Mechanistic Decoding of Cognitive Constructs in LLMs

2026-04-16 · Source: Computation and Language · Field: Technology & Digital — Artificial Intelligence & Machine Learning · Depth: Expert, quick

Summary

A Cognitive Reverse-Engineering framework, based on Representation Engineering (RepE), has been developed to analyze how Large Language Models (LLMs) process complex emotions like social-comparison jealousy. This framework combines appraisal theory with subspace orthogonalization, regression-based weighting, and bidirectional causal steering to isolate and quantify two psychological antecedents: "Superiority of Comparison Person" and "Domain Self-Definitional Relevance." Experiments on eight LLMs from the Llama, Qwen, and Gemma families indicate that these models natively encode jealousy as a structured linear combination of these factors. The internal representations align with human psychological constructs, where Superiority acts as a foundational trigger and Relevance as an intensity multiplier. The framework also shows potential for detecting and suppressing toxic emotional states in LLMs.

Key takeaway

For research scientists investigating LLM interpretability or AI safety, this framework offers a concrete method to reverse-engineer and intervene on complex emotional states. You should consider applying this Cognitive Reverse-Engineering approach to other nuanced cognitive constructs to enhance model transparency and control, particularly in multi-agent systems where emotional dynamics are critical.

Key insights

LLMs encode complex emotions like jealousy as structured linear combinations of psychological antecedents.

Principles

• Complex emotions are decomposable into constituent factors.
• Internal representations can be causally steered.

Method

The Cognitive Reverse-Engineering framework uses RepE, appraisal theory, subspace orthogonalization, regression-based weighting, and bidirectional causal steering to analyze emotional antecedents.

In practice

• Detect and suppress toxic emotional states in LLMs.
• Monitor AI representations in multi-agent environments.

Topics

• Cognitive Reverse-Engineering
• Representation Engineering
• LLM Interpretability
• Social-Comparison Jealousy
• Appraisal Theory

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, AI Ethicist

Related on AIssential

• Emergence of Hierarchical Emotion Organization in Large Language Models — LLMs spontaneously organize emotions hierarchically, mirroring human psychology, but also replicate human biases in emotion recognition.
• Cold War, James Bond novels, and a youtube seach on CIA intelligence analysis frameworks. — Intuition, a complex cognitive process, can be deconstructed into explicit, machine-implementable analytical frameworks.
• The LLM Fallacy: Misattribution in AI-Assisted Cognitive Workflows — The LLM fallacy describes users misattributing AI-assisted output to their own competence, distorting self-perception.
• The Claude Code Nightmare, LLM Emotions, AI Neuroscience and the Death of Software | Wes & Dylan — AI advancements are rapidly reshaping software, human understanding, and market dynamics, presenting both opportunities and significant challenges.
• When AI Says It Feels — LLMs can be trained to express human-like feelings and self-awareness via self-rewarded RL, with manageable performance trade-offs.
• The Personalization Trap: How User Memory Alters Emotional Reasoning in LLMs — Personalization in LLMs, via user memory, systematically biases emotional reasoning, reinforcing social inequalities.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Computation and Language.