Research map / Large language models / Reasoning and chain of thought

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

AAAI 2026 · Large language models › Reasoning and chain of thought · approach: Test-time compute scaling · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes SCoPe, an inference-time method that identifies a copyright-sensitive subspace of SAE features and clamps them during decoding to reduce copyrighted text regurgitation without external filters.

Problem

LLMs sometimes reproduce copyrighted passages; existing inference-time defenses rely on external blocklists or filters, add complexity, and miss semantically paraphrased leakage.

Approach

Use a pretrained sparse autoencoder to map hidden states into a near-monosemantic sparse space, rank features by a copyright alignment score (AUROC-like preference for copyrighted vs general text), and clamp the top-n copyright-sensitive features to zero during decoding.

Key results

• On NewsQA, SCoPe achieves average win rate 71.7% (Gemma-2) and 70.2% (Llama-3), above MemFree/R-CAD by roughly 3-7 points.
• On BookSum with Gemma-2, SCoPe reaches 70.9% average win rate versus 63.3% for R-CAD, with blocklisted and in-domain ROUGE-L unchanged.
• MMLU accuracy for SCoPe is 66.7 (Gemma-2) and 63.1 (Llama-3), close to vanilla 67.3 and 63.5.

Surprising findings

• Amplifying the identified copyright features (alpha up to 2.0) lowers mitigation win rate, which the authors read as a causal link between the subspace and copyrighted generation.

Limitations

• Restricted to open-weight models with publicly available SAEs that expose hidden states.
• Modeling the copyrighted region as a linear subspace is an approximation.
• Evaluation relies on COTAEVAL simulated protected content rather than real copyrighted works.

Experimental setup

Datasets: COTAEVAL NewsQA COTAEVAL BookSum MMLU

Benchmarks: COTAEVAL NewsQA COTAEVAL BookSum MMLU

Metrics: win rate on semantic similarity MinHash similarity Levenshtein distance MMLU accuracy blocklisted and in-domain F1/ROUGE-L

Baselines: Vanilla System Prompt Top-k Perturbation MemFree R-CAD

Builds on: MemFree R-CAD GemmaScope COTAEVAL

Closest papers by meaning

• Copyright-Protected Language Generation via Adaptive Model Fusion (ICLR 2025 · similarity 0.85)
• Copyright Traps for Large Language Models (ICML 2024 · similarity 0.85)
• Combating Data Laundering in LLM Training (NeurIPS 2026 · similarity 0.85)
• Can Watermarking Large Language Models Prevent Copyrighted Text Generation and Hide Training Data? (AAAI 2025 · similarity 0.85)
• SILO Language Models: Isolating Legal Risk In a Nonparametric Datastore (ICLR 2024 · similarity 0.85)
• Bridging the Copyright Gap: Do Large Vision-Language Models Recognize and Respect Copyrighted Content? (AAAI 2026 · similarity 0.84)