Research map / Large language models / Instruction and fine-tuning
TiTok: Transfer Token-level Knowledge via Contrastive Excess to Transplant LoRA
ChanJoo Jung, Jaehyung Kim
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 token-level excess-score token selection on top of random samples (ablation) | TiTok · BBH (avg 4 transfers) | accuracy: 0.458 -> 0.463 | Improved |
| add excess-score sample filtering (no token selection) (ablation) | TiTok · BBH (avg 4 transfers) | accuracy: 0.458 -> 0.470 | Improved |
| combine sample filtering and token selection (ablation) | TiTok · BBH (avg 4 transfers) | accuracy: 0.470 -> 0.483 | Improved |
| train on top 20% excess tokens instead of bottom 20% (ablation) | TiTok Mistral 7B -> Llama3 8B · BBH | accuracy: 0.468 -> 0.482 | Improved |
| use all aligned token cases instead of one-to-one only (ablation) | TiTok tokenizer alignment · BBH (Mistral 7B -> Llama3 8B) | accuracy: 0.472 -> 0.482 | Improved |
| lower token ratio k% in weak-to-strong transfer (ablation) | TiTok Llama2 7B -> Llama3 8B · BBH | accuracy | Improved |
| lower token ratio k% in weak-to-strong transfer (ablation) | TiTok Llama2 7B -> Llama3 8B · News Headline | ROUGE-L | Hurt |
| apply TiTok on external cross-task data | TiTok Mistral 7B -> Mistral 7B · News Headline cross-task | ROUGE-L: 0.101 -> 0.120 | Improved |
| apply TiTok transfer to a 70B model | Llama-3.3-70B · BBH (4-bit) | accuracy: 0.593 -> 0.621 | Improved |
| TiTok vs TransLoRA on MMLU in same-model transfer | TiTok Mistral 7B -> Mistral 7B · MMLU | accuracy: 0.534 -> 0.561 | Improved |
| TiTok vs vanilla on MMLU for Llama3 3B -> Llama3 8B | TiTok · MMLU | accuracy: 0.469 -> 0.478 | Mixed |
| generate synthetic queries with source expert instead of target model (ablation) | TiTok · BBH | accuracy | Improved |
Closest papers by meaning
- Seeking Neural Nuggets: Knowledge Transfer in Large Language Models from a Parametric Perspective (ICLR 2024 · similarity 0.89)
- $\textit{Trans-LoRA}$: towards data-free Transferable Parameter Efficient Finetuning (NeurIPS 2024 · similarity 0.87)
- KaSA: Knowledge-Aware Singular-Value Adaptation of Large Language Models (ICLR 2025 · similarity 0.87)
- LLaVA-MoD: Making LLaVA Tiny via MoE-Knowledge Distillation (ICLR 2025 · similarity 0.87)
- Text-to-LoRA: Instant Transformer Adaption (ICML 2025 · similarity 0.87)
- Multi-Level Optimal Transport for Universal Cross-Tokenizer Knowledge Distillation on Language Models (AAAI 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.
