Domain Restriction via Multi SAE Layer Transitions

2026-05-12

AI Summary

This work addresses the vulnerability of general-purpose large language models (LLMs) to out-of-distribution (OOD) inputs in domain-specific applications, which often leads to unintended behaviors. The authors propose a lightweight method that, for the first time, leverages the dynamic evolution of representations from multi-layer sparse autoencoders (SAEs) to model the transformation of internal activations across LLM layers. This approach enables fine-grained capture of domain-specific features for effective OOD detection. Evaluated on Gemma-2 2B and 9B models, the method significantly improves OOD detection performance while simultaneously enhancing controllability and interpretability within the target domain. Furthermore, it reveals systematic patterns in how domain-related semantics evolve throughout the LLM’s processing pipeline.

Abstract

The general-purpose nature of Large Language Models (LLMs) presents a significant challenge for domain-specific applications, often leading to out-of-domain (OOD) interactions that undermine the provider's intent. Existing methods for detecting such scenarios treat the LLM as an uninterpretable black box and overlook the internal processing of inputs. In this work we show that layer transitions provide a promising avenue for extracting domain-specific signature. Specifically, we present several lightweight ways of learning on internal dynamics encoded using a sparse autoencoder (SAE) that exhibit great capability in distinguishing OOD texts. Building on top of SAEs representation transitions enables us to better interpret the LLM internal evolution of input processing and shed light on its decisions. We provide a comprehensive analysis of the method and benchmark it with the gemma-2 2B and 9B models. Our results emphasize the efficacy of the internal process in capturing fine-grained input-related details.

Problem

• domain restriction
• out-of-domain detection
• large language models
• internal dynamics
• sparse autoencoder

Innovation

• Sparse Autoencoder
• Layer Transitions
• Out-of-Domain Detection
• Interpretable LLM
• Internal Representation Dynamics