Research map / Label-efficient and robust learning / Knowledge distillation
CTPD: Cross Tokenizer Preference Distillation
Truong Nguyen, Phi Van Dat, Ngan Nguyen, Linh Ngo Van, Trung Le, Thanh Hong Nguyen
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 CTPD instead of TIS-DPO | Llama-3.1-8B student · Six-benchmark average | accuracy (%): 66.16 -> 67.42 | Improved |
| use CTPD instead of TIS-DPO | Llama-3.2-1B student · Six-benchmark average | accuracy (%): 42.60 -> 43.26 | Improved |
| use CTPD instead of TIS-DPO | Llama-3.1-8B student · GSM8k | accuracy (%): 54.31 -> 57.47 | Improved |
| use CTPD instead of TIS-DPO | Llama-3.1-8B student · TruthfulQA | accuracy (%): 53.86 -> 55.22 | Improved |
| use CTPD instead of cross-tokenizer KD (Multi-Level OT) | Llama-3.1-8B student · Six-benchmark average | accuracy (%): 64.58 -> 67.42 | Improved |
| use CTPD instead of DSKD | Llama-3.2-1B student · Six-benchmark average | accuracy (%): 41.12 -> 43.26 | Improved |
| add teacher-estimated span weights instead of random weights (ablation) | CTPD · Six-benchmark average | accuracy (%): 54.80 -> 67.42 | Improved |
| teacher-estimated weights instead of length-averaged weights (ablation) | CTPD · Six-benchmark average | accuracy (%): 65.47 -> 67.42 | Improved |
| teacher-estimated weights instead of student-estimated weights (ablation) | CTPD · Six-benchmark average | accuracy (%): 65.88 -> 67.42 | Improved |
| teacher-estimated weights instead of teacher-student SFT estimate (ablation) | CTPD · Six-benchmark average | accuracy (%): 64.51 -> 67.42 | Improved |
| use teacher instead of student as reference model (ablation) | CTPD · Six-benchmark average | accuracy (%): 65.27 -> 67.42 | Improved |
| use CTPD instead of DPO | Llama-3.1-8B student · MMLU | accuracy (%): 65.26 -> 66.65 | Mixed |
Closest papers by meaning
- Advantage-Guided Distillation for Preference Alignment in Small Language Models (ICLR 2025 · similarity 0.91)
- BLEUBERI: BLEU is a surprisingly effective reward for instruction following (NeurIPS 2025 · similarity 0.89)
- SimCT: Recovering Lost Supervision for Cross-Tokenizer On-Policy Distillation (NeurIPS 2026 · similarity 0.89)
- HelpSteer3-Preference: Open Human-Annotated Preference Data across Diverse Tasks and Languages (NeurIPS 2025 · similarity 0.88)
- Improving Context-Aware Preference Modeling for Language Models (NeurIPS 2024 · similarity 0.88)
- SIMPLEMIX: Frustratingly Simple Mixing of Off- and On-policy Data in Language Model Preference Learning (ICML 2025 · similarity 0.88)
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.
