Research map / Large language models / Instruction and fine-tuning
Exploring the Basin-Like Loss Landscape in Large Language Models
Huanran Chen, Zeming Wei, Yao Huang, Yichi Zhang, Yinpeng Dong, Jun Zhu
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 |
|---|---|---|---|
| increase model size (Qwen2.5-0.5B to 7B to 32B) (ablation) | Qwen2.5 family · most-case landscape (basic, math, safety, coding) | basin size: 32B basin nearly twice the size of 7B | Improved |
| evaluate with generative 0-1 benchmarks instead of likelihood-based benchmarks (ablation) | loss landscape visualization · LLM benchmarks | landscape shape | Mixed |
| fine-tune on data with larger distribution gap from original data (Alpaca vs benign) (ablation) | Qwen2.5-7B SFT · SFT-case loss landscape | landscape width | Hurt |
| fine-tune on adversarial AdvBench data for 10 steps (ablation) | Qwen2.5-7B · safety capability | safety | Hurt |
| move along optimized worst-case direction instead of random direction (ablation) | LLMs · most-case vs worst-case landscape | capability retention | Hurt |
| add Gaussian-augmented (GO) optimizer in pre-training | GPT2-127M pre-training on OpenWebText · basin size | basin size | Improved |
| add GO optimizer (sigma=0.01) instead of Adam | GPT2-127M pre-training · early pre-training loss | pre-training loss | Hurt |
| use GO optimizer instead of SAM or continuous dropout | GPT2-127M fine-tuned on Alpaca · SFT degradation | performance degradation after SFT | Improved |
| suppress degradation under Gaussian noise via GO training | GPT2-127M · subsequent Alpaca fine-tuning | fine-tuning degradation | Improved |
| longer pre-training training steps (ablation) | GPT2-127M · basin evolution | basin width | Improved |
Closest papers by meaning
- Navigating the Safety Landscape: Measuring Risks in Finetuning Large Language Models (NeurIPS 2024 · similarity 0.90)
- Strong Model Collapse (ICLR 2025 · similarity 0.88)
- Remarkable Robustness of LLMs: Stages of Inference? (NeurIPS 2025 · similarity 0.88)
- Large language models suffer from a curse of ambiguity (NeurIPS 2026 · similarity 0.87)
- An Empirical Study on Noisy Data and LLM Pretraining Loss Divergence (NeurIPS 2026 · similarity 0.87)
- LLMs on the Line: Data Determines Loss-to-Loss Scaling Laws (ICML 2025 · similarity 0.87)
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.
