Cognitive Computing Frameworks for Scalable Deception Detection in Textual Data

Detecting deception in emotionally grounded natural language remains a significant challenge due to the subtlety and context dependence of deceptive intent. In this work, we use a structured behavioral dataset in which participants produce truthful and deceptive statements under emotional and social constraints. To maintain label accuracy and semantic consistency, we propose a multilayer validation pipeline combining selfconsistency prompting with feedback-guided revision, implemented through the CoTAM (Chain-of-Thought Assisted Modification) method. Our results demonstrate that this framework enhances deception detection by leveraging a sentence decomposition strategy that highlights subtle emotional and strategic cues, improving interpretability for both models and human annotators.

The study compared three LLMs (Gemma-3-27B, LLaMA-3-8B, Mistral-7B) and one transformer baseline (mDeBERTa-v3-base) across one-shot baseline, CoTAM feedback pipeline, and decomposition-enhanced configurations using a 510-input behavioral dataset. Paired t-tests assessed statistical significance, and inference timing was measured for both simple prediction and multi-stage feedback scenarios.

Background & Context

Detecting deception in emotionally grounded natural language remains one of the most significant challenges in natural language processing due to the subtlety and context-dependent nature of deceptive intent. Traditional deception detection approaches have relied on either feature-based linguistic analysis or fine-tuned pre-trained language models, but both struggle with the nuanced emotional and semantic cues that characterize real-world lying behavior.

This 2025 study introduces an innovative cognitive computing framework that addresses a critical gap: most NLP-driven deception detection methods remain limited by dataset design choices that reduce ecological validity, and lack methods that are both linguistically grounded and computationally scalable. The work is particularly timely given the rapid advancement of large language models (LLMs) and their potential application to real-world security, fraud detection, and digital trust systems.

Research Design & Methodology

The study employed a structured behavioral dataset comprising 510 inputs in which participants produced truthful and deceptive statements under emotional and social constraints. This ecologically valid design captures the complex interplay between emotion and deception that characterizes real-world lying behavior.

The core methodological innovation is the CoTAM (Chain-of-Thought Assisted Modification) framework, which implements a multilayer validation pipeline. Key design elements include:

• Semantic decomposition breaking down messages into smaller meaning units to capture emotional and structural patterns linked to deception
• A reflective validation system where LLMs iteratively check and refine their own judgments through prediction, feedback, and revision stages
• A modular framework architecture where each reasoning step can be scaled or optimized independently
• Evaluation combining standard quantitative metrics (F1 and Macro-F1) with qualitative analysis of emotional and semantic coherence

Three major configurations were compared: one-shot baseline inference, the multistep CoTAM feedback pipeline, and a decomposition-enhanced variation, plus hybrid systems combining multiple models. Models tested included Gemma-3-27B, LLaMA-3-8B, Mistral-7B, and the transformer baseline mDeBERTa-v3-base.

Results & Key Findings

The baseline one-shot inference revealed substantial weaknesses in LLMs' ability to detect deception without iterative reasoning. Gemma-3-27B achieved an F1_lie of only 0.19, compared to 0.35 for LLaMA-3-8B and 0.26 for Mistral-7B, indicating that all three models adopted conservative behavior, favoring prediction of 'truthful' over 'lie', with LLaMA-3-8B outperforming Gemma by roughly 16 percentage points.

Key statistical findings:

• Gemma-3-27B showed severe imbalance with 57.2% precision versus 52.3% recall, while Mistral-7B obtained 54.4% precision versus 52.2% recall, confirming conservative tendencies that drove down F1-lie scores
• Paired t-tests confirmed that gains from CoTAM and decomposition strategies were statistically significant (p < 0.05) for Gemma and Mistral, demonstrating that iterative reasoning and sentence-level decomposition meaningfully enhance deception detection
• Simple inference averaged 0.173s per query for Gemma-27B, 0.068s for LLaMA-8B, and 0.073s for Mistral-7B, with decomposed statements adding negligible latency (<0.02s)
• In feedback settings, Gemma-27B required 7.7s per input (1530 queries for 510 statements), LLaMA-8B required 2.8s, and Mistral-7B 1.9s, yet decomposition improved reasoning robustness without additional computational cost

The CoTAM framework enhanced deception detection by leveraging sentence decomposition strategy that highlights subtle emotional and strategic cues, improving interpretability for both models and human annotators.

Discussion & Significance

The framework not only improves deception detection performance but also supports real-world applications such as misinformation control, fraud detection, and digital trust systems through its focus on scalability, flexibility, and systematic design. The finding that iterative reasoning dramatically outperforms single-pass classification challenges conventional approaches to LLM-based deception detection.

The study's most important theoretical contribution is demonstrating that unlike previous methods using single-pass classification, the iterative reasoning strategy with prediction subject to critique and revision reduces the risk of overfitting to superficial lexical cues and increases connection between model decisions and linguistic evidence. This addresses a fundamental weakness in transformer-based deception detection systems.

The computational efficiency findings are particularly significant for practical deployment. The compositional analysis strategy identifies and organizes emotional and semantic cues while remaining efficient enough for distributed computing environments, making the approach viable for large-scale real-time applications.

Limitations & Considerations

In the baseline one-shot scenario with minimal context (one truthful and one deceptive example), all LLMs showed conservative behavior favoring truthful predictions, with Gemma's particularly low F1_lie of 0.19 highlighting that a single forward pass without feedback is insufficient for nuanced deception classification. This suggests the framework's effectiveness may be limited in extremely low-resource scenarios.

The study acknowledges that many NLP-driven deception detection methods remain limited by dataset design choices affecting ecological validity, and methodological choices from dataset design to modality selection affect performance and generalizability. While the behavioral dataset with emotional and social constraints represents an improvement, generalization to other deception contexts (e.g., written vs. spoken, different languages, cultural contexts) requires further validation.

Practical Applications

The CoTAM framework offers immediate applications for security and verification systems. The modular architecture allows deployment at different scales, from edge computing devices requiring simple inference (under 0.1s per query) to high-stakes forensic applications where the feedback pipeline's improved accuracy justifies longer processing times (2-8s per query). The 510-input test set provides a realistic benchmark for operational contexts.

For polygraph examiners and forensic investigators, the framework's emphasis on emotional and semantic decomposition aligns with established interview protocols that focus on statement analysis and emotional consistency. The interpretability improvements—making model reasoning visible through iterative refinement—could enhance expert review of automated deception assessments, bridging the gap between algorithmic detection and human judgment in high-stakes investigative contexts.