Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

UniCog: Uncovering Cognitive Abilities of LLMs through Latent Mind Space Analysis

Jiayu Liu, Yinhe Long, Zhenya Huang, Enhong Chen

TL;DR

UniCog presents a latent-mind framework to analyze LLM cognition by embedding diverse cognitive abilities into a sparse, disentangled latent space and modeling $X\sim p_\theta(X|Z)$ with $Z\sim p(Z)$. By using a language-space surrogate likelihood and a sparse posterior, the method reveals a Pareto structure where a shared reasoning core coexists with ability-specific signatures, and shows that erroneous reasoning amplifies latent activations. The authors demonstrate the surrogate’s fidelity, identify measurable signals of reasoning correctness, and propose a latent-informed candidate prioritization that improves reasoning accuracy by up to $7.5\%$ across multiple benchmarks and model families. This work provides a cognitively grounded perspective on LLM reasoning and offers a practical, plug-in mechanism to enhance reasoning reliability, with potential for real-time activation steering in future work.

Abstract

A growing body of research suggests that the cognitive processes of large language models (LLMs) differ fundamentally from those of humans. However, existing interpretability methods remain limited in explaining how cognitive abilities are engaged during LLM reasoning. In this paper, we propose UniCog, a unified framework that analyzes LLM cognition via a latent mind space. Formulated as a latent variable model, UniCog encodes diverse abilities from dense model activations into sparse, disentangled latent dimensions. Through extensive analysis on six advanced LLMs, including DeepSeek-V3.2 and GPT-4o, we reveal a Pareto principle of LLM cognition, where a shared reasoning core is complemented by ability-specific signatures. Furthermore, we discover that reasoning failures often manifest as anomalous intensity in latent activations. These findings opens a new paradigm in LLM analysis, providing a cognition grounded view of reasoning dynamics. Finally, leveraging these insights, we introduce a latent-informed candidate prioritization strategy, which improves reasoning performance by up to 7.5% across challenging benchmarks. Our code is available at https://github.com/milksalute/unicog.

UniCog: Uncovering Cognitive Abilities of LLMs through Latent Mind Space Analysis

TL;DR

UniCog presents a latent-mind framework to analyze LLM cognition by embedding diverse cognitive abilities into a sparse, disentangled latent space and modeling with . By using a language-space surrogate likelihood and a sparse posterior, the method reveals a Pareto structure where a shared reasoning core coexists with ability-specific signatures, and shows that erroneous reasoning amplifies latent activations. The authors demonstrate the surrogate’s fidelity, identify measurable signals of reasoning correctness, and propose a latent-informed candidate prioritization that improves reasoning accuracy by up to across multiple benchmarks and model families. This work provides a cognitively grounded perspective on LLM reasoning and offers a practical, plug-in mechanism to enhance reasoning reliability, with potential for real-time activation steering in future work.

Abstract

A growing body of research suggests that the cognitive processes of large language models (LLMs) differ fundamentally from those of humans. However, existing interpretability methods remain limited in explaining how cognitive abilities are engaged during LLM reasoning. In this paper, we propose UniCog, a unified framework that analyzes LLM cognition via a latent mind space. Formulated as a latent variable model, UniCog encodes diverse abilities from dense model activations into sparse, disentangled latent dimensions. Through extensive analysis on six advanced LLMs, including DeepSeek-V3.2 and GPT-4o, we reveal a Pareto principle of LLM cognition, where a shared reasoning core is complemented by ability-specific signatures. Furthermore, we discover that reasoning failures often manifest as anomalous intensity in latent activations. These findings opens a new paradigm in LLM analysis, providing a cognition grounded view of reasoning dynamics. Finally, leveraging these insights, we introduce a latent-informed candidate prioritization strategy, which improves reasoning performance by up to 7.5% across challenging benchmarks. Our code is available at https://github.com/milksalute/unicog.
Paper Structure (30 sections, 7 equations, 10 figures, 3 tables)

This paper contains 30 sections, 7 equations, 10 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Cognitive Abilities of LLMs
  4. Interpretability of LLMs
  5. Our Framework
  6. Latent Variable Formulation
  7. Transformer-Based Parameterization
  8. Analysis on Cognitive Abilities of LLMs
  9. Experimental Setup
  10. Framework Training.
  11. Evaluation Benchmark.
  12. Our Findings
  13. Our surrogate model $\mathcal{M}_{\text{like}}$ can accurately simulate the generation of diverse LLMs.
  14. Pareto principle between latent mind dimensions and cognitive abilities across all LLMs.
  15. Cognitive variants amplify the scale of latent minds.
  16. ...and 15 more sections

Figures (10)

  • Figure 1: Comparison of previous interpretability methods and our UniCog framework.
  • Figure 2: Illustration of our UniCog framework.
  • Figure 3: Surrogate likelihood and posterior modeling based on language space.
  • Figure 4: (a) Perplexity of reasoning outputs sampled from LLMs evaluated by our surrogate model (left y-axis). BLEU/ROUGE scores of reconstructed outputs relative to the original outputs (right y-axis). (b) Average activation strength relative to root mind for each cognitive variant. (c) Normalized distances among cognitive variants.
  • Figure 5: Average activation strength ratio of dimensions $\{3576, 11211, 12000\}$ in incorrect versus correct cases across different cognitive variants.
  • ...and 5 more figures