Lifted Representation Hypothesis in Language Models

2026-06-02

AI Summary

This study investigates how large language models perform type-level categorization of inputs under frozen parameters and why certain predictions change in synchrony, revealing an inherent tension between type-level generalization and local correction. To explain these phenomena, the paper proposes the “elevated state hypothesis,” positing that the model activates shared latent components—elevated states—for behaviorally indistinguishable inputs to support computation. By formalizing the relationship between generalization and correction, and integrating behavioral experiments, formal modeling, and a newly constructed NMR-Type dataset, the work systematically evaluates model behavior across full fine-tuning, LoRA, and in-context learning paradigms. Results demonstrate that models tend to overgeneralize broad rules to subtypes yet struggle with localized updates, providing behavioral evidence for elevated states and offering the first mechanistic account of type-level generalization and non-monotonic revision in large language models.

Abstract

Large language models (LLMs) often answer queries by mapping individual observations to more general rule-like structures. However, it remains unclear how these structures are stored, selected, and revised. To study this process, we propose thelifted representation hypothesis: LLMs update memory through shared latent structures rather than isolated instance-level facts. This view frames lifting as an efficient use of symmetry across instances, and shattering as the refinement of coarse lifted structures into more specific subtypes. We evaluate LLMs'lifting and shattering through controlled exception-learning experiments across in-context learning, LoRA, and full fine-tuning. We find that LLMs are vulnerable to shattering failures when data are governed by nested rules and exceptions, while lifting often occurs prematurely. These results highlight the need to study the relation between data and rule structures in LLMs.

Problem

• lifted state
• type-level generalization
• non-monotonic revision
• in-context learning
• fine-tuning

Innovation

• Lifted State Hypothesis
• non-monotonic revision
• type-level generalization
• computation-relative type
• state reuse