affect. (1980) — link
  7. Judgment under Uncertainty: Heuristics and Biases (1974) — link
  8. Rational choice and the structure of the environment. (1956) — link
  9. Administrative Behavior: A Study of Decision Making Processes in Administrative Organization (1947) — link
  10. The Rise of ‘Pink Slime’ Journalism: 1,265 U.S. Fake Local News Sites and the Foreign Networks Behind Many of Them (2024)
  11. Will we run out of data? An analysis of the limits of scaling datasets in Machine Learning (2022)
  12. … and the Cross-Section of Expected Returns (2013) — link
  13. Prospect Theory : An Analysis of Decision under Risk Author ( s ) : (2007)
  14. Rhythms of the brain (2006) — link
  15. Variation across Speech and Writing (1988)
  16. Aspects of the Theory of Syn-tax (1965)
  17. Stage 4 (Core Feature Verification
  18. “Universal Map of Thought”: implies that Fear and Uncertainty are always neighbors
  19. Perturbed_Content = Apply_Sentence_Oscillation(Raw_Content , Physics_Params
  20. Simulate rhythmic variation
  21. Simulate associative leaps (heuristic -driven semantic jumps) if probability threshold is met
  22. A “Compression Tool for Thought”: like MP3 for audio, it compresses thoughts into vectors and plays them back as text

0 of 30 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.