Research map / Large language models / Language, safety and interpretability
Language Confusion Gate: Language-Aware Decoding Through Model Self-Distillation
Collin Zhang, Fei Huang, Chenhan Yuan, Junyang Lin
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 |
|---|---|---|---|
| add norm adjustment to gate self-distillation (ablation) | LCG on Llama3.1-8B · FLORES-NO-LATIN | Latin confusion rate (%): 5.7 -> 2.9 | Improved |
| add norm adjustment to gate self-distillation (ablation) | LCG on Qwen3-30B · FLORES-NO-LATIN | Latin confusion rate (%): 0.7 -> 0.4 | Improved |
| add norm adjustment to gate self-distillation (ablation) | LCG on Qwen3-8B · FLORES-NO-LATIN | Latin confusion rate (%): 6.2 -> 2.0 | Improved |
| add norm adjustment to gate self-distillation (ablation) | LCG on Qwen3-8B · FLORES-NO-LATIN | CJ confusion rate (%): 0.5 -> 0.1 | Improved |
| add LCG-adjusted gate at decoding | Qwen3-30B-A3B-Instruct-2507 · FLORES-NO-LATIN | CJ confusion rate (%): 1.0 -> 0.0 | Improved |
| add LCG-adjusted gate at decoding | Qwen3-8B · FLORES-NO-LATIN | Latin confusion rate (%): 12.1 -> 2.0 | Improved |
| add LCG-adjusted gate at decoding | Llama3.1-8B · FLORES-NO-LATIN | Latin confusion rate (%): 8.4 -> 2.9 | Improved |
| add LCG-adjusted gate at decoding | Qwen3-30B · INCLUDE | CJ confusion rate (%): 2.21 -> 0.11 | Improved |
| add LCG-adjusted gate at decoding | Qwen3-30B · INCLUDE | accuracy: 71.12 -> 70.83 | No clear effect |
| add LCG-adjusted gate at decoding | Qwen3-8B (thinking) · Humaneval-XL | Pass@10: 86.67 -> 84.36 | Hurt |
| add LCG-adjusted gate at decoding | GPT-OSS (thinking) · Humaneval-XL | CJ confusion rate (%): 0.38 -> 0.06 | Improved |
| add LCG-adjusted gate at decoding | Qwen3-8B · FLORES-WITH-LATIN | code-switch rate (%): 46.34 -> 25.90 | Hurt |
Closest papers by meaning
- Model-Aware Tokenizer Transfer (NeurIPS 2026 · similarity 0.87)
- Adaptive Draft-Verification for Efficient Large Language Model Decoding (AAAI 2025 · similarity 0.87)
- Cross-model Control: Improving Multiple Large Language Models in One-time Training (NeurIPS 2024 · similarity 0.87)
- Over-Tokenized Transformer: Vocabulary is Generally Worth Scaling (ICML 2025 · similarity 0.86)
- Self-contradictory Hallucinations of Large Language Models: Evaluation, Detection and Mitigation (ICLR 2024 · similarity 0.86)
- Learning How Hard to Think: Input-Adaptive Allocation of LM Computation (ICLR 2025 · 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.
