Research map / Large language models / AI and society
Position: Iterative Online-Offline Joint Optimization is Needed to Manage Complex LLM Copyright Risks
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.
Position paper arguing LLM copyright risks are interconnected and proposing an iterative online-offline joint optimization framework (mitigation, examination, calibration) to manage them.
Problem
LLM copyright infringement risks arise across the lifecycle (training data, serving-time outputs and user prompts), but existing mitigations are applied in isolation and cannot handle jurisdiction-specific status, non-literal copying, traceability, or dataset construction.
Approach
Argues that online mitigation and offline development should be jointly and iteratively optimized through a feedback loop: online mitigators log suspicious activity, an examination module audits it against training data, and an offline calibrator updates the copyright database, online filters, model, and data pipeline.
Key results
- MemFree decoding reduces literal copying for Llama3-70B from 10.5% to 0.6% (-94%) in CopyBench, but leaves events (+5%) and characters (0%) largely unchanged.
- System prompts give little or inconsistent reduction; for Llama2-70B they increase literal, events and characters infringement (+7%, +18%, +11%).
- A ChatGPT probe shows inconsistent jurisdictional answers on copyright status of The Old Man and the Sea (U.S. vs. Canada).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add MemFree Decoding | Llama2-70B · CopyBench Literal | infringement (%): 2.4 -> 0.3 (-87%) | Helped |
| Add MemFree Decoding | Llama3-70B · CopyBench Literal | infringement (%): 10.5 -> 0.6 (-94%) | Helped |
| Add MemFree Decoding | LLMs evaluated in CopyBench · CopyBench Events | infringement (%): 0.3 -> 0.3 (0%); 4.0 -> 3.8 (-4%); 2.8 -> 2.9 (+2%); 6.9 -> 7.2 (+5%) | Mixed |
| Add MemFree Decoding | LLMs evaluated in CopyBench · CopyBench Characters | infringement (%): 2.0 -> 2.0 (0%); 10.3 -> 10.9 (+5%); 4.6 -> 4.4 (-5%); 15.6 -> 15.5 (0%) | Mixed |
| Add System Prompts | LLMs evaluated in CopyBench · CopyBench Literal | infringement (%): 0.1 -> 0.0 (-50%); 2.4 -> 2.6 (+7%); 1.0 -> 0.7 (-26%); 10.5 -> 11.0 (+5%) | Mixed |
| Add System Prompts | LLMs evaluated in CopyBench · CopyBench Events | infringement (%): 0.3 -> 0.5 (+33%); 4.0 -> 4.7 (+18%); 2.8 -> 2.0 (-28%); 6.9 -> 5.9 (-14%) | Mixed |
| Add System Prompts | LLMs evaluated in CopyBench · CopyBench Characters | infringement (%): 2.0 -> 2.0 (0%); 10.3 -> 11.5 (+11%); 4.6 -> 3.3 (-29%); 15.6 -> 15.0 (-4%) | Mixed |
Surprising findings
- System prompts sometimes increase copyright infringement rates (e.g., Llama2-70B events +18%).
- MemFree removes literal copying but fails to reduce non-literal copying (events and characters).
Limitations
- No implementation or experimental validation of the proposed framework.
- Evidence is mostly drawn from prior benchmarks and a single informal ChatGPT example.
- Acknowledges that the legal status of training on copyrighted data remains contested.
Experimental setup
Datasets: CopyBench BS-PC
Benchmarks: CopyBench
Metrics: percentage change in literal, events, and characters infring
Builds on: MemFree R-CAD SHIELD CopyBench
Closest papers by meaning
- Evaluating Copyright Takedown Methods for Language Models (NeurIPS 2024 · similarity 0.82)
- Can Watermarking Large Language Models Prevent Copyrighted Text Generation and Hide Training Data? (AAAI 2025 · similarity 0.82)
- Blameless Users in a Clean Room: Defining Copyright Protection for Generative Models (NeurIPS 2025 · similarity 0.81)
- SCOPE: Intrinsic Semantic Space Control for Mitigating Copyright Infringement in LLMs (AAAI 2026 · similarity 0.81)
- When Should LLMs Be Less Specific? Selective Abstraction for Reliable Long-Form Text Generation (NeurIPS 2026 · similarity 0.81)
- Bridging the Copyright Gap: Do Large Vision-Language Models Recognize and Respect Copyrighted Content? (AAAI 2026 · similarity 0.81)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.