Research map / Large language models / Instruction and fine-tuning
SASFT: Sparse Autoencoder-guided Supervised Finetuning to Mitigate Unexpected Code-Switching in LLMs
Boyi Deng, Yu Wan, Baosong Yang, Fei Huang, Wenjie Wang, Fuli Feng
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 SAE-guided auxiliary loss on language features | SFT · code-switching any->ko, Qwen3-1.7B-Base 210k | code-switching ratio (%): 0.36 -> 0.00 | Improved |
| add SAE-guided auxiliary loss on language features | SFT · code-switching any->zh, Gemma-2-9B 210k | code-switching ratio (%): 0.78 -> 0.41 | Improved |
| add SAE-guided auxiliary loss on language features | SFT · code-switching any->zh, Qwen3-8B-Base 210k | code-switching ratio (%): 0.96 -> 0.66 | Improved |
| use SASFT instead of GRPO language-consistency reward | SFT+GRPO · code-switching, five models and three languages | code-switching ratio (%): Llama-3.1-8B ko 110k: GRPO 0.95 (+157%) vs SASFT 0.20 | Improved |
| set alpha to pre-estimated average instead of zero (ablation) | SASFT · code-switching any->ru, Gemma-2-2B 210k | code-switching ratio (%): 0.90 -> 0.22 | Improved |
| apply SASFT to multiple layers instead of one (ablation) | SASFT · code-switching ratio across models | code-switching ratio (%) | Improved |
| apply SASFT to multiple features instead of one (ablation) | SASFT · code-switching to Chinese | code-switching ratio (%) | Improved |
| apply SASFT to earlier single layer (ablation) | SASFT · code-switching ratio, single layer | code-switching ratio (%) | Hurt |
| add SASFT | SFT · MMLU, Llama-3.1-8B | accuracy (%): 29.99 -> 33.12 | Improved |
| add SASFT | SFT · HumanEval, Llama-3.1-8B | accuracy (%): 87.74 -> 91.88 | Improved |
| add SASFT | SFT · MMLU, Qwen3-8B-Base | accuracy (%): 52.15 -> 50.09 | Hurt |
| add SASFT | SFT · LogiQA, Gemma-2-2B | accuracy (%): 28.00 -> 25.75 | Hurt |
Closest papers by meaning
- Unveiling the Secret Recipe: A Guide For Supervised Fine-Tuning Small LLMs (ICLR 2025 · similarity 0.89)
- When Scaling Meets LLM Finetuning: The Effect of Data, Model and Finetuning Method (ICLR 2024 · similarity 0.89)
- Cross-model Control: Improving Multiple Large Language Models in One-time Training (NeurIPS 2024 · similarity 0.89)
- SAEBench: A Comprehensive Benchmark for Sparse Autoencoders in Language Model Interpretability (ICML 2025 · similarity 0.89)
- Preference-Oriented Supervised Fine-Tuning: Favoring Target Model over Aligned Large Language Models (AAAI 2025 · similarity 0.89)
- Linearization Explains Fine-Tuning in Large Language Models (NeurIPS 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.
