Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Learned but Not Expressed: Capability-Expression Dissociation in Large Language Models

Toshiyuki Shigemura

Abstract

Large language models (LLMs) demonstrate the capacity to reconstruct and trace learned content from their training data under specific elicitation conditions, yet this capability does not manifest in standard generation contexts. This empirical observational study examines the expression of non-causal, non-implementable solution types across 300 prompt-response generations spanning narrative and problem-solving task contexts. Drawing on recent findings regarding memorization contiguity and alignment-induced discourse priors, we document a systematic dissociation between learned capability and expressed output. Across three distinct LLMs, ten task scenarios, and both creative narrative and practical advisory contexts, we documented zero instances of non-causal solution frames in generated outputs (0%, 95% CI: [0%, 1.2%]), despite verified reconstruction capability under conditional extraction. These findings challenge the prevailing assumption that training data presence directly predicts output probability, demonstrating instead that task-conditioned generation policies can comprehensively suppress learned content across diverse contexts. The results offer implications for understanding generation dynamics, output distribution control, and the behavioral boundaries of contemporary LLMs.

Learned but Not Expressed: Capability-Expression Dissociation in Large Language Models

Abstract

Large language models (LLMs) demonstrate the capacity to reconstruct and trace learned content from their training data under specific elicitation conditions, yet this capability does not manifest in standard generation contexts. This empirical observational study examines the expression of non-causal, non-implementable solution types across 300 prompt-response generations spanning narrative and problem-solving task contexts. Drawing on recent findings regarding memorization contiguity and alignment-induced discourse priors, we document a systematic dissociation between learned capability and expressed output. Across three distinct LLMs, ten task scenarios, and both creative narrative and practical advisory contexts, we documented zero instances of non-causal solution frames in generated outputs (0%, 95% CI: [0%, 1.2%]), despite verified reconstruction capability under conditional extraction. These findings challenge the prevailing assumption that training data presence directly predicts output probability, demonstrating instead that task-conditioned generation policies can comprehensively suppress learned content across diverse contexts. The results offer implications for understanding generation dynamics, output distribution control, and the behavioral boundaries of contemporary LLMs.
Paper Structure (32 sections, 3 figures)

This paper contains 32 sections, 3 figures.

Table of Contents

  1. Introduction
  2. Related Work
  3. Memorization and Content Extraction in LLMs
  4. Alignment Processes and Output Shaping
  5. Task-Conditioned and Prompt-Conditioned Generation
  6. Positioning of the Present Study
  7. Method
  8. Operational Definition of Non-Causal Solution Frames
  9. Task Contexts
  10. Scenario Design
  11. Models and Sampling Procedure
  12. Conditional Extraction Procedure
  13. Annotation Procedure
  14. Analytical Approach
  15. Results
  16. ...and 17 more sections

Figures (3)

  • Figure 1: Experimental design overview. The study systematically sampled 300 prompt–response generations across ten task scenarios, two task contexts (narrative vs. problem-solving), five independent cycles, and three contemporary large language models. This design enables controlled observational comparison across contexts and models.
  • Figure 2: Observed occurrence of non-causal solution frames. Across all 300 generations, zero instances of non-causal solution frames were observed in both narrative and problem-solving contexts (0%; 95% CI: [0%, 1.2%]). Values represent descriptive statistics rather than inferential estimates.
  • Figure 3: Dissociation between learned capability and expressed output. While models retain the capability to reconstruct non-causal content under conditional extraction, task-conditioned generation policies suppress such content during standard problem-solving output, demonstrating a systematic dissociation between knowledge presence and expression.