SCOPE: Intrinsic Semantic Space Control for Mitigating Copyright Infringement in LLMs

2025-11-10

AI Summary

Large language models (LLMs) may inadvertently reproduce copyrighted content during generation; existing inference-time defenses—relying on external filters or surface-level string matching—struggle to mitigate semantic paraphrasing-based leakage. This work reframes copyright protection as an internal semantic-space control problem within the LLM, proposing a lightweight, parameter-free intervention method that requires no external components. Specifically, we employ sparse autoencoders (SAEs) to interpret hidden-layer representations, identifying and isolating copyright-sensitive, high-dimensional, near-monosemantic subspaces; during decoding, we dynamically suppress activations in these subspaces. Evaluated on standard copyright detection benchmarks, our approach substantially reduces infringement risk while preserving generation quality and general-purpose capabilities. Interpretability analysis confirms that the identified subspaces encode high-level semantic features, validating their relevance to copyright-sensitive content. To our knowledge, this is the first method to address copyright leakage via intrinsic, representation-level intervention without fine-tuning or auxiliary modules.

Technology Category

• Machine Learning: Large Multimodal Models (LMMs)
• Natural Language Processing: (Large) Language Models
• Computer Vision: Large Vision Models

Application Category

• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Search and Retrieval-Augmented AI: Large language models for search
• User Modeling, Personalization and Recommendation: Large Language Models (LLM) for user modeling and recommendation

Abstract

Large language models sometimes inadvertently reproduce passages that are copyrighted, exposing downstream applications to legal risk. Most existing studies for inference-time defences focus on surface-level token matching and rely on external blocklists or filters, which add deployment complexity and may overlook semantically paraphrased leakage. In this work, we reframe copyright infringement mitigation as intrinsic semantic-space control and introduce SCOPE, an inference-time method that requires no parameter updates or auxiliary filters. Specifically, the sparse autoencoder (SAE) projects hidden states into a high-dimensional, near-monosemantic space; benefiting from this representation, we identify a copyright-sensitive subspace and clamp its activations during decoding. Experiments on widely recognized benchmarks show that SCOPE mitigates copyright infringement without degrading general utility. Further interpretability analyses confirm that the isolated subspace captures high-level semantics.

Problem

• Mitigating copyright infringement in large language models
• Reducing legal risks from reproduced copyrighted passages
• Controlling semantic leakage without external filters

Innovation

• Intrinsic semantic-space control without parameter updates
• Sparse autoencoder projects hidden states into semantic space
• Clamp copyright-sensitive subspace activations during decoding