Cognitive Text Encoder (CTE)

Updated 5 December 2025

CTE is a neural model that encodes text by aligning its representations with cognitive signals such as EEG, eye-tracking, and fMRI.
It utilizes dual-layer simulation and adversarial training to integrate structured cognitive vectors into text generation for improved human-like nuances.
Empirical results show that CTEs achieve higher fidelity in reproducing human textual characteristics and support applications in cognitive simulation and brain–language interfaces.

A Cognitive Text Encoder (CTE) is a neural model designed to encode or generate text representations that are explicitly aligned with, derived from, or enriched by human cognitive processes and signals. This paradigm aims to transcend surface-level statistical imitation, instead mapping between text and structured intermediate representations grounded in cognitive states, neurophysiology, or behavioral patterns. CTEs are instantiated in various settings: as invertible codecs within cognitive simulation frameworks, as encoders trained to align with brain or behavioral measurements, and as modules that reconstruct text from brain signals. This article surveys the principal architectures, mathematical formulations, algorithmic details, and empirical results underlying several prominent lines of CTE research.

1. Conceptual Role and Definitions

The CTE framework arises in contexts where the goal is not merely to maximize statistical fluency, but to reinject or decode cognitively grounded details absent from conventional text generators. In the Prompt-driven Cognitive Computing Framework (PMCSF), the CTE is paired with a Cognitive State Decoder (CSD) to form a “cognitive codec” (Jiang, 1 Dec 2025). The CSD reads natural language text and extracts a structured, 17-dimensional cognitive state vector, while the CTE re-materializes this vector into natural language—explicitly preserving long-tail, irregular human-typical imperfections via a layered simulation process. In this architecture, the CTE is the generative or "encoding" step that ensures text retains traceable signatures of human bounded rationality.

In neural alignment approaches, such as those in “CogAlign” (Ren et al., 2021), the CTE refers to a text encoder whose hidden representations are aligned with human cognitive processing signals (e.g., eye-tracking, EEG). These CTEs ingest pure text at inference, but their internal states are regularized during training to mirror those observed in cognitive modalities. In brain decoding, as exemplified by UniCoRN (Xi et al., 2023), the CTE is the model component mapping raw neural time series (fMRI, EEG) into latent embeddings, which condition a downstream LLM to reconstruct stimulus text.

2. Representative Architectures and Operational Principles

PMCSF: Dual-Layer Simulation Pipeline

In PMCSF (Jiang, 1 Dec 2025), the CTE is architected atop a base LLM, structured into two simulation layers:

Macro Layer (“Satisficing Search”): This layer anchors text generation using the supplied 17-dimensional cognitive state vector $\mathbf z \in \mathbb{R}^{17}$ , alongside a prompt template. The module internally selects "cognitive priors" (e.g., dominant emotional or bias dimensions), segmenting reasoning into roles—Architect (planning), Narrator (surface realization), Punchline (injecting surprising elements).
Micro Layer (Cognitive Perturbation Operators): This stage applies computational perturbations to base model output, including:
- Sentence Length Oscillation: stochastic variation according to $L_s(n) = \lfloor L + A\sin(\omega n + \phi) + \epsilon \rfloor$ , where $L$ is baseline length, $A$ amplitude, $\omega$ frequency, and $\epsilon \sim \mathcal{N}(0,\sigma^2)$ .
- Token Probability Perturbation: implements hesitancy or suboptimal word choice using temperature and bias masks on the LLM’s next-token distribution: $P'(w_t|w_{<t}) \propto [P(w_t|w_{<t})]^{1/\tau(\mathbf z)} \times M_\text{bias}(w_t;\mathbf z)$ .
- Semantic Leap: enforces occasional topic jumps by constraining lexical similarity at paragraph boundaries: $\cos(E(w), C_\text{prev}) < \theta_\text{leap}$ .
- Final logical verification can inject “flaws” to break over-smoothed output.

CogAlign: Shared Representation Alignment

CogAlign (Ren et al., 2021) employs a tripartite Bi-LSTM network, where two private encoders process text and cognitive modality inputs respectively, and a single shared encoder is adversarially trained to form modality-agnostic representations. Modality discrimination is penalized via a gradient reversal layer; a text-aware attention module attenuates noise in cognitive features. The shared encoder is operationalized as the Cognitive Text Encoder—its output hidden states, even on pure text, inhabit a manifold aligned with eye-tracking and EEG-derived spaces.

UniCoRN: Neural Signal to Text Generation

In UniCoRN (Xi et al., 2023), the CTE is the composite encoder that transforms multivariate brain signals to latent embeddings fed into a pretrained text decoder (BART). This encoder is split into snapshot (e.g., 3D-CNN for fMRI, Transformer for EEG) and temporal sequence (stacked Transformer) modules, coupled by linear projection. The decoded text is thus conditioned strictly on cognitive signals, with no explicit use of textual context during encoding.

3. Mathematical Formulations

Key CTE mechanisms are mathematically grounded as follows:

Model	Core Cognitive Representation	Primary Loss Functions
PMCSF	$\mathbf z \in \mathbb{R}^{17}$	$\mathcal{L}_\text{CE}$ (recon), $\mathcal{L}_\text{cog}$ (vec)
CogAlign	Hidden states $\mathbf{h}^S_i$ aligned to cognitive signals	$L_\text{task}$ , $L_\text{adv}$ (adversarial)
UniCoRN	Embeddings $E^e_t$ from neural signals	$\mathcal{L}_\text{snap}$ , $\mathcal{L}_\text{series}$ , $\mathcal{L}_\text{CE}$

PMCSF: The combination loss is

$\mathcal{L} = \mathcal{L}_{\text{CE}} + \lambda \mathcal{L}_{\text{cog}}, \quad \lambda > 0$

where $\mathcal{L}_{\text{CE}} = -\sum_{t=1}^n \log P'(x_t|x_{<t},\mathbf z)$ and $\mathcal{L}_{\text{cog}} = \|\mathrm{CSD}(x) - \mathbf z\|^2$ .

CogAlign: The adversarial loss on the shared encoder employs gradient reversal:

$L = L_{\text{task}} + \lambda L_{\text{adv}}$

with $L_{\text{adv}}$ being the cross-entropy that penalizes success at modality discrimination.

UniCoRN: Multiphase objectives:

$\begin{align*} \mathcal{L}_\text{snap} &= \frac{1}{N} \sum_{k=1}^N \|\mathcal{D}_r(\mathcal{E}_r(f_k)) - f_k\|_1 \ \mathcal{L}_\text{series} &= \frac{1}{M} \sum_{j=1}^M \|\mathcal{D}_r(\mathcal{E}_s(E^i_{j\sim j+\mathcal T-1})) - E^i_{j\sim j+\mathcal T-1}\|_1 \ \mathcal{L}_\text{CE} &= -\sum_{t=1}^n \log p(w_t|w_{<t}, E_{k\sim k+\mathcal T-1}) \end{align*}$

4. Algorithmic Details and Pseudocode

The CTE simulation pipeline in PMCSF can be abstracted as follows (Jiang, 1 Dec 2025):

function CognitiveTextEncoder(z, PromptTemplate):
    MacroAnchoringPhase:
        Extract reasoning skeleton from PromptTemplate
        Assign operator: {Architect, Narrator, Punchline}
    MicroPerturbationPhase:
        FinalText = ""
        for segment in skeleton:
            BaseGeneration = LLM.generate(segment, operator)
            Perturbed1 = ApplySentenceOscillation(BaseGeneration, L_s)
            if random() < AssociativeLeapProb(z):
                Perturbed2 = ApplySemanticLeap(Perturbed1)
            else:
                Perturbed2 = Perturbed1
            Perturbed3 = ApplyTokenPerturbation(Perturbed2, τ(z), M_bias)
            if not LogicalToleranceCheck(Perturbed3):
                Perturbed3 = InjectHumanFlaw(Perturbed3)
            FinalText += Perturbed3
    return FinalText

This structure enforces a strict sequence of macro-level skeleton generation and micro-level stochastic perturbation to achieve cognitive-fidelity text.

In CogAlign, the CTE is realized as the shared encoder in a bi-modal framework, with modality alignment enforced during alternating input of textual and cognitive signals (Ren et al., 2021). In UniCoRN, the CTE’s encoder is subject to explicit stepwise spatial and temporal reconstruction, before final alignment with a generative LLM (Xi et al., 2023).

5. Empirical Results and Evaluation

Empirical results across CTE instantiations demonstrate significant gains in text-cognitive alignment, statistical irregularity, and functional outcomes.

PMCSF CTE (Jiang, 1 Dec 2025):
- Achieves Jensen–Shannon divergence of 0.0614 from human text (vs. 0.4431 for standard LLMs).
- High fidelity for style metrics: sentence-length SD, adjective density, noun-verb ratio, etc.
- Micro-statistical volatility: CTE-A exhibits coefficient of variation ≈58.69% (rejects normality), approaching human-like distribution skew.
- Cross-model consistency (CTE+CSD invariants): ICCs > 0.9 across model families.
- Downstream impact: strategies using CTE data yield 47.4% lower max drawdown and 8.6% Defensive Alpha in A-share market simulations.
CogAlign (Ren et al., 2021):
- On ZuCo corpus, CTE-derived encodings grant +0.48 F1 (NER), +2.17 F1 (sentiment), +0.81 F1 (relation extraction) compared to best baselines.
- Transfer learning: improvements of ≈1–2% F1 even on datasets without cognitive annotations at inference.
UniCoRN (Xi et al., 2023):
- BLEU-4 = 34.77% (fMRI2text, random-time), 37.78% (by-subject split); ROUGE-1 F1 up to 59.52%.
- BLEU-4 = 37.04% (EEG2text).
- The CTE generalizes across modalities (fMRI, EEG) and provides open-vocabulary decoding.

6. Modalities, Cognitive Vectors, and Preprocessing

CTEs leverage diverse input modalities:

Structured Cognitive Vectors (PMCSF): $\mathbf z$ partitions into 8 basic emotions, 4 regulatory/arousal variables, 5 domain-specialized states (e.g., FOMO, Regret).
Behavioral/Neurophysiological signals (CogAlign, UniCoRN): Eye-tracking (FFD, FPD, TFD, re-fixations), EEG (averaged spectral bands), and fMRI (preprocessed volumetric time series).
Preprocessing: ZuCo corpus for eye/EEG, Narratives pipeline for fMRI (realignment, normalization, z-scoring).

The CTE in each paradigm is designed to inject, reconstruct, or align to these signals at either encoding or decoding, often via explicit loss terms and model modules.

7. Limitations and Open Problems

Several limitations and prospective research directions are documented:

PMCSF: Lacks a fully autonomous CTE training regime; cognitive vector recovery is enforced via explicit loop closure with CSD.
CogAlign: Noise in real cognitive features and the need for task-matched attention representations remain challenges.
UniCoRN: Real-time decoding is hindered by fMRI temporal resolution and batch processing; current encoders may mis-prioritize filler words; multimodal and spatio-temporal encoding architectures are prospective enhancements (Xi et al., 2023).

A plausible implication is that as CTEs are extended to richer signals (ECoG, combined MEG/EEG), and as cognitive perturbation operators become more diverse, the fidelity and generality of cognitively conditioned text generation and understanding will increase. This suggests technical pathways toward more robust mitigation of “synthetic collapse” and greater interpretability in brain–language interfaces.

References

“The Necessity of Imperfection: Reversing Model Collapse via Simulating Cognitive Boundedness” (Jiang, 1 Dec 2025)
“CogAlign: Learning to Align Textual Neural Representations to Cognitive Language Processing Signals” (Ren et al., 2021)
“UniCoRN: Unified Cognitive Signal ReconstructioN bridging cognitive signals and human language” (Xi et al., 2023)