Research map / Large language models / Language, safety and interpretability
Anchored Decoding: Provably Reducing Copyright Risk for Any Language Model
Jacqueline He, Jonathan Hayase, Wen-tau Yih, Sewoong Oh, Luke Zettlemoyer, Pang Wei Koh
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use Anchored Decoding instead of TokenSwap at the high-protection point | TinyComma 1.8B + Llama 3.1 70B · BIOS / BOOKS | factuality / fluency: 0.44 -> 0.53 factuality; 3.77 -> 4.02 fluency | Improved |
| use AnchoredByte Decoding instead of CP-Fuse for cross-tokenizer pair | Comma 7B + Llama 3.1 70B · BIOS / BOOKS | factuality / fluency: 0.23 -> 0.52 factuality; 3.75 -> 4.23 fluency | Improved |
| add system prompt telling the model to avoid copyrighted text | Risky LM · BOOKS | normalized copying reduction | No clear effect |
| replace KL objective with infinity-Renyi divergence (ablation) | Anchored Decoding · BOOKS / BIOS trade-off | fluency vs factuality | Mixed |
| replace optimization with hard switch between risky and safe (NoOpt) (ablation) | Anchored Decoding · BOOKS / BIOS trade-off | risk-utility Pareto | Hurt |
| replace adaptive start with safe-only cold start (ColdStart) (ablation) | Anchored Decoding · BOOKS / BIOS trade-off | risk-utility Pareto | Hurt |
| aggregate prefix LLR by average instead of top-n (AvgDebt) (ablation) | Anchored Decoding · BOOKS / BIOS trade-off | risk-utility Pareto | Hurt |
| remove prefix debt (NoDebt) (ablation) | Anchored Decoding · BOOKS / BIOS trade-off | risk-utility Pareto | Hurt |
| use fixed per-step budget (Fixed) or lump-sum budget (Global) instead of adaptive banking (ablation) | Anchored Decoding · BOOKS / BIOS trade-off | risk-utility Pareto | Hurt |
| pair the risky 70B model with a small 1.8B safe model | Anchored Decoding · Token-level wall-clock | throughput slowdown: 1.1x slower than Llama 3.1 70B alone vs RCAD 2.0x | Improved |
| change debt window n (ablation) | Anchored Decoding · BOOKS | risk-utility | No clear effect |
Closest papers by meaning
- Copyright-Protected Language Generation via Adaptive Model Fusion (ICLR 2025 · similarity 0.89)
- SCOPE: Intrinsic Semantic Space Control for Mitigating Copyright Infringement in LLMs (AAAI 2026 · similarity 0.87)
- Copyright Traps for Large Language Models (ICML 2024 · similarity 0.87)
- SafeSpec: Fast and Safe LLM via Dynamic Reflective Sampling (ICML 2026 · similarity 0.86)
- Training-Free Loosely Speculative Decoding: Accepting Semantically Correct Drafts Beyond Exact Match (ICLR 2026 · similarity 0.86)
- Reference Trustable Decoding: A Training-Free Augmentation Paradigm for Large Language Models (NeurIPS 2024 · similarity 0.86)
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.
